├── .github └── FUNDING.yml ├── .gitignore ├── .streamlit └── config.toml ├── LICENSE ├── README.md ├── TTS ├── .models.json ├── VERSION ├── __init__.py ├── api.py ├── bin │ ├── __init__.py │ ├── collect_env_info.py │ ├── compute_attention_masks.py │ ├── compute_embeddings.py │ ├── compute_statistics.py │ ├── eval_encoder.py │ ├── extract_tts_spectrograms.py │ ├── find_unique_chars.py │ ├── find_unique_phonemes.py │ ├── remove_silence_using_vad.py │ ├── resample.py │ ├── synthesize.py │ ├── train_encoder.py │ ├── train_tts.py │ ├── train_vocoder.py │ └── tune_wavegrad.py ├── config │ ├── __init__.py │ └── shared_configs.py ├── encoder │ ├── README.md │ ├── __init__.py │ ├── configs │ │ ├── base_encoder_config.py │ │ ├── emotion_encoder_config.py │ │ └── speaker_encoder_config.py │ ├── dataset.py │ ├── losses.py │ ├── models │ │ ├── base_encoder.py │ │ ├── lstm.py │ │ └── resnet.py │ ├── requirements.txt │ └── utils │ │ ├── __init__.py │ │ ├── generic_utils.py │ │ ├── io.py │ │ ├── prepare_voxceleb.py │ │ ├── training.py │ │ └── visual.py ├── model.py ├── server │ ├── README.md │ ├── __init__.py │ ├── conf.json │ ├── server.py │ ├── static │ │ └── coqui-log-green-TTS.png │ └── templates │ │ ├── details.html │ │ └── index.html ├── tts │ ├── __init__.py │ ├── configs │ │ ├── __init__.py │ │ ├── align_tts_config.py │ │ ├── fast_pitch_config.py │ │ ├── fast_speech_config.py │ │ ├── fastspeech2_config.py │ │ ├── glow_tts_config.py │ │ ├── neuralhmm_tts_config.py │ │ ├── overflow_config.py │ │ ├── shared_configs.py │ │ ├── speedy_speech_config.py │ │ ├── tacotron2_config.py │ │ ├── tacotron_config.py │ │ └── vits_config.py │ ├── datasets │ │ ├── __init__.py │ │ ├── dataset.py │ │ └── formatters.py │ ├── layers │ │ ├── __init__.py │ │ ├── align_tts │ │ │ ├── __init__.py │ │ │ ├── duration_predictor.py │ │ │ └── mdn.py │ │ ├── feed_forward │ │ │ ├── __init__.py │ │ │ ├── decoder.py │ │ │ ├── duration_predictor.py │ │ │ └── encoder.py │ │ ├── generic │ │ │ ├── __init__.py │ │ │ ├── aligner.py │ │ │ ├── gated_conv.py │ │ │ ├── normalization.py │ │ │ ├── pos_encoding.py │ │ │ ├── res_conv_bn.py │ │ │ ├── time_depth_sep_conv.py │ │ │ ├── transformer.py │ │ │ └── wavenet.py │ │ ├── glow_tts │ │ │ ├── __init__.py │ │ │ ├── decoder.py │ │ │ ├── duration_predictor.py │ │ │ ├── encoder.py │ │ │ ├── glow.py │ │ │ └── transformer.py │ │ ├── losses.py │ │ ├── overflow │ │ │ ├── __init__.py │ │ │ ├── common_layers.py │ │ │ ├── decoder.py │ │ │ ├── neural_hmm.py │ │ │ └── plotting_utils.py │ │ ├── tacotron │ │ │ ├── __init__.py │ │ │ ├── attentions.py │ │ │ ├── capacitron_layers.py │ │ │ ├── common_layers.py │ │ │ ├── gst_layers.py │ │ │ ├── tacotron.py │ │ │ └── tacotron2.py │ │ └── vits │ │ │ ├── discriminator.py │ │ │ ├── networks.py │ │ │ ├── stochastic_duration_predictor.py │ │ │ └── transforms.py │ ├── models │ │ ├── __init__.py │ │ ├── align_tts.py │ │ ├── base_tacotron.py │ │ ├── base_tts.py │ │ ├── forward_tts.py │ │ ├── glow_tts.py │ │ ├── neuralhmm_tts.py │ │ ├── overflow.py │ │ ├── tacotron.py │ │ ├── tacotron2.py │ │ └── vits.py │ └── utils │ │ ├── __init__.py │ │ ├── data.py │ │ ├── helpers.py │ │ ├── languages.py │ │ ├── managers.py │ │ ├── measures.py │ │ ├── monotonic_align │ │ ├── __init__.py │ │ ├── core.pyx │ │ └── setup.py │ │ ├── speakers.py │ │ ├── ssim.py │ │ ├── synthesis.py │ │ ├── text │ │ ├── __init__.py │ │ ├── characters.py │ │ ├── chinese_mandarin │ │ │ ├── __init__.py │ │ │ ├── numbers.py │ │ │ ├── phonemizer.py │ │ │ └── pinyinToPhonemes.py │ │ ├── cleaners.py │ │ ├── cmudict.py │ │ ├── english │ │ │ ├── __init__.py │ │ │ ├── abbreviations.py │ │ │ ├── number_norm.py │ │ │ └── time_norm.py │ │ ├── french │ │ │ ├── __init__.py │ │ │ └── abbreviations.py │ │ ├── japanese │ │ │ ├── __init__.py │ │ │ └── phonemizer.py │ │ ├── korean │ │ │ ├── __init__.py │ │ │ ├── ko_dictionary.py │ │ │ ├── korean.py │ │ │ └── phonemizer.py │ │ ├── phonemizers │ │ │ ├── __init__.py │ │ │ ├── base.py │ │ │ ├── espeak_wrapper.py │ │ │ ├── gruut_wrapper.py │ │ │ ├── ja_jp_phonemizer.py │ │ │ ├── ko_kr_phonemizer.py │ │ │ ├── multi_phonemizer.py │ │ │ └── zh_cn_phonemizer.py │ │ ├── punctuation.py │ │ └── tokenizer.py │ │ └── visual.py ├── utils │ ├── __init__.py │ ├── audio │ │ ├── __init__.py │ │ ├── numpy_transforms.py │ │ ├── processor.py │ │ └── torch_transforms.py │ ├── callbacks.py │ ├── capacitron_optimizer.py │ ├── distribute.py │ ├── download.py │ ├── downloaders.py │ ├── generic_utils.py │ ├── io.py │ ├── manage.py │ ├── radam.py │ ├── samplers.py │ ├── synthesizer.py │ ├── training.py │ └── vad.py └── vocoder │ ├── README.md │ ├── __init__.py │ ├── configs │ ├── __init__.py │ ├── fullband_melgan_config.py │ ├── hifigan_config.py │ ├── melgan_config.py │ ├── multiband_melgan_config.py │ ├── parallel_wavegan_config.py │ ├── shared_configs.py │ ├── univnet_config.py │ ├── wavegrad_config.py │ └── wavernn_config.py │ ├── datasets │ ├── __init__.py │ ├── gan_dataset.py │ ├── preprocess.py │ ├── wavegrad_dataset.py │ └── wavernn_dataset.py │ ├── layers │ ├── __init__.py │ ├── hifigan.py │ ├── losses.py │ ├── lvc_block.py │ ├── melgan.py │ ├── parallel_wavegan.py │ ├── pqmf.py │ ├── qmf.dat │ ├── upsample.py │ └── wavegrad.py │ ├── models │ ├── __init__.py │ ├── base_vocoder.py │ ├── fullband_melgan_generator.py │ ├── gan.py │ ├── hifigan_discriminator.py │ ├── hifigan_generator.py │ ├── melgan_discriminator.py │ ├── melgan_generator.py │ ├── melgan_multiscale_discriminator.py │ ├── multiband_melgan_generator.py │ ├── parallel_wavegan_discriminator.py │ ├── parallel_wavegan_generator.py │ ├── random_window_discriminator.py │ ├── univnet_discriminator.py │ ├── univnet_generator.py │ ├── wavegrad.py │ └── wavernn.py │ ├── pqmf_output.wav │ └── utils │ ├── __init__.py │ ├── distribution.py │ └── generic_utils.py ├── app.py ├── clonner_output └── sample.txt ├── helpers.py ├── language_model └── sample.txt ├── model_list.json ├── output └── sample.ttx ├── packages.txt ├── requirements.txt └── setup.py /.github/FUNDING.yml: -------------------------------------------------------------------------------- 1 | # These are supported funding model platforms 2 | 3 | github: everydaycodings 4 | patreon: # Replace with a single Patreon username 5 | open_collective: # Replace with a single Open Collective username 6 | ko_fi: # Replace with a single Ko-fi username 7 | tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel 8 | community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry 9 | liberapay: # Replace with a single Liberapay username 10 | issuehunt: # Replace with a single IssueHunt username 11 | otechie: # Replace with a single Otechie username 12 | lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry 13 | custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2'] 14 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | docs/ 2 | images/ 3 | notebooks/ 4 | language_model/*/ 5 | output/*.wav 6 | recipes/ 7 | tests/ 8 | test.ipynb 9 | extras/ 10 | .ipynb_checkpoints/ 11 | test_sound/ 12 | language_model_test/ 13 | downloader.ipynb 14 | __pycache__ 15 | temp.zip 16 | clonner_output/.wav -------------------------------------------------------------------------------- /.streamlit/config.toml: -------------------------------------------------------------------------------- 1 | [theme] 2 | base="dark" 3 | 4 | [server] 5 | maxUploadSize = 1028 6 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2023 Kumar Saksham 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # MimicMania 2 | 3 | MiniMania is a web application that provides text-to-speech (TTS) and voice cloning capabilities. The application utilizes deep learning models to generate high-quality speech output in multiple languages, with options to customize the pitch, speed, and volume of the generated audio. MiniMania's TTS engine is built on top of the Tacotron 2 and WaveGlow models, while its voice cloning capabilities are based on the FastSpeech and MelGAN models. 4 | 5 | ## Features 6 | 7 | - Easy-to-use interface: MimicMania has a user-friendly interface that allows users to quickly generate speech in their desired language and voice. 8 | - Multiple languages: MimicMania supports multiple languages, including English, Spanish, French, and more. 9 | - Multiple voices: MimicMania provides a range of voices for each language, giving users a wide variety of options to choose from. 10 | - Customizable parameters: Users can adjust the speed, pitch, and volume of the generated speech to fit their specific needs. 11 | - Voice cloning: MimicMania's voice cloning technology allows users to clone their own voice, making it easier than ever to create personalized voiceovers. 12 | 13 | 14 | ## WebApp Demo 15 | 16 | Click the Mega Link To see the Demo: [Link For The Video](https://mega.nz/file/5ShTiARB#aO4ecf518xnBnj1HKx98y4vw4ozQwSRFgKLifxFJO-E) 17 | 18 | 19 | ## Things to Be Downloaded 20 | 21 | Before you can install and use MimicMania, you will need to download and install the following: 22 | 23 | - **ffmpeg:** MimicMania requires the ffmpeg library for audio encoding and decoding. To install ffmpeg, execute the command `apt-get install ffmpeg` in your terminal. 24 | - **espeak-ng:** MimicMania uses the espeak-ng text-to-speech engine for generating speech in various languages. To install espeak-ng, execute the command `sudo apt-get install espeak-ng` in your terminal. 25 | - **espeak:** MimicMania uses the espeak-ng text-to-speech engine for generating speech in various languages. To install espeak-ng, execute the command `sudo apt-get install espeak` in your terminal. 26 | 27 | Additionally, you will also need to download the required module and Python dependencies as described in the next section. 28 | 29 | ## Installation 30 | 31 | To install and run MimicMania, follow these steps: 32 | 33 | 1. Clone this repository to your local machine using `git clone https://github.com/everydaycodings/MimicMania.git` 34 | 2. Navigate to the project directory using `cd MimicMania` 35 | 3. Run the command `python setup.py` to download the required modules. This process may take some time, as the required module is around 6 GB. 36 | 4. Run the command `pip install -r requirements.txt` to download all the Python dependencies. 37 | 5. Type the command `streamlit run app.py` to start the web application. 38 | 39 | 40 | 41 | ## Contributing 42 | 43 | If you'd like to contribute to MimicMania, please fork the repository and create a pull request. We welcome contributions of all kinds, including bug fixes, new features, and documentation improvements. 44 | 45 | ## Credits 46 | 47 | MimicMania was developed by **Kumar Saksham(everydaycodings)** with the help of various open source resources. 48 | 49 | We would like to extend a special thank you to [coqui-ai/TTS](https://github.com/coqui-ai/TTS) for providing their text-to-speech model as a resource for our project. 50 | 51 | MimicMania is licensed under the [MIT License](https://opensource.org/licenses/MIT). 52 | 53 | ## Support 54 | 55 | If you have any questions or issues with MimicMania, please contact us at [everydaycodings@gmail.com](mailto:everydaycodings@gmail.com) or reach out to us on Twitter [@everydaycodings](https://twitter.com/everydaycodings) or Medium [@everydaycodings](https://medium.com/@everydaycodings). 56 | 57 | We're always happy to help! 58 | -------------------------------------------------------------------------------- /TTS/VERSION: -------------------------------------------------------------------------------- 1 | 0.11.1 -------------------------------------------------------------------------------- /TTS/__init__.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | with open(os.path.join(os.path.dirname(__file__), "VERSION"), "r", encoding="utf-8") as f: 4 | version = f.read().strip() 5 | 6 | __version__ = version 7 | -------------------------------------------------------------------------------- /TTS/bin/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/bin/__init__.py -------------------------------------------------------------------------------- /TTS/bin/collect_env_info.py: -------------------------------------------------------------------------------- 1 | """Get detailed info about the working environment.""" 2 | import os 3 | import platform 4 | import sys 5 | 6 | import numpy 7 | import torch 8 | 9 | sys.path += [os.path.abspath(".."), os.path.abspath(".")] 10 | import json 11 | 12 | import TTS 13 | 14 | 15 | def system_info(): 16 | return { 17 | "OS": platform.system(), 18 | "architecture": platform.architecture(), 19 | "version": platform.version(), 20 | "processor": platform.processor(), 21 | "python": platform.python_version(), 22 | } 23 | 24 | 25 | def cuda_info(): 26 | return { 27 | "GPU": [torch.cuda.get_device_name(i) for i in range(torch.cuda.device_count())], 28 | "available": torch.cuda.is_available(), 29 | "version": torch.version.cuda, 30 | } 31 | 32 | 33 | def package_info(): 34 | return { 35 | "numpy": numpy.__version__, 36 | "PyTorch_version": torch.__version__, 37 | "PyTorch_debug": torch.version.debug, 38 | "TTS": TTS.__version__, 39 | } 40 | 41 | 42 | def main(): 43 | details = {"System": system_info(), "CUDA": cuda_info(), "Packages": package_info()} 44 | print(json.dumps(details, indent=4, sort_keys=True)) 45 | 46 | 47 | if __name__ == "__main__": 48 | main() 49 | -------------------------------------------------------------------------------- /TTS/bin/compute_statistics.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | 4 | import argparse 5 | import glob 6 | import os 7 | 8 | import numpy as np 9 | from tqdm import tqdm 10 | 11 | # from TTS.utils.io import load_config 12 | from TTS.config import load_config 13 | from TTS.tts.datasets import load_tts_samples 14 | from TTS.utils.audio import AudioProcessor 15 | 16 | 17 | def main(): 18 | """Run preprocessing process.""" 19 | parser = argparse.ArgumentParser(description="Compute mean and variance of spectrogtram features.") 20 | parser.add_argument("config_path", type=str, help="TTS config file path to define audio processin parameters.") 21 | parser.add_argument("out_path", type=str, help="save path (directory and filename).") 22 | parser.add_argument( 23 | "--data_path", 24 | type=str, 25 | required=False, 26 | help="folder including the target set of wavs overriding dataset config.", 27 | ) 28 | args, overrides = parser.parse_known_args() 29 | 30 | CONFIG = load_config(args.config_path) 31 | CONFIG.parse_known_args(overrides, relaxed_parser=True) 32 | 33 | # load config 34 | CONFIG.audio.signal_norm = False # do not apply earlier normalization 35 | CONFIG.audio.stats_path = None # discard pre-defined stats 36 | 37 | # load audio processor 38 | ap = AudioProcessor(**CONFIG.audio.to_dict()) 39 | 40 | # load the meta data of target dataset 41 | if args.data_path: 42 | dataset_items = glob.glob(os.path.join(args.data_path, "**", "*.wav"), recursive=True) 43 | else: 44 | dataset_items = load_tts_samples(CONFIG.datasets)[0] # take only train data 45 | print(f" > There are {len(dataset_items)} files.") 46 | 47 | mel_sum = 0 48 | mel_square_sum = 0 49 | linear_sum = 0 50 | linear_square_sum = 0 51 | N = 0 52 | for item in tqdm(dataset_items): 53 | # compute features 54 | wav = ap.load_wav(item if isinstance(item, str) else item["audio_file"]) 55 | linear = ap.spectrogram(wav) 56 | mel = ap.melspectrogram(wav) 57 | 58 | # compute stats 59 | N += mel.shape[1] 60 | mel_sum += mel.sum(1) 61 | linear_sum += linear.sum(1) 62 | mel_square_sum += (mel**2).sum(axis=1) 63 | linear_square_sum += (linear**2).sum(axis=1) 64 | 65 | mel_mean = mel_sum / N 66 | mel_scale = np.sqrt(mel_square_sum / N - mel_mean**2) 67 | linear_mean = linear_sum / N 68 | linear_scale = np.sqrt(linear_square_sum / N - linear_mean**2) 69 | 70 | output_file_path = args.out_path 71 | stats = {} 72 | stats["mel_mean"] = mel_mean 73 | stats["mel_std"] = mel_scale 74 | stats["linear_mean"] = linear_mean 75 | stats["linear_std"] = linear_scale 76 | 77 | print(f" > Avg mel spec mean: {mel_mean.mean()}") 78 | print(f" > Avg mel spec scale: {mel_scale.mean()}") 79 | print(f" > Avg linear spec mean: {linear_mean.mean()}") 80 | print(f" > Avg linear spec scale: {linear_scale.mean()}") 81 | 82 | # set default config values for mean-var scaling 83 | CONFIG.audio.stats_path = output_file_path 84 | CONFIG.audio.signal_norm = True 85 | # remove redundant values 86 | del CONFIG.audio.max_norm 87 | del CONFIG.audio.min_level_db 88 | del CONFIG.audio.symmetric_norm 89 | del CONFIG.audio.clip_norm 90 | stats["audio_config"] = CONFIG.audio.to_dict() 91 | np.save(output_file_path, stats, allow_pickle=True) 92 | print(f" > stats saved to {output_file_path}") 93 | 94 | 95 | if __name__ == "__main__": 96 | main() 97 | -------------------------------------------------------------------------------- /TTS/bin/eval_encoder.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from argparse import RawTextHelpFormatter 3 | 4 | import torch 5 | from tqdm import tqdm 6 | 7 | from TTS.config import load_config 8 | from TTS.tts.datasets import load_tts_samples 9 | from TTS.tts.utils.speakers import SpeakerManager 10 | 11 | 12 | def compute_encoder_accuracy(dataset_items, encoder_manager): 13 | class_name_key = encoder_manager.encoder_config.class_name_key 14 | map_classid_to_classname = getattr(encoder_manager.encoder_config, "map_classid_to_classname", None) 15 | 16 | class_acc_dict = {} 17 | 18 | # compute embeddings for all wav_files 19 | for item in tqdm(dataset_items): 20 | class_name = item[class_name_key] 21 | wav_file = item["audio_file"] 22 | 23 | # extract the embedding 24 | embedd = encoder_manager.compute_embedding_from_clip(wav_file) 25 | if encoder_manager.encoder_criterion is not None and map_classid_to_classname is not None: 26 | embedding = torch.FloatTensor(embedd).unsqueeze(0) 27 | if encoder_manager.use_cuda: 28 | embedding = embedding.cuda() 29 | 30 | class_id = encoder_manager.encoder_criterion.softmax.inference(embedding).item() 31 | predicted_label = map_classid_to_classname[str(class_id)] 32 | else: 33 | predicted_label = None 34 | 35 | if class_name is not None and predicted_label is not None: 36 | is_equal = int(class_name == predicted_label) 37 | if class_name not in class_acc_dict: 38 | class_acc_dict[class_name] = [is_equal] 39 | else: 40 | class_acc_dict[class_name].append(is_equal) 41 | else: 42 | raise RuntimeError("Error: class_name or/and predicted_label are None") 43 | 44 | acc_avg = 0 45 | for key, values in class_acc_dict.items(): 46 | acc = sum(values) / len(values) 47 | print("Class", key, "Accuracy:", acc) 48 | acc_avg += acc 49 | 50 | print("Average Accuracy:", acc_avg / len(class_acc_dict)) 51 | 52 | 53 | if __name__ == "__main__": 54 | parser = argparse.ArgumentParser( 55 | description="""Compute the accuracy of the encoder.\n\n""" 56 | """ 57 | Example runs: 58 | python TTS/bin/eval_encoder.py emotion_encoder_model.pth emotion_encoder_config.json dataset_config.json 59 | """, 60 | formatter_class=RawTextHelpFormatter, 61 | ) 62 | parser.add_argument("model_path", type=str, help="Path to model checkpoint file.") 63 | parser.add_argument( 64 | "config_path", 65 | type=str, 66 | help="Path to model config file.", 67 | ) 68 | 69 | parser.add_argument( 70 | "config_dataset_path", 71 | type=str, 72 | help="Path to dataset config file.", 73 | ) 74 | parser.add_argument("--use_cuda", type=bool, help="flag to set cuda.", default=True) 75 | parser.add_argument("--eval", type=bool, help="compute eval.", default=True) 76 | 77 | args = parser.parse_args() 78 | 79 | c_dataset = load_config(args.config_dataset_path) 80 | 81 | meta_data_train, meta_data_eval = load_tts_samples(c_dataset.datasets, eval_split=args.eval) 82 | items = meta_data_train + meta_data_eval 83 | 84 | enc_manager = SpeakerManager( 85 | encoder_model_path=args.model_path, encoder_config_path=args.config_path, use_cuda=args.use_cuda 86 | ) 87 | 88 | compute_encoder_accuracy(items, enc_manager) 89 | -------------------------------------------------------------------------------- /TTS/bin/find_unique_chars.py: -------------------------------------------------------------------------------- 1 | """Find all the unique characters in a dataset""" 2 | import argparse 3 | from argparse import RawTextHelpFormatter 4 | 5 | from TTS.config import load_config 6 | from TTS.tts.datasets import load_tts_samples 7 | 8 | 9 | def main(): 10 | # pylint: disable=bad-option-value 11 | parser = argparse.ArgumentParser( 12 | description="""Find all the unique characters or phonemes in a dataset.\n\n""" 13 | """ 14 | Example runs: 15 | 16 | python TTS/bin/find_unique_chars.py --config_path config.json 17 | """, 18 | formatter_class=RawTextHelpFormatter, 19 | ) 20 | parser.add_argument("--config_path", type=str, help="Path to dataset config file.", required=True) 21 | args = parser.parse_args() 22 | 23 | c = load_config(args.config_path) 24 | 25 | # load all datasets 26 | train_items, eval_items = load_tts_samples( 27 | c.datasets, eval_split=True, eval_split_max_size=c.eval_split_max_size, eval_split_size=c.eval_split_size 28 | ) 29 | 30 | items = train_items + eval_items 31 | 32 | texts = "".join(item["text"] for item in items) 33 | chars = set(texts) 34 | lower_chars = filter(lambda c: c.islower(), chars) 35 | chars_force_lower = [c.lower() for c in chars] 36 | chars_force_lower = set(chars_force_lower) 37 | 38 | print(f" > Number of unique characters: {len(chars)}") 39 | print(f" > Unique characters: {''.join(sorted(chars))}") 40 | print(f" > Unique lower characters: {''.join(sorted(lower_chars))}") 41 | print(f" > Unique all forced to lower characters: {''.join(sorted(chars_force_lower))}") 42 | 43 | 44 | if __name__ == "__main__": 45 | main() 46 | -------------------------------------------------------------------------------- /TTS/bin/find_unique_phonemes.py: -------------------------------------------------------------------------------- 1 | """Find all the unique characters in a dataset""" 2 | import argparse 3 | import multiprocessing 4 | from argparse import RawTextHelpFormatter 5 | 6 | from tqdm.contrib.concurrent import process_map 7 | 8 | from TTS.config import load_config 9 | from TTS.tts.datasets import load_tts_samples 10 | from TTS.tts.utils.text.phonemizers import Gruut 11 | 12 | 13 | def compute_phonemes(item): 14 | text = item["text"] 15 | ph = phonemizer.phonemize(text).replace("|", "") 16 | return set(list(ph)) 17 | 18 | 19 | def main(): 20 | # pylint: disable=W0601 21 | global c, phonemizer 22 | # pylint: disable=bad-option-value 23 | parser = argparse.ArgumentParser( 24 | description="""Find all the unique characters or phonemes in a dataset.\n\n""" 25 | """ 26 | Example runs: 27 | 28 | python TTS/bin/find_unique_phonemes.py --config_path config.json 29 | """, 30 | formatter_class=RawTextHelpFormatter, 31 | ) 32 | parser.add_argument("--config_path", type=str, help="Path to dataset config file.", required=True) 33 | args = parser.parse_args() 34 | 35 | c = load_config(args.config_path) 36 | 37 | # load all datasets 38 | train_items, eval_items = load_tts_samples( 39 | c.datasets, eval_split=True, eval_split_max_size=c.eval_split_max_size, eval_split_size=c.eval_split_size 40 | ) 41 | items = train_items + eval_items 42 | print("Num items:", len(items)) 43 | 44 | language_list = [item["language"] for item in items] 45 | is_lang_def = all(language_list) 46 | 47 | if not c.phoneme_language or not is_lang_def: 48 | raise ValueError("Phoneme language must be defined in config.") 49 | 50 | if not language_list.count(language_list[0]) == len(language_list): 51 | raise ValueError( 52 | "Currently, just one phoneme language per config file is supported !! Please split the dataset config into different configs and run it individually for each language !!" 53 | ) 54 | 55 | phonemizer = Gruut(language=language_list[0], keep_puncs=True) 56 | 57 | phonemes = process_map(compute_phonemes, items, max_workers=multiprocessing.cpu_count(), chunksize=15) 58 | phones = [] 59 | for ph in phonemes: 60 | phones.extend(ph) 61 | 62 | phones = set(phones) 63 | lower_phones = filter(lambda c: c.islower(), phones) 64 | phones_force_lower = [c.lower() for c in phones] 65 | phones_force_lower = set(phones_force_lower) 66 | 67 | print(f" > Number of unique phonemes: {len(phones)}") 68 | print(f" > Unique phonemes: {''.join(sorted(phones))}") 69 | print(f" > Unique lower phonemes: {''.join(sorted(lower_phones))}") 70 | print(f" > Unique all forced to lower phonemes: {''.join(sorted(phones_force_lower))}") 71 | 72 | 73 | if __name__ == "__main__": 74 | main() 75 | -------------------------------------------------------------------------------- /TTS/bin/remove_silence_using_vad.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import glob 3 | import os 4 | import pathlib 5 | 6 | from tqdm import tqdm 7 | 8 | from TTS.utils.vad import get_vad_model_and_utils, remove_silence 9 | 10 | 11 | def adjust_path_and_remove_silence(audio_path): 12 | output_path = audio_path.replace(os.path.join(args.input_dir, ""), os.path.join(args.output_dir, "")) 13 | # ignore if the file exists 14 | if os.path.exists(output_path) and not args.force: 15 | return output_path 16 | 17 | # create all directory structure 18 | pathlib.Path(output_path).parent.mkdir(parents=True, exist_ok=True) 19 | # remove the silence and save the audio 20 | output_path, is_speech = remove_silence( 21 | model_and_utils, 22 | audio_path, 23 | output_path, 24 | trim_just_beginning_and_end=args.trim_just_beginning_and_end, 25 | use_cuda=args.use_cuda, 26 | ) 27 | 28 | return output_path, is_speech 29 | 30 | 31 | def preprocess_audios(): 32 | files = sorted(glob.glob(os.path.join(args.input_dir, args.glob), recursive=True)) 33 | print("> Number of files: ", len(files)) 34 | if not args.force: 35 | print("> Ignoring files that already exist in the output idrectory.") 36 | 37 | if args.trim_just_beginning_and_end: 38 | print("> Trimming just the beginning and the end with nonspeech parts.") 39 | else: 40 | print("> Trimming all nonspeech parts.") 41 | 42 | filtered_files = [] 43 | if files: 44 | # create threads 45 | # num_threads = multiprocessing.cpu_count() 46 | # process_map(adjust_path_and_remove_silence, files, max_workers=num_threads, chunksize=15) 47 | for f in tqdm(files): 48 | output_path, is_speech = adjust_path_and_remove_silence(f) 49 | if not is_speech: 50 | filtered_files.append(output_path) 51 | 52 | # write files that do not have speech 53 | with open(os.path.join(args.output_dir, "filtered_files.txt"), "w", encoding="utf-8") as f: 54 | for file in filtered_files: 55 | f.write(file + "\n") 56 | else: 57 | print("> No files Found !") 58 | 59 | 60 | if __name__ == "__main__": 61 | parser = argparse.ArgumentParser( 62 | description="python TTS/bin/remove_silence_using_vad.py -i=VCTK-Corpus/ -o=VCTK-Corpus-removed-silence/ -g=wav48_silence_trimmed/*/*_mic1.flac --trim_just_beginning_and_end True" 63 | ) 64 | parser.add_argument("-i", "--input_dir", type=str, default="../VCTK-Corpus", help="Dataset root dir") 65 | parser.add_argument( 66 | "-o", "--output_dir", type=str, default="../VCTK-Corpus-removed-silence", help="Output Dataset dir" 67 | ) 68 | parser.add_argument("-f", "--force", default=False, action="store_true", help="Force the replace of exists files") 69 | parser.add_argument( 70 | "-g", 71 | "--glob", 72 | type=str, 73 | default="**/*.wav", 74 | help="path in glob format for acess wavs from input_dir. ex: wav48/*/*.wav", 75 | ) 76 | parser.add_argument( 77 | "-t", 78 | "--trim_just_beginning_and_end", 79 | type=bool, 80 | default=True, 81 | help="If True this script will trim just the beginning and end nonspeech parts. If False all nonspeech parts will be trim. Default True", 82 | ) 83 | parser.add_argument( 84 | "-c", 85 | "--use_cuda", 86 | type=bool, 87 | default=False, 88 | help="If True use cuda", 89 | ) 90 | args = parser.parse_args() 91 | # load the model and utils 92 | model_and_utils = get_vad_model_and_utils(use_cuda=args.use_cuda) 93 | preprocess_audios() 94 | -------------------------------------------------------------------------------- /TTS/bin/resample.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import glob 3 | import os 4 | from argparse import RawTextHelpFormatter 5 | from multiprocessing import Pool 6 | from shutil import copytree 7 | 8 | import librosa 9 | import soundfile as sf 10 | from tqdm import tqdm 11 | 12 | 13 | def resample_file(func_args): 14 | filename, output_sr = func_args 15 | y, sr = librosa.load(filename, sr=output_sr) 16 | sf.write(filename, y, sr) 17 | 18 | 19 | def resample_files(input_dir, output_sr, output_dir=None, file_ext="wav", n_jobs=10): 20 | if output_dir: 21 | print("Recursively copying the input folder...") 22 | copytree(input_dir, output_dir) 23 | input_dir = output_dir 24 | 25 | print("Resampling the audio files...") 26 | audio_files = glob.glob(os.path.join(input_dir, f"**/*.{file_ext}"), recursive=True) 27 | print(f"Found {len(audio_files)} files...") 28 | audio_files = list(zip(audio_files, len(audio_files) * [output_sr])) 29 | with Pool(processes=n_jobs) as p: 30 | with tqdm(total=len(audio_files)) as pbar: 31 | for _, _ in enumerate(p.imap_unordered(resample_file, audio_files)): 32 | pbar.update() 33 | 34 | print("Done !") 35 | 36 | 37 | if __name__ == "__main__": 38 | parser = argparse.ArgumentParser( 39 | description="""Resample a folder recusively with librosa 40 | Can be used in place or create a copy of the folder as an output.\n\n 41 | Example run: 42 | python TTS/bin/resample.py 43 | --input_dir /root/LJSpeech-1.1/ 44 | --output_sr 22050 45 | --output_dir /root/resampled_LJSpeech-1.1/ 46 | --file_ext wav 47 | --n_jobs 24 48 | """, 49 | formatter_class=RawTextHelpFormatter, 50 | ) 51 | 52 | parser.add_argument( 53 | "--input_dir", 54 | type=str, 55 | default=None, 56 | required=True, 57 | help="Path of the folder containing the audio files to resample", 58 | ) 59 | 60 | parser.add_argument( 61 | "--output_sr", 62 | type=int, 63 | default=22050, 64 | required=False, 65 | help="Samlple rate to which the audio files should be resampled", 66 | ) 67 | 68 | parser.add_argument( 69 | "--output_dir", 70 | type=str, 71 | default=None, 72 | required=False, 73 | help="Path of the destination folder. If not defined, the operation is done in place", 74 | ) 75 | 76 | parser.add_argument( 77 | "--file_ext", 78 | type=str, 79 | default="wav", 80 | required=False, 81 | help="Extension of the audio files to resample", 82 | ) 83 | 84 | parser.add_argument( 85 | "--n_jobs", type=int, default=None, help="Number of threads to use, by default it uses all cores" 86 | ) 87 | 88 | args = parser.parse_args() 89 | 90 | resample_files(args.input_dir, args.output_sr, args.output_dir, args.file_ext, args.n_jobs) 91 | -------------------------------------------------------------------------------- /TTS/bin/train_tts.py: -------------------------------------------------------------------------------- 1 | import os 2 | from dataclasses import dataclass, field 3 | 4 | from trainer import Trainer, TrainerArgs 5 | 6 | from TTS.config import load_config, register_config 7 | from TTS.tts.datasets import load_tts_samples 8 | from TTS.tts.models import setup_model 9 | 10 | 11 | @dataclass 12 | class TrainTTSArgs(TrainerArgs): 13 | config_path: str = field(default=None, metadata={"help": "Path to the config file."}) 14 | 15 | 16 | def main(): 17 | """Run `tts` model training directly by a `config.json` file.""" 18 | # init trainer args 19 | train_args = TrainTTSArgs() 20 | parser = train_args.init_argparse(arg_prefix="") 21 | 22 | # override trainer args from comman-line args 23 | args, config_overrides = parser.parse_known_args() 24 | train_args.parse_args(args) 25 | 26 | # load config.json and register 27 | if args.config_path or args.continue_path: 28 | if args.config_path: 29 | # init from a file 30 | config = load_config(args.config_path) 31 | if len(config_overrides) > 0: 32 | config.parse_known_args(config_overrides, relaxed_parser=True) 33 | elif args.continue_path: 34 | # continue from a prev experiment 35 | config = load_config(os.path.join(args.continue_path, "config.json")) 36 | if len(config_overrides) > 0: 37 | config.parse_known_args(config_overrides, relaxed_parser=True) 38 | else: 39 | # init from console args 40 | from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel 41 | 42 | config_base = BaseTrainingConfig() 43 | config_base.parse_known_args(config_overrides) 44 | config = register_config(config_base.model)() 45 | 46 | # load training samples 47 | train_samples, eval_samples = load_tts_samples( 48 | config.datasets, 49 | eval_split=True, 50 | eval_split_max_size=config.eval_split_max_size, 51 | eval_split_size=config.eval_split_size, 52 | ) 53 | 54 | # init the model from config 55 | model = setup_model(config, train_samples + eval_samples) 56 | 57 | # init the trainer and 🚀 58 | trainer = Trainer( 59 | train_args, 60 | model.config, 61 | config.output_path, 62 | model=model, 63 | train_samples=train_samples, 64 | eval_samples=eval_samples, 65 | parse_command_line_args=False, 66 | ) 67 | trainer.fit() 68 | 69 | 70 | if __name__ == "__main__": 71 | main() 72 | -------------------------------------------------------------------------------- /TTS/bin/train_vocoder.py: -------------------------------------------------------------------------------- 1 | import os 2 | from dataclasses import dataclass, field 3 | 4 | from trainer import Trainer, TrainerArgs 5 | 6 | from TTS.config import load_config, register_config 7 | from TTS.utils.audio import AudioProcessor 8 | from TTS.vocoder.datasets.preprocess import load_wav_data, load_wav_feat_data 9 | from TTS.vocoder.models import setup_model 10 | 11 | 12 | @dataclass 13 | class TrainVocoderArgs(TrainerArgs): 14 | config_path: str = field(default=None, metadata={"help": "Path to the config file."}) 15 | 16 | 17 | def main(): 18 | """Run `tts` model training directly by a `config.json` file.""" 19 | # init trainer args 20 | train_args = TrainVocoderArgs() 21 | parser = train_args.init_argparse(arg_prefix="") 22 | 23 | # override trainer args from comman-line args 24 | args, config_overrides = parser.parse_known_args() 25 | train_args.parse_args(args) 26 | 27 | # load config.json and register 28 | if args.config_path or args.continue_path: 29 | if args.config_path: 30 | # init from a file 31 | config = load_config(args.config_path) 32 | if len(config_overrides) > 0: 33 | config.parse_known_args(config_overrides, relaxed_parser=True) 34 | elif args.continue_path: 35 | # continue from a prev experiment 36 | config = load_config(os.path.join(args.continue_path, "config.json")) 37 | if len(config_overrides) > 0: 38 | config.parse_known_args(config_overrides, relaxed_parser=True) 39 | else: 40 | # init from console args 41 | from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel 42 | 43 | config_base = BaseTrainingConfig() 44 | config_base.parse_known_args(config_overrides) 45 | config = register_config(config_base.model)() 46 | 47 | # load training samples 48 | if "feature_path" in config and config.feature_path: 49 | # load pre-computed features 50 | print(f" > Loading features from: {config.feature_path}") 51 | eval_samples, train_samples = load_wav_feat_data(config.data_path, config.feature_path, config.eval_split_size) 52 | else: 53 | # load data raw wav files 54 | eval_samples, train_samples = load_wav_data(config.data_path, config.eval_split_size) 55 | 56 | # setup audio processor 57 | ap = AudioProcessor(**config.audio) 58 | 59 | # init the model from config 60 | model = setup_model(config) 61 | 62 | # init the trainer and 🚀 63 | trainer = Trainer( 64 | train_args, 65 | config, 66 | config.output_path, 67 | model=model, 68 | train_samples=train_samples, 69 | eval_samples=eval_samples, 70 | training_assets={"audio_processor": ap}, 71 | parse_command_line_args=False, 72 | ) 73 | trainer.fit() 74 | 75 | 76 | if __name__ == "__main__": 77 | main() 78 | -------------------------------------------------------------------------------- /TTS/bin/tune_wavegrad.py: -------------------------------------------------------------------------------- 1 | """Search a good noise schedule for WaveGrad for a given number of inference iterations""" 2 | import argparse 3 | from itertools import product as cartesian_product 4 | 5 | import numpy as np 6 | import torch 7 | from torch.utils.data import DataLoader 8 | from tqdm import tqdm 9 | 10 | from TTS.config import load_config 11 | from TTS.utils.audio import AudioProcessor 12 | from TTS.vocoder.datasets.preprocess import load_wav_data 13 | from TTS.vocoder.datasets.wavegrad_dataset import WaveGradDataset 14 | from TTS.vocoder.models import setup_model 15 | 16 | if __name__ == "__main__": 17 | parser = argparse.ArgumentParser() 18 | parser.add_argument("--model_path", type=str, help="Path to model checkpoint.") 19 | parser.add_argument("--config_path", type=str, help="Path to model config file.") 20 | parser.add_argument("--data_path", type=str, help="Path to data directory.") 21 | parser.add_argument("--output_path", type=str, help="path for output file including file name and extension.") 22 | parser.add_argument( 23 | "--num_iter", 24 | type=int, 25 | help="Number of model inference iterations that you like to optimize noise schedule for.", 26 | ) 27 | parser.add_argument("--use_cuda", action="store_true", help="enable CUDA.") 28 | parser.add_argument("--num_samples", type=int, default=1, help="Number of datasamples used for inference.") 29 | parser.add_argument( 30 | "--search_depth", 31 | type=int, 32 | default=3, 33 | help="Search granularity. Increasing this increases the run-time exponentially.", 34 | ) 35 | 36 | # load config 37 | args = parser.parse_args() 38 | config = load_config(args.config_path) 39 | 40 | # setup audio processor 41 | ap = AudioProcessor(**config.audio) 42 | 43 | # load dataset 44 | _, train_data = load_wav_data(args.data_path, 0) 45 | train_data = train_data[: args.num_samples] 46 | dataset = WaveGradDataset( 47 | ap=ap, 48 | items=train_data, 49 | seq_len=-1, 50 | hop_len=ap.hop_length, 51 | pad_short=config.pad_short, 52 | conv_pad=config.conv_pad, 53 | is_training=True, 54 | return_segments=False, 55 | use_noise_augment=False, 56 | use_cache=False, 57 | verbose=True, 58 | ) 59 | loader = DataLoader( 60 | dataset, 61 | batch_size=1, 62 | shuffle=False, 63 | collate_fn=dataset.collate_full_clips, 64 | drop_last=False, 65 | num_workers=config.num_loader_workers, 66 | pin_memory=False, 67 | ) 68 | 69 | # setup the model 70 | model = setup_model(config) 71 | if args.use_cuda: 72 | model.cuda() 73 | 74 | # setup optimization parameters 75 | base_values = sorted(10 * np.random.uniform(size=args.search_depth)) 76 | print(f" > base values: {base_values}") 77 | exponents = 10 ** np.linspace(-6, -1, num=args.num_iter) 78 | best_error = float("inf") 79 | best_schedule = None # pylint: disable=C0103 80 | total_search_iter = len(base_values) ** args.num_iter 81 | for base in tqdm(cartesian_product(base_values, repeat=args.num_iter), total=total_search_iter): 82 | beta = exponents * base 83 | model.compute_noise_level(beta) 84 | for data in loader: 85 | mel, audio = data 86 | y_hat = model.inference(mel.cuda() if args.use_cuda else mel) 87 | 88 | if args.use_cuda: 89 | y_hat = y_hat.cpu() 90 | y_hat = y_hat.numpy() 91 | 92 | mel_hat = [] 93 | for i in range(y_hat.shape[0]): 94 | m = ap.melspectrogram(y_hat[i, 0])[:, :-1] 95 | mel_hat.append(torch.from_numpy(m)) 96 | 97 | mel_hat = torch.stack(mel_hat) 98 | mse = torch.sum((mel - mel_hat) ** 2).mean() 99 | if mse.item() < best_error: 100 | best_error = mse.item() 101 | best_schedule = {"beta": beta} 102 | print(f" > Found a better schedule. - MSE: {mse.item()}") 103 | np.save(args.output_path, best_schedule) 104 | -------------------------------------------------------------------------------- /TTS/config/__init__.py: -------------------------------------------------------------------------------- 1 | import json 2 | import os 3 | import re 4 | from typing import Dict 5 | 6 | import fsspec 7 | import yaml 8 | from coqpit import Coqpit 9 | 10 | from TTS.config.shared_configs import * 11 | from TTS.utils.generic_utils import find_module 12 | 13 | 14 | def read_json_with_comments(json_path): 15 | """for backward compat.""" 16 | # fallback to json 17 | with fsspec.open(json_path, "r", encoding="utf-8") as f: 18 | input_str = f.read() 19 | # handle comments 20 | input_str = re.sub(r"\\\n", "", input_str) 21 | input_str = re.sub(r"//.*\n", "\n", input_str) 22 | data = json.loads(input_str) 23 | return data 24 | 25 | 26 | def register_config(model_name: str) -> Coqpit: 27 | """Find the right config for the given model name. 28 | 29 | Args: 30 | model_name (str): Model name. 31 | 32 | Raises: 33 | ModuleNotFoundError: No matching config for the model name. 34 | 35 | Returns: 36 | Coqpit: config class. 37 | """ 38 | config_class = None 39 | config_name = model_name + "_config" 40 | paths = ["TTS.tts.configs", "TTS.vocoder.configs", "TTS.encoder.configs"] 41 | for path in paths: 42 | try: 43 | config_class = find_module(path, config_name) 44 | except ModuleNotFoundError: 45 | pass 46 | if config_class is None: 47 | raise ModuleNotFoundError(f" [!] Config for {model_name} cannot be found.") 48 | return config_class 49 | 50 | 51 | def _process_model_name(config_dict: Dict) -> str: 52 | """Format the model name as expected. It is a band-aid for the old `vocoder` model names. 53 | 54 | Args: 55 | config_dict (Dict): A dictionary including the config fields. 56 | 57 | Returns: 58 | str: Formatted modelname. 59 | """ 60 | model_name = config_dict["model"] if "model" in config_dict else config_dict["generator_model"] 61 | model_name = model_name.replace("_generator", "").replace("_discriminator", "") 62 | return model_name 63 | 64 | 65 | def load_config(config_path: str) -> Coqpit: 66 | """Import `json` or `yaml` files as TTS configs. First, load the input file as a `dict` and check the model name 67 | to find the corresponding Config class. Then initialize the Config. 68 | 69 | Args: 70 | config_path (str): path to the config file. 71 | 72 | Raises: 73 | TypeError: given config file has an unknown type. 74 | 75 | Returns: 76 | Coqpit: TTS config object. 77 | """ 78 | config_dict = {} 79 | ext = os.path.splitext(config_path)[1] 80 | if ext in (".yml", ".yaml"): 81 | with fsspec.open(config_path, "r", encoding="utf-8") as f: 82 | data = yaml.safe_load(f) 83 | elif ext == ".json": 84 | try: 85 | with fsspec.open(config_path, "r", encoding="utf-8") as f: 86 | data = json.load(f) 87 | except json.decoder.JSONDecodeError: 88 | # backwards compat. 89 | data = read_json_with_comments(config_path) 90 | else: 91 | raise TypeError(f" [!] Unknown config file type {ext}") 92 | config_dict.update(data) 93 | model_name = _process_model_name(config_dict) 94 | config_class = register_config(model_name.lower()) 95 | config = config_class() 96 | config.from_dict(config_dict) 97 | return config 98 | 99 | 100 | def check_config_and_model_args(config, arg_name, value): 101 | """Check the give argument in `config.model_args` if exist or in `config` for 102 | the given value. 103 | 104 | Return False if the argument does not exist in `config.model_args` or `config`. 105 | This is to patch up the compatibility between models with and without `model_args`. 106 | 107 | TODO: Remove this in the future with a unified approach. 108 | """ 109 | if hasattr(config, "model_args"): 110 | if arg_name in config.model_args: 111 | return config.model_args[arg_name] == value 112 | if hasattr(config, arg_name): 113 | return config[arg_name] == value 114 | return False 115 | 116 | 117 | def get_from_config_or_model_args(config, arg_name): 118 | """Get the given argument from `config.model_args` if exist or in `config`.""" 119 | if hasattr(config, "model_args"): 120 | if arg_name in config.model_args: 121 | return config.model_args[arg_name] 122 | return config[arg_name] 123 | 124 | 125 | def get_from_config_or_model_args_with_default(config, arg_name, def_val): 126 | """Get the given argument from `config.model_args` if exist or in `config`.""" 127 | if hasattr(config, "model_args"): 128 | if arg_name in config.model_args: 129 | return config.model_args[arg_name] 130 | if hasattr(config, arg_name): 131 | return config[arg_name] 132 | return def_val 133 | -------------------------------------------------------------------------------- /TTS/encoder/README.md: -------------------------------------------------------------------------------- 1 | ### Speaker Encoder 2 | 3 | This is an implementation of https://arxiv.org/abs/1710.10467. This model can be used for voice and speaker embedding. 4 | 5 | With the code here you can generate d-vectors for both multi-speaker and single-speaker TTS datasets, then visualise and explore them along with the associated audio files in an interactive chart. 6 | 7 | Below is an example showing embedding results of various speakers. You can generate the same plot with the provided notebook as demonstrated in [this video](https://youtu.be/KW3oO7JVa7Q). 8 | 9 | ![](umap.png) 10 | 11 | Download a pretrained model from [Released Models](https://github.com/mozilla/TTS/wiki/Released-Models) page. 12 | 13 | To run the code, you need to follow the same flow as in TTS. 14 | 15 | - Define 'config.json' for your needs. Note that, audio parameters should match your TTS model. 16 | - Example training call ```python speaker_encoder/train.py --config_path speaker_encoder/config.json --data_path ~/Data/Libri-TTS/train-clean-360``` 17 | - Generate embedding vectors ```python speaker_encoder/compute_embeddings.py --use_cuda true /model/path/best_model.pth model/config/path/config.json dataset/path/ output_path``` . This code parses all .wav files at the given dataset path and generates the same folder structure under the output path with the generated embedding files. 18 | - Watch training on Tensorboard as in TTS 19 | -------------------------------------------------------------------------------- /TTS/encoder/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/encoder/__init__.py -------------------------------------------------------------------------------- /TTS/encoder/configs/base_encoder_config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import asdict, dataclass, field 2 | from typing import Dict, List 3 | 4 | from coqpit import MISSING 5 | 6 | from TTS.config.shared_configs import BaseAudioConfig, BaseDatasetConfig, BaseTrainingConfig 7 | 8 | 9 | @dataclass 10 | class BaseEncoderConfig(BaseTrainingConfig): 11 | """Defines parameters for a Generic Encoder model.""" 12 | 13 | model: str = None 14 | audio: BaseAudioConfig = field(default_factory=BaseAudioConfig) 15 | datasets: List[BaseDatasetConfig] = field(default_factory=lambda: [BaseDatasetConfig()]) 16 | # model params 17 | model_params: Dict = field( 18 | default_factory=lambda: { 19 | "model_name": "lstm", 20 | "input_dim": 80, 21 | "proj_dim": 256, 22 | "lstm_dim": 768, 23 | "num_lstm_layers": 3, 24 | "use_lstm_with_projection": True, 25 | } 26 | ) 27 | 28 | audio_augmentation: Dict = field(default_factory=lambda: {}) 29 | 30 | # training params 31 | epochs: int = 10000 32 | loss: str = "angleproto" 33 | grad_clip: float = 3.0 34 | lr: float = 0.0001 35 | optimizer: str = "radam" 36 | optimizer_params: Dict = field(default_factory=lambda: {"betas": [0.9, 0.999], "weight_decay": 0}) 37 | lr_decay: bool = False 38 | warmup_steps: int = 4000 39 | 40 | # logging params 41 | tb_model_param_stats: bool = False 42 | steps_plot_stats: int = 10 43 | save_step: int = 1000 44 | print_step: int = 20 45 | run_eval: bool = False 46 | 47 | # data loader 48 | num_classes_in_batch: int = MISSING 49 | num_utter_per_class: int = MISSING 50 | eval_num_classes_in_batch: int = None 51 | eval_num_utter_per_class: int = None 52 | 53 | num_loader_workers: int = MISSING 54 | voice_len: float = 1.6 55 | 56 | def check_values(self): 57 | super().check_values() 58 | c = asdict(self) 59 | assert ( 60 | c["model_params"]["input_dim"] == self.audio.num_mels 61 | ), " [!] model input dimendion must be equal to melspectrogram dimension." 62 | -------------------------------------------------------------------------------- /TTS/encoder/configs/emotion_encoder_config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import asdict, dataclass 2 | 3 | from TTS.encoder.configs.base_encoder_config import BaseEncoderConfig 4 | 5 | 6 | @dataclass 7 | class EmotionEncoderConfig(BaseEncoderConfig): 8 | """Defines parameters for Emotion Encoder model.""" 9 | 10 | model: str = "emotion_encoder" 11 | map_classid_to_classname: dict = None 12 | class_name_key: str = "emotion_name" 13 | -------------------------------------------------------------------------------- /TTS/encoder/configs/speaker_encoder_config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import asdict, dataclass 2 | 3 | from TTS.encoder.configs.base_encoder_config import BaseEncoderConfig 4 | 5 | 6 | @dataclass 7 | class SpeakerEncoderConfig(BaseEncoderConfig): 8 | """Defines parameters for Speaker Encoder model.""" 9 | 10 | model: str = "speaker_encoder" 11 | class_name_key: str = "speaker_name" 12 | -------------------------------------------------------------------------------- /TTS/encoder/models/lstm.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | from TTS.encoder.models.base_encoder import BaseEncoder 5 | 6 | 7 | class LSTMWithProjection(nn.Module): 8 | def __init__(self, input_size, hidden_size, proj_size): 9 | super().__init__() 10 | self.input_size = input_size 11 | self.hidden_size = hidden_size 12 | self.proj_size = proj_size 13 | self.lstm = nn.LSTM(input_size, hidden_size, batch_first=True) 14 | self.linear = nn.Linear(hidden_size, proj_size, bias=False) 15 | 16 | def forward(self, x): 17 | self.lstm.flatten_parameters() 18 | o, (_, _) = self.lstm(x) 19 | return self.linear(o) 20 | 21 | 22 | class LSTMWithoutProjection(nn.Module): 23 | def __init__(self, input_dim, lstm_dim, proj_dim, num_lstm_layers): 24 | super().__init__() 25 | self.lstm = nn.LSTM(input_size=input_dim, hidden_size=lstm_dim, num_layers=num_lstm_layers, batch_first=True) 26 | self.linear = nn.Linear(lstm_dim, proj_dim, bias=True) 27 | self.relu = nn.ReLU() 28 | 29 | def forward(self, x): 30 | _, (hidden, _) = self.lstm(x) 31 | return self.relu(self.linear(hidden[-1])) 32 | 33 | 34 | class LSTMSpeakerEncoder(BaseEncoder): 35 | def __init__( 36 | self, 37 | input_dim, 38 | proj_dim=256, 39 | lstm_dim=768, 40 | num_lstm_layers=3, 41 | use_lstm_with_projection=True, 42 | use_torch_spec=False, 43 | audio_config=None, 44 | ): 45 | super().__init__() 46 | self.use_lstm_with_projection = use_lstm_with_projection 47 | self.use_torch_spec = use_torch_spec 48 | self.audio_config = audio_config 49 | self.proj_dim = proj_dim 50 | 51 | layers = [] 52 | # choise LSTM layer 53 | if use_lstm_with_projection: 54 | layers.append(LSTMWithProjection(input_dim, lstm_dim, proj_dim)) 55 | for _ in range(num_lstm_layers - 1): 56 | layers.append(LSTMWithProjection(proj_dim, lstm_dim, proj_dim)) 57 | self.layers = nn.Sequential(*layers) 58 | else: 59 | self.layers = LSTMWithoutProjection(input_dim, lstm_dim, proj_dim, num_lstm_layers) 60 | 61 | self.instancenorm = nn.InstanceNorm1d(input_dim) 62 | 63 | if self.use_torch_spec: 64 | self.torch_spec = self.get_torch_mel_spectrogram_class(audio_config) 65 | else: 66 | self.torch_spec = None 67 | 68 | self._init_layers() 69 | 70 | def _init_layers(self): 71 | for name, param in self.layers.named_parameters(): 72 | if "bias" in name: 73 | nn.init.constant_(param, 0.0) 74 | elif "weight" in name: 75 | nn.init.xavier_normal_(param) 76 | 77 | def forward(self, x, l2_norm=True): 78 | """Forward pass of the model. 79 | 80 | Args: 81 | x (Tensor): Raw waveform signal or spectrogram frames. If input is a waveform, `torch_spec` must be `True` 82 | to compute the spectrogram on-the-fly. 83 | l2_norm (bool): Whether to L2-normalize the outputs. 84 | 85 | Shapes: 86 | - x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})` 87 | """ 88 | with torch.no_grad(): 89 | with torch.cuda.amp.autocast(enabled=False): 90 | if self.use_torch_spec: 91 | x.squeeze_(1) 92 | x = self.torch_spec(x) 93 | x = self.instancenorm(x).transpose(1, 2) 94 | d = self.layers(x) 95 | if self.use_lstm_with_projection: 96 | d = d[:, -1] 97 | if l2_norm: 98 | d = torch.nn.functional.normalize(d, p=2, dim=1) 99 | return d 100 | -------------------------------------------------------------------------------- /TTS/encoder/requirements.txt: -------------------------------------------------------------------------------- 1 | umap-learn 2 | numpy>=1.17.0 3 | -------------------------------------------------------------------------------- /TTS/encoder/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/encoder/utils/__init__.py -------------------------------------------------------------------------------- /TTS/encoder/utils/io.py: -------------------------------------------------------------------------------- 1 | import datetime 2 | import os 3 | 4 | from TTS.utils.io import save_fsspec 5 | 6 | 7 | def save_checkpoint(model, optimizer, model_loss, out_path, current_step): 8 | checkpoint_path = "checkpoint_{}.pth".format(current_step) 9 | checkpoint_path = os.path.join(out_path, checkpoint_path) 10 | print(" | | > Checkpoint saving : {}".format(checkpoint_path)) 11 | 12 | new_state_dict = model.state_dict() 13 | state = { 14 | "model": new_state_dict, 15 | "optimizer": optimizer.state_dict() if optimizer is not None else None, 16 | "step": current_step, 17 | "loss": model_loss, 18 | "date": datetime.date.today().strftime("%B %d, %Y"), 19 | } 20 | save_fsspec(state, checkpoint_path) 21 | 22 | 23 | def save_best_model(model, optimizer, model_loss, best_loss, out_path, current_step): 24 | if model_loss < best_loss: 25 | new_state_dict = model.state_dict() 26 | state = { 27 | "model": new_state_dict, 28 | "optimizer": optimizer.state_dict(), 29 | "step": current_step, 30 | "loss": model_loss, 31 | "date": datetime.date.today().strftime("%B %d, %Y"), 32 | } 33 | best_loss = model_loss 34 | bestmodel_path = "best_model.pth" 35 | bestmodel_path = os.path.join(out_path, bestmodel_path) 36 | print("\n > BEST MODEL ({0:.5f}) : {1:}".format(model_loss, bestmodel_path)) 37 | save_fsspec(state, bestmodel_path) 38 | return best_loss 39 | -------------------------------------------------------------------------------- /TTS/encoder/utils/training.py: -------------------------------------------------------------------------------- 1 | import os 2 | from dataclasses import dataclass, field 3 | 4 | from coqpit import Coqpit 5 | from trainer import TrainerArgs, get_last_checkpoint 6 | from trainer.logging import logger_factory 7 | from trainer.logging.console_logger import ConsoleLogger 8 | 9 | from TTS.config import load_config, register_config 10 | from TTS.tts.utils.text.characters import parse_symbols 11 | from TTS.utils.generic_utils import get_experiment_folder_path, get_git_branch 12 | from TTS.utils.io import copy_model_files 13 | 14 | 15 | @dataclass 16 | class TrainArgs(TrainerArgs): 17 | config_path: str = field(default=None, metadata={"help": "Path to the config file."}) 18 | 19 | 20 | def getarguments(): 21 | train_config = TrainArgs() 22 | parser = train_config.init_argparse(arg_prefix="") 23 | return parser 24 | 25 | 26 | def process_args(args, config=None): 27 | """Process parsed comand line arguments and initialize the config if not provided. 28 | Args: 29 | args (argparse.Namespace or dict like): Parsed input arguments. 30 | config (Coqpit): Model config. If none, it is generated from `args`. Defaults to None. 31 | Returns: 32 | c (TTS.utils.io.AttrDict): Config paramaters. 33 | out_path (str): Path to save models and logging. 34 | audio_path (str): Path to save generated test audios. 35 | c_logger (TTS.utils.console_logger.ConsoleLogger): Class that does 36 | logging to the console. 37 | dashboard_logger (WandbLogger or TensorboardLogger): Class that does the dashboard Logging 38 | TODO: 39 | - Interactive config definition. 40 | """ 41 | if isinstance(args, tuple): 42 | args, coqpit_overrides = args 43 | if args.continue_path: 44 | # continue a previous training from its output folder 45 | experiment_path = args.continue_path 46 | args.config_path = os.path.join(args.continue_path, "config.json") 47 | args.restore_path, best_model = get_last_checkpoint(args.continue_path) 48 | if not args.best_path: 49 | args.best_path = best_model 50 | # init config if not already defined 51 | if config is None: 52 | if args.config_path: 53 | # init from a file 54 | config = load_config(args.config_path) 55 | else: 56 | # init from console args 57 | from TTS.config.shared_configs import BaseTrainingConfig # pylint: disable=import-outside-toplevel 58 | 59 | config_base = BaseTrainingConfig() 60 | config_base.parse_known_args(coqpit_overrides) 61 | config = register_config(config_base.model)() 62 | # override values from command-line args 63 | config.parse_known_args(coqpit_overrides, relaxed_parser=True) 64 | experiment_path = args.continue_path 65 | if not experiment_path: 66 | experiment_path = get_experiment_folder_path(config.output_path, config.run_name) 67 | audio_path = os.path.join(experiment_path, "test_audios") 68 | config.output_log_path = experiment_path 69 | # setup rank 0 process in distributed training 70 | dashboard_logger = None 71 | if args.rank == 0: 72 | new_fields = {} 73 | if args.restore_path: 74 | new_fields["restore_path"] = args.restore_path 75 | new_fields["github_branch"] = get_git_branch() 76 | # if model characters are not set in the config file 77 | # save the default set to the config file for future 78 | # compatibility. 79 | if config.has("characters") and config.characters is None: 80 | used_characters = parse_symbols() 81 | new_fields["characters"] = used_characters 82 | copy_model_files(config, experiment_path, new_fields) 83 | dashboard_logger = logger_factory(config, experiment_path) 84 | c_logger = ConsoleLogger() 85 | return config, experiment_path, audio_path, c_logger, dashboard_logger 86 | 87 | 88 | def init_arguments(): 89 | train_config = TrainArgs() 90 | parser = train_config.init_argparse(arg_prefix="") 91 | return parser 92 | 93 | 94 | def init_training(config: Coqpit = None): 95 | """Initialization of a training run.""" 96 | parser = init_arguments() 97 | args = parser.parse_known_args() 98 | config, OUT_PATH, AUDIO_PATH, c_logger, dashboard_logger = process_args(args, config) 99 | return args[0], config, OUT_PATH, AUDIO_PATH, c_logger, dashboard_logger 100 | -------------------------------------------------------------------------------- /TTS/encoder/utils/visual.py: -------------------------------------------------------------------------------- 1 | import matplotlib 2 | import matplotlib.pyplot as plt 3 | import numpy as np 4 | import umap 5 | 6 | matplotlib.use("Agg") 7 | 8 | 9 | colormap = ( 10 | np.array( 11 | [ 12 | [76, 255, 0], 13 | [0, 127, 70], 14 | [255, 0, 0], 15 | [255, 217, 38], 16 | [0, 135, 255], 17 | [165, 0, 165], 18 | [255, 167, 255], 19 | [0, 255, 255], 20 | [255, 96, 38], 21 | [142, 76, 0], 22 | [33, 0, 127], 23 | [0, 0, 0], 24 | [183, 183, 183], 25 | ], 26 | dtype=np.float, 27 | ) 28 | / 255 29 | ) 30 | 31 | 32 | def plot_embeddings(embeddings, num_classes_in_batch): 33 | num_utter_per_class = embeddings.shape[0] // num_classes_in_batch 34 | 35 | # if necessary get just the first 10 classes 36 | if num_classes_in_batch > 10: 37 | num_classes_in_batch = 10 38 | embeddings = embeddings[: num_classes_in_batch * num_utter_per_class] 39 | 40 | model = umap.UMAP() 41 | projection = model.fit_transform(embeddings) 42 | ground_truth = np.repeat(np.arange(num_classes_in_batch), num_utter_per_class) 43 | colors = [colormap[i] for i in ground_truth] 44 | fig, ax = plt.subplots(figsize=(16, 10)) 45 | _ = ax.scatter(projection[:, 0], projection[:, 1], c=colors) 46 | plt.gca().set_aspect("equal", "datalim") 47 | plt.title("UMAP projection") 48 | plt.tight_layout() 49 | plt.savefig("umap") 50 | return fig 51 | -------------------------------------------------------------------------------- /TTS/model.py: -------------------------------------------------------------------------------- 1 | from abc import abstractmethod 2 | from typing import Dict 3 | 4 | import torch 5 | from coqpit import Coqpit 6 | from trainer import TrainerModel 7 | 8 | # pylint: skip-file 9 | 10 | 11 | class BaseTrainerModel(TrainerModel): 12 | """BaseTrainerModel model expanding TrainerModel with required functions by 🐸TTS. 13 | 14 | Every new 🐸TTS model must inherit it. 15 | """ 16 | 17 | @staticmethod 18 | @abstractmethod 19 | def init_from_config(config: Coqpit): 20 | """Init the model and all its attributes from the given config. 21 | 22 | Override this depending on your model. 23 | """ 24 | ... 25 | 26 | @abstractmethod 27 | def inference(self, input: torch.Tensor, aux_input={}) -> Dict: 28 | """Forward pass for inference. 29 | 30 | It must return a dictionary with the main model output and all the auxiliary outputs. The key ```model_outputs``` 31 | is considered to be the main output and you can add any other auxiliary outputs as you want. 32 | 33 | We don't use `*kwargs` since it is problematic with the TorchScript API. 34 | 35 | Args: 36 | input (torch.Tensor): [description] 37 | aux_input (Dict): Auxiliary inputs like speaker embeddings, durations etc. 38 | 39 | Returns: 40 | Dict: [description] 41 | """ 42 | outputs_dict = {"model_outputs": None} 43 | ... 44 | return outputs_dict 45 | 46 | @abstractmethod 47 | def load_checkpoint( 48 | self, config: Coqpit, checkpoint_path: str, eval: bool = False, strict: bool = True, cache=False 49 | ) -> None: 50 | """Load a model checkpoint gile and get ready for training or inference. 51 | 52 | Args: 53 | config (Coqpit): Model configuration. 54 | checkpoint_path (str): Path to the model checkpoint file. 55 | eval (bool, optional): If true, init model for inference else for training. Defaults to False. 56 | strict (bool, optional): Match all checkpoint keys to model's keys. Defaults to True. 57 | cache (bool, optional): If True, cache the file locally for subsequent calls. It is cached under `get_user_data_dir()/tts_cache`. Defaults to False. 58 | """ 59 | ... 60 | -------------------------------------------------------------------------------- /TTS/server/README.md: -------------------------------------------------------------------------------- 1 | # :frog: TTS demo server 2 | Before you use the server, make sure you [install](https://github.com/coqui-ai/TTS/tree/dev#install-tts)) :frog: TTS properly. Then, you can follow the steps below. 3 | 4 | **Note:** If you install :frog:TTS using ```pip```, you can also use the ```tts-server``` end point on the terminal. 5 | 6 | Examples runs: 7 | 8 | List officially released models. 9 | ```python TTS/server/server.py --list_models ``` 10 | 11 | Run the server with the official models. 12 | ```python TTS/server/server.py --model_name tts_models/en/ljspeech/tacotron2-DCA --vocoder_name vocoder_models/en/ljspeech/multiband-melgan``` 13 | 14 | Run the server with the official models on a GPU. 15 | ```CUDA_VISIBLE_DEVICES="0" python TTS/server/server.py --model_name tts_models/en/ljspeech/tacotron2-DCA --vocoder_name vocoder_models/en/ljspeech/multiband-melgan --use_cuda True``` 16 | 17 | Run the server with a custom models. 18 | ```python TTS/server/server.py --tts_checkpoint /path/to/tts/model.pth --tts_config /path/to/tts/config.json --vocoder_checkpoint /path/to/vocoder/model.pth --vocoder_config /path/to/vocoder/config.json``` 19 | -------------------------------------------------------------------------------- /TTS/server/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/server/__init__.py -------------------------------------------------------------------------------- /TTS/server/conf.json: -------------------------------------------------------------------------------- 1 | { 2 | "tts_path":"/media/erogol/data_ssd/Models/libri_tts/5049/", // tts model root folder 3 | "tts_file":"best_model.pth", // tts checkpoint file 4 | "tts_config":"config.json", // tts config.json file 5 | "tts_speakers": null, // json file listing speaker ids. null if no speaker embedding. 6 | "vocoder_config":null, 7 | "vocoder_file": null, 8 | "is_wavernn_batched":true, 9 | "port": 5002, 10 | "use_cuda": true, 11 | "debug": true 12 | } 13 | -------------------------------------------------------------------------------- /TTS/server/static/coqui-log-green-TTS.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/server/static/coqui-log-green-TTS.png -------------------------------------------------------------------------------- /TTS/server/templates/details.html: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | TTS engine 12 | 13 | 14 | 17 | 18 | 19 | 30 | 31 | 32 | 33 | Fork me on GitHub 35 | 36 | {% if show_details == true %} 37 | 38 |
39 | Model details 40 |
41 | 42 |
43 |
44 | CLI arguments: 45 | 46 | 47 | 48 | 49 | 50 | 51 | {% for key, value in args.items() %} 52 | 53 | 54 | 55 | 56 | 57 | 58 | {% endfor %} 59 |
CLI key Value
{{ key }}{{ value }}
60 |
61 |

62 | 63 |
64 | 65 | {% if model_config != None %} 66 | 67 |
68 | Model config: 69 | 70 | 71 | 72 | 73 | 74 | 75 | 76 | 77 | {% for key, value in model_config.items() %} 78 | 79 | 80 | 81 | 82 | 83 | 84 | {% endfor %} 85 | 86 |
Key Value
{{ key }}{{ value }}
87 |
88 | 89 | {% endif %} 90 | 91 |

92 | 93 | 94 | 95 |
96 | {% if vocoder_config != None %} 97 |
98 | Vocoder model config: 99 | 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | {% for key, value in vocoder_config.items() %} 108 | 109 | 110 | 111 | 112 | 113 | 114 | {% endfor %} 115 | 116 | 117 |
Key Value
{{ key }}{{ value }}
118 |
119 | {% endif %} 120 |

121 | 122 | {% else %} 123 |
124 | Please start server with --show_details=true to see details. 125 |
126 | 127 | {% endif %} 128 | 129 | 130 | 131 | -------------------------------------------------------------------------------- /TTS/tts/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/__init__.py -------------------------------------------------------------------------------- /TTS/tts/configs/__init__.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import os 3 | from inspect import isclass 4 | 5 | # import all files under configs/ 6 | # configs_dir = os.path.dirname(__file__) 7 | # for file in os.listdir(configs_dir): 8 | # path = os.path.join(configs_dir, file) 9 | # if not file.startswith("_") and not file.startswith(".") and (file.endswith(".py") or os.path.isdir(path)): 10 | # config_name = file[: file.find(".py")] if file.endswith(".py") else file 11 | # module = importlib.import_module("TTS.tts.configs." + config_name) 12 | # for attribute_name in dir(module): 13 | # attribute = getattr(module, attribute_name) 14 | 15 | # if isclass(attribute): 16 | # # Add the class to this package's variables 17 | # globals()[attribute_name] = attribute 18 | -------------------------------------------------------------------------------- /TTS/tts/configs/tacotron2_config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import dataclass 2 | 3 | from TTS.tts.configs.tacotron_config import TacotronConfig 4 | 5 | 6 | @dataclass 7 | class Tacotron2Config(TacotronConfig): 8 | """Defines parameters for Tacotron2 based models. 9 | 10 | Example: 11 | 12 | >>> from TTS.tts.configs.tacotron2_config import Tacotron2Config 13 | >>> config = Tacotron2Config() 14 | 15 | Check `TacotronConfig` for argument descriptions. 16 | """ 17 | 18 | model: str = "tacotron2" 19 | out_channels: int = 80 20 | encoder_in_features: int = 512 21 | decoder_in_features: int = 512 22 | -------------------------------------------------------------------------------- /TTS/tts/layers/__init__.py: -------------------------------------------------------------------------------- 1 | from TTS.tts.layers.losses import * 2 | -------------------------------------------------------------------------------- /TTS/tts/layers/align_tts/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/align_tts/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/align_tts/duration_predictor.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from TTS.tts.layers.generic.pos_encoding import PositionalEncoding 4 | from TTS.tts.layers.generic.transformer import FFTransformerBlock 5 | 6 | 7 | class DurationPredictor(nn.Module): 8 | def __init__(self, num_chars, hidden_channels, hidden_channels_ffn, num_heads): 9 | super().__init__() 10 | self.embed = nn.Embedding(num_chars, hidden_channels) 11 | self.pos_enc = PositionalEncoding(hidden_channels, dropout_p=0.1) 12 | self.FFT = FFTransformerBlock(hidden_channels, num_heads, hidden_channels_ffn, 2, 0.1) 13 | self.out_layer = nn.Conv1d(hidden_channels, 1, 1) 14 | 15 | def forward(self, text, text_lengths): 16 | # B, L -> B, L 17 | emb = self.embed(text) 18 | emb = self.pos_enc(emb.transpose(1, 2)) 19 | x = self.FFT(emb, text_lengths) 20 | x = self.out_layer(x).squeeze(-1) 21 | return x 22 | -------------------------------------------------------------------------------- /TTS/tts/layers/align_tts/mdn.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | 4 | class MDNBlock(nn.Module): 5 | """Mixture of Density Network implementation 6 | https://arxiv.org/pdf/2003.01950.pdf 7 | """ 8 | 9 | def __init__(self, in_channels, out_channels): 10 | super().__init__() 11 | self.out_channels = out_channels 12 | self.conv1 = nn.Conv1d(in_channels, in_channels, 1) 13 | self.norm = nn.LayerNorm(in_channels) 14 | self.relu = nn.ReLU() 15 | self.dropout = nn.Dropout(0.1) 16 | self.conv2 = nn.Conv1d(in_channels, out_channels, 1) 17 | 18 | def forward(self, x): 19 | o = self.conv1(x) 20 | o = o.transpose(1, 2) 21 | o = self.norm(o) 22 | o = o.transpose(1, 2) 23 | o = self.relu(o) 24 | o = self.dropout(o) 25 | mu_sigma = self.conv2(o) 26 | # TODO: check this sigmoid 27 | # mu = torch.sigmoid(mu_sigma[:, :self.out_channels//2, :]) 28 | mu = mu_sigma[:, : self.out_channels // 2, :] 29 | log_sigma = mu_sigma[:, self.out_channels // 2 :, :] 30 | return mu, log_sigma 31 | -------------------------------------------------------------------------------- /TTS/tts/layers/feed_forward/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/feed_forward/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/feed_forward/duration_predictor.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from TTS.tts.layers.generic.res_conv_bn import Conv1dBN 4 | 5 | 6 | class DurationPredictor(nn.Module): 7 | """Speedy Speech duration predictor model. 8 | Predicts phoneme durations from encoder outputs. 9 | 10 | Note: 11 | Outputs interpreted as log(durations) 12 | To get actual durations, do exp transformation 13 | 14 | conv_BN_4x1 -> conv_BN_3x1 -> conv_BN_1x1 -> conv_1x1 15 | 16 | Args: 17 | hidden_channels (int): number of channels in the inner layers. 18 | """ 19 | 20 | def __init__(self, hidden_channels): 21 | super().__init__() 22 | 23 | self.layers = nn.ModuleList( 24 | [ 25 | Conv1dBN(hidden_channels, hidden_channels, 4, 1), 26 | Conv1dBN(hidden_channels, hidden_channels, 3, 1), 27 | Conv1dBN(hidden_channels, hidden_channels, 1, 1), 28 | nn.Conv1d(hidden_channels, 1, 1), 29 | ] 30 | ) 31 | 32 | def forward(self, x, x_mask): 33 | """ 34 | Shapes: 35 | x: [B, C, T] 36 | x_mask: [B, 1, T] 37 | """ 38 | o = x 39 | for layer in self.layers: 40 | o = layer(o) * x_mask 41 | return o 42 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/generic/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/generic/aligner.py: -------------------------------------------------------------------------------- 1 | from typing import Tuple 2 | 3 | import torch 4 | from torch import nn 5 | 6 | 7 | class AlignmentNetwork(torch.nn.Module): 8 | """Aligner Network for learning alignment between the input text and the model output with Gaussian Attention. 9 | 10 | :: 11 | 12 | query -> conv1d -> relu -> conv1d -> relu -> conv1d -> L2_dist -> softmax -> alignment 13 | key -> conv1d -> relu -> conv1d -----------------------^ 14 | 15 | Args: 16 | in_query_channels (int): Number of channels in the query network. Defaults to 80. 17 | in_key_channels (int): Number of channels in the key network. Defaults to 512. 18 | attn_channels (int): Number of inner channels in the attention layers. Defaults to 80. 19 | temperature (float): Temperature for the softmax. Defaults to 0.0005. 20 | """ 21 | 22 | def __init__( 23 | self, 24 | in_query_channels=80, 25 | in_key_channels=512, 26 | attn_channels=80, 27 | temperature=0.0005, 28 | ): 29 | super().__init__() 30 | self.temperature = temperature 31 | self.softmax = torch.nn.Softmax(dim=3) 32 | self.log_softmax = torch.nn.LogSoftmax(dim=3) 33 | 34 | self.key_layer = nn.Sequential( 35 | nn.Conv1d( 36 | in_key_channels, 37 | in_key_channels * 2, 38 | kernel_size=3, 39 | padding=1, 40 | bias=True, 41 | ), 42 | torch.nn.ReLU(), 43 | nn.Conv1d(in_key_channels * 2, attn_channels, kernel_size=1, padding=0, bias=True), 44 | ) 45 | 46 | self.query_layer = nn.Sequential( 47 | nn.Conv1d( 48 | in_query_channels, 49 | in_query_channels * 2, 50 | kernel_size=3, 51 | padding=1, 52 | bias=True, 53 | ), 54 | torch.nn.ReLU(), 55 | nn.Conv1d(in_query_channels * 2, in_query_channels, kernel_size=1, padding=0, bias=True), 56 | torch.nn.ReLU(), 57 | nn.Conv1d(in_query_channels, attn_channels, kernel_size=1, padding=0, bias=True), 58 | ) 59 | 60 | def forward( 61 | self, queries: torch.tensor, keys: torch.tensor, mask: torch.tensor = None, attn_prior: torch.tensor = None 62 | ) -> Tuple[torch.tensor, torch.tensor]: 63 | """Forward pass of the aligner encoder. 64 | Shapes: 65 | - queries: :math:`[B, C, T_de]` 66 | - keys: :math:`[B, C_emb, T_en]` 67 | - mask: :math:`[B, T_de]` 68 | Output: 69 | attn (torch.tensor): :math:`[B, 1, T_en, T_de]` soft attention mask. 70 | attn_logp (torch.tensor): :math:`[ßB, 1, T_en , T_de]` log probabilities. 71 | """ 72 | key_out = self.key_layer(keys) 73 | query_out = self.query_layer(queries) 74 | attn_factor = (query_out[:, :, :, None] - key_out[:, :, None]) ** 2 75 | attn_logp = -self.temperature * attn_factor.sum(1, keepdim=True) 76 | if attn_prior is not None: 77 | attn_logp = self.log_softmax(attn_logp) + torch.log(attn_prior[:, None] + 1e-8) 78 | if mask is not None: 79 | attn_logp.data.masked_fill_(~mask.bool().unsqueeze(2), -float("inf")) 80 | attn = self.softmax(attn_logp) 81 | return attn, attn_logp 82 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/gated_conv.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from .normalization import LayerNorm 4 | 5 | 6 | class GatedConvBlock(nn.Module): 7 | """Gated convolutional block as in https://arxiv.org/pdf/1612.08083.pdf 8 | Args: 9 | in_out_channels (int): number of input/output channels. 10 | kernel_size (int): convolution kernel size. 11 | dropout_p (float): dropout rate. 12 | """ 13 | 14 | def __init__(self, in_out_channels, kernel_size, dropout_p, num_layers): 15 | super().__init__() 16 | # class arguments 17 | self.dropout_p = dropout_p 18 | self.num_layers = num_layers 19 | # define layers 20 | self.conv_layers = nn.ModuleList() 21 | self.norm_layers = nn.ModuleList() 22 | self.layers = nn.ModuleList() 23 | for _ in range(num_layers): 24 | self.conv_layers += [nn.Conv1d(in_out_channels, 2 * in_out_channels, kernel_size, padding=kernel_size // 2)] 25 | self.norm_layers += [LayerNorm(2 * in_out_channels)] 26 | 27 | def forward(self, x, x_mask): 28 | o = x 29 | res = x 30 | for idx in range(self.num_layers): 31 | o = nn.functional.dropout(o, p=self.dropout_p, training=self.training) 32 | o = self.conv_layers[idx](o * x_mask) 33 | o = self.norm_layers[idx](o) 34 | o = nn.functional.glu(o, dim=1) 35 | o = res + o 36 | res = o 37 | return o 38 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/normalization.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | 5 | class LayerNorm(nn.Module): 6 | def __init__(self, channels, eps=1e-4): 7 | """Layer norm for the 2nd dimension of the input. 8 | Args: 9 | channels (int): number of channels (2nd dimension) of the input. 10 | eps (float): to prevent 0 division 11 | 12 | Shapes: 13 | - input: (B, C, T) 14 | - output: (B, C, T) 15 | """ 16 | super().__init__() 17 | self.channels = channels 18 | self.eps = eps 19 | 20 | self.gamma = nn.Parameter(torch.ones(1, channels, 1) * 0.1) 21 | self.beta = nn.Parameter(torch.zeros(1, channels, 1)) 22 | 23 | def forward(self, x): 24 | mean = torch.mean(x, 1, keepdim=True) 25 | variance = torch.mean((x - mean) ** 2, 1, keepdim=True) 26 | x = (x - mean) * torch.rsqrt(variance + self.eps) 27 | x = x * self.gamma + self.beta 28 | return x 29 | 30 | 31 | class LayerNorm2(nn.Module): 32 | """Layer norm for the 2nd dimension of the input using torch primitive. 33 | Args: 34 | channels (int): number of channels (2nd dimension) of the input. 35 | eps (float): to prevent 0 division 36 | 37 | Shapes: 38 | - input: (B, C, T) 39 | - output: (B, C, T) 40 | """ 41 | 42 | def __init__(self, channels, eps=1e-5): 43 | super().__init__() 44 | self.channels = channels 45 | self.eps = eps 46 | 47 | self.gamma = nn.Parameter(torch.ones(channels)) 48 | self.beta = nn.Parameter(torch.zeros(channels)) 49 | 50 | def forward(self, x): 51 | x = x.transpose(1, -1) 52 | x = torch.nn.functional.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps) 53 | return x.transpose(1, -1) 54 | 55 | 56 | class TemporalBatchNorm1d(nn.BatchNorm1d): 57 | """Normalize each channel separately over time and batch.""" 58 | 59 | def __init__(self, channels, affine=True, track_running_stats=True, momentum=0.1): 60 | super().__init__(channels, affine=affine, track_running_stats=track_running_stats, momentum=momentum) 61 | 62 | def forward(self, x): 63 | return super().forward(x.transpose(2, 1)).transpose(2, 1) 64 | 65 | 66 | class ActNorm(nn.Module): 67 | """Activation Normalization bijector as an alternative to Batch Norm. It computes 68 | mean and std from a sample data in advance and it uses these values 69 | for normalization at training. 70 | 71 | Args: 72 | channels (int): input channels. 73 | ddi (False): data depended initialization flag. 74 | 75 | Shapes: 76 | - inputs: (B, C, T) 77 | - outputs: (B, C, T) 78 | """ 79 | 80 | def __init__(self, channels, ddi=False, **kwargs): # pylint: disable=unused-argument 81 | super().__init__() 82 | self.channels = channels 83 | self.initialized = not ddi 84 | 85 | self.logs = nn.Parameter(torch.zeros(1, channels, 1)) 86 | self.bias = nn.Parameter(torch.zeros(1, channels, 1)) 87 | 88 | def forward(self, x, x_mask=None, reverse=False, **kwargs): # pylint: disable=unused-argument 89 | if x_mask is None: 90 | x_mask = torch.ones(x.size(0), 1, x.size(2)).to(device=x.device, dtype=x.dtype) 91 | x_len = torch.sum(x_mask, [1, 2]) 92 | if not self.initialized: 93 | self.initialize(x, x_mask) 94 | self.initialized = True 95 | 96 | if reverse: 97 | z = (x - self.bias) * torch.exp(-self.logs) * x_mask 98 | logdet = None 99 | else: 100 | z = (self.bias + torch.exp(self.logs) * x) * x_mask 101 | logdet = torch.sum(self.logs) * x_len # [b] 102 | 103 | return z, logdet 104 | 105 | def store_inverse(self): 106 | pass 107 | 108 | def set_ddi(self, ddi): 109 | self.initialized = not ddi 110 | 111 | def initialize(self, x, x_mask): 112 | with torch.no_grad(): 113 | denom = torch.sum(x_mask, [0, 2]) 114 | m = torch.sum(x * x_mask, [0, 2]) / denom 115 | m_sq = torch.sum(x * x * x_mask, [0, 2]) / denom 116 | v = m_sq - (m**2) 117 | logs = 0.5 * torch.log(torch.clamp_min(v, 1e-6)) 118 | 119 | bias_init = (-m * torch.exp(-logs)).view(*self.bias.shape).to(dtype=self.bias.dtype) 120 | logs_init = (-logs).view(*self.logs.shape).to(dtype=self.logs.dtype) 121 | 122 | self.bias.data.copy_(bias_init) 123 | self.logs.data.copy_(logs_init) 124 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/pos_encoding.py: -------------------------------------------------------------------------------- 1 | import math 2 | 3 | import torch 4 | from torch import nn 5 | 6 | 7 | class PositionalEncoding(nn.Module): 8 | """Sinusoidal positional encoding for non-recurrent neural networks. 9 | Implementation based on "Attention Is All You Need" 10 | 11 | Args: 12 | channels (int): embedding size 13 | dropout_p (float): dropout rate applied to the output. 14 | max_len (int): maximum sequence length. 15 | use_scale (bool): whether to use a learnable scaling coefficient. 16 | """ 17 | 18 | def __init__(self, channels, dropout_p=0.0, max_len=5000, use_scale=False): 19 | super().__init__() 20 | if channels % 2 != 0: 21 | raise ValueError( 22 | "Cannot use sin/cos positional encoding with " "odd channels (got channels={:d})".format(channels) 23 | ) 24 | self.use_scale = use_scale 25 | if use_scale: 26 | self.scale = torch.nn.Parameter(torch.ones(1)) 27 | pe = torch.zeros(max_len, channels) 28 | position = torch.arange(0, max_len).unsqueeze(1) 29 | div_term = torch.pow(10000, torch.arange(0, channels, 2).float() / channels) 30 | pe[:, 0::2] = torch.sin(position.float() * div_term) 31 | pe[:, 1::2] = torch.cos(position.float() * div_term) 32 | pe = pe.unsqueeze(0).transpose(1, 2) 33 | self.register_buffer("pe", pe) 34 | if dropout_p > 0: 35 | self.dropout = nn.Dropout(p=dropout_p) 36 | self.channels = channels 37 | 38 | def forward(self, x, mask=None, first_idx=None, last_idx=None): 39 | """ 40 | Shapes: 41 | x: [B, C, T] 42 | mask: [B, 1, T] 43 | first_idx: int 44 | last_idx: int 45 | """ 46 | 47 | x = x * math.sqrt(self.channels) 48 | if first_idx is None: 49 | if self.pe.size(2) < x.size(2): 50 | raise RuntimeError( 51 | f"Sequence is {x.size(2)} but PositionalEncoding is" 52 | f" limited to {self.pe.size(2)}. See max_len argument." 53 | ) 54 | if mask is not None: 55 | pos_enc = self.pe[:, :, : x.size(2)] * mask 56 | else: 57 | pos_enc = self.pe[:, :, : x.size(2)] 58 | if self.use_scale: 59 | x = x + self.scale * pos_enc 60 | else: 61 | x = x + pos_enc 62 | else: 63 | if self.use_scale: 64 | x = x + self.scale * self.pe[:, :, first_idx:last_idx] 65 | else: 66 | x = x + self.pe[:, :, first_idx:last_idx] 67 | if hasattr(self, "dropout"): 68 | x = self.dropout(x) 69 | return x 70 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/time_depth_sep_conv.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | 5 | class TimeDepthSeparableConv(nn.Module): 6 | """Time depth separable convolution as in https://arxiv.org/pdf/1904.02619.pdf 7 | It shows competative results with less computation and memory footprint.""" 8 | 9 | def __init__(self, in_channels, hid_channels, out_channels, kernel_size, bias=True): 10 | super().__init__() 11 | 12 | self.in_channels = in_channels 13 | self.out_channels = out_channels 14 | self.hid_channels = hid_channels 15 | self.kernel_size = kernel_size 16 | 17 | self.time_conv = nn.Conv1d( 18 | in_channels, 19 | 2 * hid_channels, 20 | kernel_size=1, 21 | stride=1, 22 | padding=0, 23 | bias=bias, 24 | ) 25 | self.norm1 = nn.BatchNorm1d(2 * hid_channels) 26 | self.depth_conv = nn.Conv1d( 27 | hid_channels, 28 | hid_channels, 29 | kernel_size, 30 | stride=1, 31 | padding=(kernel_size - 1) // 2, 32 | groups=hid_channels, 33 | bias=bias, 34 | ) 35 | self.norm2 = nn.BatchNorm1d(hid_channels) 36 | self.time_conv2 = nn.Conv1d( 37 | hid_channels, 38 | out_channels, 39 | kernel_size=1, 40 | stride=1, 41 | padding=0, 42 | bias=bias, 43 | ) 44 | self.norm3 = nn.BatchNorm1d(out_channels) 45 | 46 | def forward(self, x): 47 | x_res = x 48 | x = self.time_conv(x) 49 | x = self.norm1(x) 50 | x = nn.functional.glu(x, dim=1) 51 | x = self.depth_conv(x) 52 | x = self.norm2(x) 53 | x = x * torch.sigmoid(x) 54 | x = self.time_conv2(x) 55 | x = self.norm3(x) 56 | x = x_res + x 57 | return x 58 | 59 | 60 | class TimeDepthSeparableConvBlock(nn.Module): 61 | def __init__(self, in_channels, hid_channels, out_channels, num_layers, kernel_size, bias=True): 62 | super().__init__() 63 | assert (kernel_size - 1) % 2 == 0 64 | assert num_layers > 1 65 | 66 | self.layers = nn.ModuleList() 67 | layer = TimeDepthSeparableConv( 68 | in_channels, hid_channels, out_channels if num_layers == 1 else hid_channels, kernel_size, bias 69 | ) 70 | self.layers.append(layer) 71 | for idx in range(num_layers - 1): 72 | layer = TimeDepthSeparableConv( 73 | hid_channels, 74 | hid_channels, 75 | out_channels if (idx + 1) == (num_layers - 1) else hid_channels, 76 | kernel_size, 77 | bias, 78 | ) 79 | self.layers.append(layer) 80 | 81 | def forward(self, x, mask): 82 | for layer in self.layers: 83 | x = layer(x * mask) 84 | return x 85 | -------------------------------------------------------------------------------- /TTS/tts/layers/generic/transformer.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | from torch import nn 4 | 5 | 6 | class FFTransformer(nn.Module): 7 | def __init__(self, in_out_channels, num_heads, hidden_channels_ffn=1024, kernel_size_fft=3, dropout_p=0.1): 8 | super().__init__() 9 | self.self_attn = nn.MultiheadAttention(in_out_channels, num_heads, dropout=dropout_p) 10 | 11 | padding = (kernel_size_fft - 1) // 2 12 | self.conv1 = nn.Conv1d(in_out_channels, hidden_channels_ffn, kernel_size=kernel_size_fft, padding=padding) 13 | self.conv2 = nn.Conv1d(hidden_channels_ffn, in_out_channels, kernel_size=kernel_size_fft, padding=padding) 14 | 15 | self.norm1 = nn.LayerNorm(in_out_channels) 16 | self.norm2 = nn.LayerNorm(in_out_channels) 17 | 18 | self.dropout1 = nn.Dropout(dropout_p) 19 | self.dropout2 = nn.Dropout(dropout_p) 20 | 21 | def forward(self, src, src_mask=None, src_key_padding_mask=None): 22 | """😦 ugly looking with all the transposing""" 23 | src = src.permute(2, 0, 1) 24 | src2, enc_align = self.self_attn(src, src, src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask) 25 | src = src + self.dropout1(src2) 26 | src = self.norm1(src + src2) 27 | # T x B x D -> B x D x T 28 | src = src.permute(1, 2, 0) 29 | src2 = self.conv2(F.relu(self.conv1(src))) 30 | src2 = self.dropout2(src2) 31 | src = src + src2 32 | src = src.transpose(1, 2) 33 | src = self.norm2(src) 34 | src = src.transpose(1, 2) 35 | return src, enc_align 36 | 37 | 38 | class FFTransformerBlock(nn.Module): 39 | def __init__(self, in_out_channels, num_heads, hidden_channels_ffn, num_layers, dropout_p): 40 | super().__init__() 41 | self.fft_layers = nn.ModuleList( 42 | [ 43 | FFTransformer( 44 | in_out_channels=in_out_channels, 45 | num_heads=num_heads, 46 | hidden_channels_ffn=hidden_channels_ffn, 47 | dropout_p=dropout_p, 48 | ) 49 | for _ in range(num_layers) 50 | ] 51 | ) 52 | 53 | def forward(self, x, mask=None, g=None): # pylint: disable=unused-argument 54 | """ 55 | TODO: handle multi-speaker 56 | Shapes: 57 | - x: :math:`[B, C, T]` 58 | - mask: :math:`[B, 1, T] or [B, T]` 59 | """ 60 | if mask is not None and mask.ndim == 3: 61 | mask = mask.squeeze(1) 62 | # mask is negated, torch uses 1s and 0s reversely. 63 | mask = ~mask.bool() 64 | alignments = [] 65 | for layer in self.fft_layers: 66 | x, align = layer(x, src_key_padding_mask=mask) 67 | alignments.append(align.unsqueeze(1)) 68 | alignments = torch.cat(alignments, 1) 69 | return x 70 | 71 | 72 | class FFTDurationPredictor: 73 | def __init__( 74 | self, in_channels, hidden_channels, num_heads, num_layers, dropout_p=0.1, cond_channels=None 75 | ): # pylint: disable=unused-argument 76 | self.fft = FFTransformerBlock(in_channels, num_heads, hidden_channels, num_layers, dropout_p) 77 | self.proj = nn.Linear(in_channels, 1) 78 | 79 | def forward(self, x, mask=None, g=None): # pylint: disable=unused-argument 80 | """ 81 | Shapes: 82 | - x: :math:`[B, C, T]` 83 | - mask: :math:`[B, 1, T]` 84 | 85 | TODO: Handle the cond input 86 | """ 87 | x = self.fft(x, mask=mask) 88 | x = self.proj(x) 89 | return x 90 | -------------------------------------------------------------------------------- /TTS/tts/layers/glow_tts/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/glow_tts/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/glow_tts/duration_predictor.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | from ..generic.normalization import LayerNorm 5 | 6 | 7 | class DurationPredictor(nn.Module): 8 | """Glow-TTS duration prediction model. 9 | 10 | :: 11 | 12 | [2 x (conv1d_kxk -> relu -> layer_norm -> dropout)] -> conv1d_1x1 -> durs 13 | 14 | Args: 15 | in_channels (int): Number of channels of the input tensor. 16 | hidden_channels (int): Number of hidden channels of the network. 17 | kernel_size (int): Kernel size for the conv layers. 18 | dropout_p (float): Dropout rate used after each conv layer. 19 | """ 20 | 21 | def __init__(self, in_channels, hidden_channels, kernel_size, dropout_p, cond_channels=None, language_emb_dim=None): 22 | super().__init__() 23 | 24 | # add language embedding dim in the input 25 | if language_emb_dim: 26 | in_channels += language_emb_dim 27 | 28 | # class arguments 29 | self.in_channels = in_channels 30 | self.filter_channels = hidden_channels 31 | self.kernel_size = kernel_size 32 | self.dropout_p = dropout_p 33 | # layers 34 | self.drop = nn.Dropout(dropout_p) 35 | self.conv_1 = nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2) 36 | self.norm_1 = LayerNorm(hidden_channels) 37 | self.conv_2 = nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2) 38 | self.norm_2 = LayerNorm(hidden_channels) 39 | # output layer 40 | self.proj = nn.Conv1d(hidden_channels, 1, 1) 41 | if cond_channels is not None and cond_channels != 0: 42 | self.cond = nn.Conv1d(cond_channels, in_channels, 1) 43 | 44 | if language_emb_dim != 0 and language_emb_dim is not None: 45 | self.cond_lang = nn.Conv1d(language_emb_dim, in_channels, 1) 46 | 47 | def forward(self, x, x_mask, g=None, lang_emb=None): 48 | """ 49 | Shapes: 50 | - x: :math:`[B, C, T]` 51 | - x_mask: :math:`[B, 1, T]` 52 | - g: :math:`[B, C, 1]` 53 | """ 54 | if g is not None: 55 | x = x + self.cond(g) 56 | 57 | if lang_emb is not None: 58 | x = x + self.cond_lang(lang_emb) 59 | 60 | x = self.conv_1(x * x_mask) 61 | x = torch.relu(x) 62 | x = self.norm_1(x) 63 | x = self.drop(x) 64 | x = self.conv_2(x * x_mask) 65 | x = torch.relu(x) 66 | x = self.norm_2(x) 67 | x = self.drop(x) 68 | x = self.proj(x * x_mask) 69 | return x * x_mask 70 | -------------------------------------------------------------------------------- /TTS/tts/layers/overflow/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/overflow/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/overflow/decoder.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | from TTS.tts.layers.glow_tts.decoder import Decoder as GlowDecoder 5 | from TTS.tts.utils.helpers import sequence_mask 6 | 7 | 8 | class Decoder(nn.Module): 9 | """Uses glow decoder with some modifications. 10 | :: 11 | 12 | Squeeze -> ActNorm -> InvertibleConv1x1 -> AffineCoupling -> Unsqueeze 13 | 14 | Args: 15 | in_channels (int): channels of input tensor. 16 | hidden_channels (int): hidden decoder channels. 17 | kernel_size (int): Coupling block kernel size. (Wavenet filter kernel size.) 18 | dilation_rate (int): rate to increase dilation by each layer in a decoder block. 19 | num_flow_blocks (int): number of decoder blocks. 20 | num_coupling_layers (int): number coupling layers. (number of wavenet layers.) 21 | dropout_p (float): wavenet dropout rate. 22 | sigmoid_scale (bool): enable/disable sigmoid scaling in coupling layer. 23 | """ 24 | 25 | def __init__( 26 | self, 27 | in_channels, 28 | hidden_channels, 29 | kernel_size, 30 | dilation_rate, 31 | num_flow_blocks, 32 | num_coupling_layers, 33 | dropout_p=0.0, 34 | num_splits=4, 35 | num_squeeze=2, 36 | sigmoid_scale=False, 37 | c_in_channels=0, 38 | ): 39 | super().__init__() 40 | 41 | self.glow_decoder = GlowDecoder( 42 | in_channels, 43 | hidden_channels, 44 | kernel_size, 45 | dilation_rate, 46 | num_flow_blocks, 47 | num_coupling_layers, 48 | dropout_p, 49 | num_splits, 50 | num_squeeze, 51 | sigmoid_scale, 52 | c_in_channels, 53 | ) 54 | self.n_sqz = num_squeeze 55 | 56 | def forward(self, x, x_len, g=None, reverse=False): 57 | """ 58 | Input shapes: 59 | - x: :math:`[B, C, T]` 60 | - x_len :math:`[B]` 61 | - g: :math:`[B, C]` 62 | 63 | Output shapes: 64 | - x: :math:`[B, C, T]` 65 | - x_len :math:`[B]` 66 | - logget_tot :math:`[B]` 67 | """ 68 | x, x_len, x_max_len = self.preprocess(x, x_len, x_len.max()) 69 | x_mask = torch.unsqueeze(sequence_mask(x_len, x_max_len), 1).to(x.dtype) 70 | x, logdet_tot = self.glow_decoder(x, x_mask, g, reverse) 71 | return x, x_len, logdet_tot 72 | 73 | def preprocess(self, y, y_lengths, y_max_length): 74 | if y_max_length is not None: 75 | y_max_length = torch.div(y_max_length, self.n_sqz, rounding_mode="floor") * self.n_sqz 76 | y = y[:, :, :y_max_length] 77 | y_lengths = torch.div(y_lengths, self.n_sqz, rounding_mode="floor") * self.n_sqz 78 | return y, y_lengths, y_max_length 79 | 80 | def store_inverse(self): 81 | self.glow_decoder.store_inverse() 82 | -------------------------------------------------------------------------------- /TTS/tts/layers/overflow/plotting_utils.py: -------------------------------------------------------------------------------- 1 | from typing import Any 2 | 3 | import matplotlib.pyplot as plt 4 | import numpy as np 5 | import torch 6 | 7 | 8 | def validate_numpy_array(value: Any): 9 | r""" 10 | Validates the input and makes sure it returns a numpy array (i.e on CPU) 11 | 12 | Args: 13 | value (Any): the input value 14 | 15 | Raises: 16 | TypeError: if the value is not a numpy array or torch tensor 17 | 18 | Returns: 19 | np.ndarray: numpy array of the value 20 | """ 21 | if isinstance(value, np.ndarray): 22 | pass 23 | elif isinstance(value, list): 24 | value = np.array(value) 25 | elif torch.is_tensor(value): 26 | value = value.cpu().numpy() 27 | else: 28 | raise TypeError("Value must be a numpy array, a torch tensor or a list") 29 | 30 | return value 31 | 32 | 33 | def get_spec_from_most_probable_state(log_alpha_scaled, means, decoder=None): 34 | """Get the most probable state means from the log_alpha_scaled. 35 | 36 | Args: 37 | log_alpha_scaled (torch.Tensor): Log alpha scaled values. 38 | - Shape: :math:`(T, N)` 39 | means (torch.Tensor): Means of the states. 40 | - Shape: :math:`(N, T, D_out)` 41 | decoder (torch.nn.Module): Decoder module to decode the latent to melspectrogram. Defaults to None. 42 | """ 43 | max_state_numbers = torch.max(log_alpha_scaled, dim=1)[1] 44 | max_len = means.shape[0] 45 | n_mel_channels = means.shape[2] 46 | max_state_numbers = max_state_numbers.unsqueeze(1).unsqueeze(1).expand(max_len, 1, n_mel_channels) 47 | means = torch.gather(means, 1, max_state_numbers).squeeze(1).to(log_alpha_scaled.dtype) 48 | if decoder is not None: 49 | mel = ( 50 | decoder(means.T.unsqueeze(0), torch.tensor([means.shape[0]], device=means.device), reverse=True)[0] 51 | .squeeze(0) 52 | .T 53 | ) 54 | else: 55 | mel = means 56 | return mel 57 | 58 | 59 | def plot_transition_probabilities_to_numpy(states, transition_probabilities, output_fig=False): 60 | """Generates trainsition probabilities plot for the states and the probability of transition. 61 | 62 | Args: 63 | states (torch.IntTensor): the states 64 | transition_probabilities (torch.FloatTensor): the transition probabilities 65 | """ 66 | states = validate_numpy_array(states) 67 | transition_probabilities = validate_numpy_array(transition_probabilities) 68 | 69 | fig, ax = plt.subplots(figsize=(30, 3)) 70 | ax.plot(transition_probabilities, "o") 71 | ax.set_title("Transition probability of state") 72 | ax.set_xlabel("hidden state") 73 | ax.set_ylabel("probability") 74 | ax.set_xticks([i for i in range(len(transition_probabilities))]) # pylint: disable=unnecessary-comprehension 75 | ax.set_xticklabels([int(x) for x in states], rotation=90) 76 | plt.tight_layout() 77 | if not output_fig: 78 | plt.close() 79 | return fig 80 | -------------------------------------------------------------------------------- /TTS/tts/layers/tacotron/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/layers/tacotron/__init__.py -------------------------------------------------------------------------------- /TTS/tts/layers/vits/discriminator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from torch.nn.modules.conv import Conv1d 4 | 5 | from TTS.vocoder.models.hifigan_discriminator import DiscriminatorP, MultiPeriodDiscriminator 6 | 7 | 8 | class DiscriminatorS(torch.nn.Module): 9 | """HiFiGAN Scale Discriminator. Channel sizes are different from the original HiFiGAN. 10 | 11 | Args: 12 | use_spectral_norm (bool): if `True` swith to spectral norm instead of weight norm. 13 | """ 14 | 15 | def __init__(self, use_spectral_norm=False): 16 | super().__init__() 17 | norm_f = nn.utils.spectral_norm if use_spectral_norm else nn.utils.weight_norm 18 | self.convs = nn.ModuleList( 19 | [ 20 | norm_f(Conv1d(1, 16, 15, 1, padding=7)), 21 | norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), 22 | norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), 23 | norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), 24 | norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), 25 | norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), 26 | ] 27 | ) 28 | self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) 29 | 30 | def forward(self, x): 31 | """ 32 | Args: 33 | x (Tensor): input waveform. 34 | 35 | Returns: 36 | Tensor: discriminator scores. 37 | List[Tensor]: list of features from the convolutiona layers. 38 | """ 39 | feat = [] 40 | for l in self.convs: 41 | x = l(x) 42 | x = torch.nn.functional.leaky_relu(x, 0.1) 43 | feat.append(x) 44 | x = self.conv_post(x) 45 | feat.append(x) 46 | x = torch.flatten(x, 1, -1) 47 | return x, feat 48 | 49 | 50 | class VitsDiscriminator(nn.Module): 51 | """VITS discriminator wrapping one Scale Discriminator and a stack of Period Discriminator. 52 | 53 | :: 54 | waveform -> ScaleDiscriminator() -> scores_sd, feats_sd --> append() -> scores, feats 55 | |--> MultiPeriodDiscriminator() -> scores_mpd, feats_mpd ^ 56 | 57 | Args: 58 | use_spectral_norm (bool): if `True` swith to spectral norm instead of weight norm. 59 | """ 60 | 61 | def __init__(self, periods=(2, 3, 5, 7, 11), use_spectral_norm=False): 62 | super().__init__() 63 | self.nets = nn.ModuleList() 64 | self.nets.append(DiscriminatorS(use_spectral_norm=use_spectral_norm)) 65 | self.nets.extend([DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]) 66 | 67 | def forward(self, x, x_hat=None): 68 | """ 69 | Args: 70 | x (Tensor): ground truth waveform. 71 | x_hat (Tensor): predicted waveform. 72 | 73 | Returns: 74 | List[Tensor]: discriminator scores. 75 | List[List[Tensor]]: list of list of features from each layers of each discriminator. 76 | """ 77 | x_scores = [] 78 | x_hat_scores = [] if x_hat is not None else None 79 | x_feats = [] 80 | x_hat_feats = [] if x_hat is not None else None 81 | for net in self.nets: 82 | x_score, x_feat = net(x) 83 | x_scores.append(x_score) 84 | x_feats.append(x_feat) 85 | if x_hat is not None: 86 | x_hat_score, x_hat_feat = net(x_hat) 87 | x_hat_scores.append(x_hat_score) 88 | x_hat_feats.append(x_hat_feat) 89 | return x_scores, x_feats, x_hat_scores, x_hat_feats 90 | -------------------------------------------------------------------------------- /TTS/tts/models/__init__.py: -------------------------------------------------------------------------------- 1 | from typing import Dict, List, Union 2 | 3 | from TTS.utils.generic_utils import find_module 4 | 5 | 6 | def setup_model(config: "Coqpit", samples: Union[List[List], List[Dict]] = None) -> "BaseTTS": 7 | print(" > Using model: {}".format(config.model)) 8 | # fetch the right model implementation. 9 | if "base_model" in config and config["base_model"] is not None: 10 | MyModel = find_module("TTS.tts.models", config.base_model.lower()) 11 | else: 12 | MyModel = find_module("TTS.tts.models", config.model.lower()) 13 | model = MyModel.init_from_config(config, samples) 14 | return model 15 | -------------------------------------------------------------------------------- /TTS/tts/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/data.py: -------------------------------------------------------------------------------- 1 | import bisect 2 | 3 | import numpy as np 4 | import torch 5 | 6 | 7 | def _pad_data(x, length): 8 | _pad = 0 9 | assert x.ndim == 1 10 | return np.pad(x, (0, length - x.shape[0]), mode="constant", constant_values=_pad) 11 | 12 | 13 | def prepare_data(inputs): 14 | max_len = max((len(x) for x in inputs)) 15 | return np.stack([_pad_data(x, max_len) for x in inputs]) 16 | 17 | 18 | def _pad_tensor(x, length): 19 | _pad = 0.0 20 | assert x.ndim == 2 21 | x = np.pad(x, [[0, 0], [0, length - x.shape[1]]], mode="constant", constant_values=_pad) 22 | return x 23 | 24 | 25 | def prepare_tensor(inputs, out_steps): 26 | max_len = max((x.shape[1] for x in inputs)) 27 | remainder = max_len % out_steps 28 | pad_len = max_len + (out_steps - remainder) if remainder > 0 else max_len 29 | return np.stack([_pad_tensor(x, pad_len) for x in inputs]) 30 | 31 | 32 | def _pad_stop_target(x: np.ndarray, length: int, pad_val=1) -> np.ndarray: 33 | """Pad stop target array. 34 | 35 | Args: 36 | x (np.ndarray): Stop target array. 37 | length (int): Length after padding. 38 | pad_val (int, optional): Padding value. Defaults to 1. 39 | 40 | Returns: 41 | np.ndarray: Padded stop target array. 42 | """ 43 | assert x.ndim == 1 44 | return np.pad(x, (0, length - x.shape[0]), mode="constant", constant_values=pad_val) 45 | 46 | 47 | def prepare_stop_target(inputs, out_steps): 48 | """Pad row vectors with 1.""" 49 | max_len = max((x.shape[0] for x in inputs)) 50 | remainder = max_len % out_steps 51 | pad_len = max_len + (out_steps - remainder) if remainder > 0 else max_len 52 | return np.stack([_pad_stop_target(x, pad_len) for x in inputs]) 53 | 54 | 55 | def pad_per_step(inputs, pad_len): 56 | return np.pad(inputs, [[0, 0], [0, 0], [0, pad_len]], mode="constant", constant_values=0.0) 57 | 58 | 59 | def get_length_balancer_weights(items: list, num_buckets=10): 60 | # get all durations 61 | audio_lengths = np.array([item["audio_length"] for item in items]) 62 | # create the $num_buckets buckets classes based in the dataset max and min length 63 | max_length = int(max(audio_lengths)) 64 | min_length = int(min(audio_lengths)) 65 | step = int((max_length - min_length) / num_buckets) + 1 66 | buckets_classes = [i + step for i in range(min_length, (max_length - step) + num_buckets + 1, step)] 67 | # add each sample in their respective length bucket 68 | buckets_names = np.array( 69 | [buckets_classes[bisect.bisect_left(buckets_classes, item["audio_length"])] for item in items] 70 | ) 71 | # count and compute the weights_bucket for each sample 72 | unique_buckets_names = np.unique(buckets_names).tolist() 73 | bucket_ids = [unique_buckets_names.index(l) for l in buckets_names] 74 | bucket_count = np.array([len(np.where(buckets_names == l)[0]) for l in unique_buckets_names]) 75 | weight_bucket = 1.0 / bucket_count 76 | dataset_samples_weight = np.array([weight_bucket[l] for l in bucket_ids]) 77 | # normalize 78 | dataset_samples_weight = dataset_samples_weight / np.linalg.norm(dataset_samples_weight) 79 | return torch.from_numpy(dataset_samples_weight).float() 80 | -------------------------------------------------------------------------------- /TTS/tts/utils/languages.py: -------------------------------------------------------------------------------- 1 | import os 2 | from typing import Any, Dict, List 3 | 4 | import fsspec 5 | import numpy as np 6 | import torch 7 | from coqpit import Coqpit 8 | 9 | from TTS.config import check_config_and_model_args 10 | from TTS.tts.utils.managers import BaseIDManager 11 | 12 | 13 | class LanguageManager(BaseIDManager): 14 | """Manage the languages for multi-lingual 🐸TTS models. Load a datafile and parse the information 15 | in a way that can be queried by language. 16 | 17 | Args: 18 | language_ids_file_path (str, optional): Path to the metafile that maps language names to ids used by 19 | TTS models. Defaults to "". 20 | config (Coqpit, optional): Coqpit config that contains the language information in the datasets filed. 21 | Defaults to None. 22 | 23 | Examples: 24 | >>> manager = LanguageManager(language_ids_file_path=language_ids_file_path) 25 | >>> language_id_mapper = manager.language_ids 26 | """ 27 | 28 | def __init__( 29 | self, 30 | language_ids_file_path: str = "", 31 | config: Coqpit = None, 32 | ): 33 | super().__init__(id_file_path=language_ids_file_path) 34 | 35 | if config: 36 | self.set_language_ids_from_config(config) 37 | 38 | @property 39 | def num_languages(self) -> int: 40 | return len(list(self.name_to_id.keys())) 41 | 42 | @property 43 | def language_names(self) -> List: 44 | return list(self.name_to_id.keys()) 45 | 46 | @staticmethod 47 | def parse_language_ids_from_config(c: Coqpit) -> Dict: 48 | """Set language id from config. 49 | 50 | Args: 51 | c (Coqpit): Config 52 | 53 | Returns: 54 | Tuple[Dict, int]: Language ID mapping and the number of languages. 55 | """ 56 | languages = set({}) 57 | for dataset in c.datasets: 58 | if "language" in dataset: 59 | languages.add(dataset["language"]) 60 | else: 61 | raise ValueError(f"Dataset {dataset['name']} has no language specified.") 62 | return {name: i for i, name in enumerate(sorted(list(languages)))} 63 | 64 | def set_language_ids_from_config(self, c: Coqpit) -> None: 65 | """Set language IDs from config samples. 66 | 67 | Args: 68 | c (Coqpit): Config. 69 | """ 70 | self.name_to_id = self.parse_language_ids_from_config(c) 71 | 72 | @staticmethod 73 | def parse_ids_from_data(items: List, parse_key: str) -> Any: 74 | raise NotImplementedError 75 | 76 | def set_ids_from_data(self, items: List, parse_key: str) -> Any: 77 | raise NotImplementedError 78 | 79 | def save_ids_to_file(self, file_path: str) -> None: 80 | """Save language IDs to a json file. 81 | 82 | Args: 83 | file_path (str): Path to the output file. 84 | """ 85 | self._save_json(file_path, self.name_to_id) 86 | 87 | @staticmethod 88 | def init_from_config(config: Coqpit) -> "LanguageManager": 89 | """Initialize the language manager from a Coqpit config. 90 | 91 | Args: 92 | config (Coqpit): Coqpit config. 93 | """ 94 | language_manager = None 95 | if check_config_and_model_args(config, "use_language_embedding", True): 96 | if config.get("language_ids_file", None): 97 | language_manager = LanguageManager(language_ids_file_path=config.language_ids_file) 98 | language_manager = LanguageManager(config=config) 99 | return language_manager 100 | 101 | 102 | def _set_file_path(path): 103 | """Find the language_ids.json under the given path or the above it. 104 | Intended to band aid the different paths returned in restored and continued training.""" 105 | path_restore = os.path.join(os.path.dirname(path), "language_ids.json") 106 | path_continue = os.path.join(path, "language_ids.json") 107 | fs = fsspec.get_mapper(path).fs 108 | if fs.exists(path_restore): 109 | return path_restore 110 | if fs.exists(path_continue): 111 | return path_continue 112 | return None 113 | 114 | 115 | def get_language_balancer_weights(items: list): 116 | language_names = np.array([item["language"] for item in items]) 117 | unique_language_names = np.unique(language_names).tolist() 118 | language_ids = [unique_language_names.index(l) for l in language_names] 119 | language_count = np.array([len(np.where(language_names == l)[0]) for l in unique_language_names]) 120 | weight_language = 1.0 / language_count 121 | # get weight for each sample 122 | dataset_samples_weight = np.array([weight_language[l] for l in language_ids]) 123 | # normalize 124 | dataset_samples_weight = dataset_samples_weight / np.linalg.norm(dataset_samples_weight) 125 | return torch.from_numpy(dataset_samples_weight).float() 126 | -------------------------------------------------------------------------------- /TTS/tts/utils/measures.py: -------------------------------------------------------------------------------- 1 | def alignment_diagonal_score(alignments, binary=False): 2 | """ 3 | Compute how diagonal alignment predictions are. It is useful 4 | to measure the alignment consistency of a model 5 | Args: 6 | alignments (torch.Tensor): batch of alignments. 7 | binary (bool): if True, ignore scores and consider attention 8 | as a binary mask. 9 | Shape: 10 | - alignments : :math:`[B, T_de, T_en]` 11 | """ 12 | maxs = alignments.max(dim=1)[0] 13 | if binary: 14 | maxs[maxs > 0] = 1 15 | return maxs.mean(dim=1).mean(dim=0).item() 16 | -------------------------------------------------------------------------------- /TTS/tts/utils/monotonic_align/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/monotonic_align/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/monotonic_align/core.pyx: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | cimport cython 4 | cimport numpy as np 5 | 6 | from cython.parallel import prange 7 | 8 | 9 | @cython.boundscheck(False) 10 | @cython.wraparound(False) 11 | cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_x, int t_y, float max_neg_val) nogil: 12 | cdef int x 13 | cdef int y 14 | cdef float v_prev 15 | cdef float v_cur 16 | cdef float tmp 17 | cdef int index = t_x - 1 18 | 19 | for y in range(t_y): 20 | for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)): 21 | if x == y: 22 | v_cur = max_neg_val 23 | else: 24 | v_cur = value[x, y-1] 25 | if x == 0: 26 | if y == 0: 27 | v_prev = 0. 28 | else: 29 | v_prev = max_neg_val 30 | else: 31 | v_prev = value[x-1, y-1] 32 | value[x, y] = max(v_cur, v_prev) + value[x, y] 33 | 34 | for y in range(t_y - 1, -1, -1): 35 | path[index, y] = 1 36 | if index != 0 and (index == y or value[index, y-1] < value[index-1, y-1]): 37 | index = index - 1 38 | 39 | 40 | @cython.boundscheck(False) 41 | @cython.wraparound(False) 42 | cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_xs, int[::1] t_ys, float max_neg_val=-1e9) nogil: 43 | cdef int b = values.shape[0] 44 | 45 | cdef int i 46 | for i in prange(b, nogil=True): 47 | maximum_path_each(paths[i], values[i], t_xs[i], t_ys[i], max_neg_val) 48 | -------------------------------------------------------------------------------- /TTS/tts/utils/monotonic_align/setup.py: -------------------------------------------------------------------------------- 1 | # from distutils.core import setup 2 | # from Cython.Build import cythonize 3 | # import numpy 4 | 5 | # setup(name='monotonic_align', 6 | # ext_modules=cythonize("core.pyx"), 7 | # include_dirs=[numpy.get_include()]) 8 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/__init__.py: -------------------------------------------------------------------------------- 1 | from TTS.tts.utils.text.tokenizer import TTSTokenizer 2 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/chinese_mandarin/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/text/chinese_mandarin/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/text/chinese_mandarin/numbers.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | 4 | # Licensed under WTFPL or the Unlicense or CC0. 5 | # This uses Python 3, but it's easy to port to Python 2 by changing 6 | # strings to u'xx'. 7 | 8 | import itertools 9 | import re 10 | 11 | 12 | def _num2chinese(num: str, big=False, simp=True, o=False, twoalt=False) -> str: 13 | """Convert numerical arabic numbers (0->9) to chinese hanzi numbers (〇 -> 九) 14 | 15 | Args: 16 | num (str): arabic number to convert 17 | big (bool, optional): use financial characters. Defaults to False. 18 | simp (bool, optional): use simplified characters instead of tradictional characters. Defaults to True. 19 | o (bool, optional): use 〇 for 'zero'. Defaults to False. 20 | twoalt (bool, optional): use 两/兩 for 'two' when appropriate. Defaults to False. 21 | 22 | Raises: 23 | ValueError: if number is more than 1e48 24 | ValueError: if 'e' exposent in number 25 | 26 | Returns: 27 | str: converted number as hanzi characters 28 | """ 29 | 30 | # check num first 31 | nd = str(num) 32 | if abs(float(nd)) >= 1e48: 33 | raise ValueError("number out of range") 34 | if "e" in nd: 35 | raise ValueError("scientific notation is not supported") 36 | c_symbol = "正负点" if simp else "正負點" 37 | if o: # formal 38 | twoalt = False 39 | if big: 40 | c_basic = "零壹贰叁肆伍陆柒捌玖" if simp else "零壹貳參肆伍陸柒捌玖" 41 | c_unit1 = "拾佰仟" 42 | c_twoalt = "贰" if simp else "貳" 43 | else: 44 | c_basic = "〇一二三四五六七八九" if o else "零一二三四五六七八九" 45 | c_unit1 = "十百千" 46 | if twoalt: 47 | c_twoalt = "两" if simp else "兩" 48 | else: 49 | c_twoalt = "二" 50 | c_unit2 = "万亿兆京垓秭穰沟涧正载" if simp else "萬億兆京垓秭穰溝澗正載" 51 | revuniq = lambda l: "".join(k for k, g in itertools.groupby(reversed(l))) 52 | nd = str(num) 53 | result = [] 54 | if nd[0] == "+": 55 | result.append(c_symbol[0]) 56 | elif nd[0] == "-": 57 | result.append(c_symbol[1]) 58 | if "." in nd: 59 | integer, remainder = nd.lstrip("+-").split(".") 60 | else: 61 | integer, remainder = nd.lstrip("+-"), None 62 | if int(integer): 63 | splitted = [integer[max(i - 4, 0) : i] for i in range(len(integer), 0, -4)] 64 | intresult = [] 65 | for nu, unit in enumerate(splitted): 66 | # special cases 67 | if int(unit) == 0: # 0000 68 | intresult.append(c_basic[0]) 69 | continue 70 | if nu > 0 and int(unit) == 2: # 0002 71 | intresult.append(c_twoalt + c_unit2[nu - 1]) 72 | continue 73 | ulist = [] 74 | unit = unit.zfill(4) 75 | for nc, ch in enumerate(reversed(unit)): 76 | if ch == "0": 77 | if ulist: # ???0 78 | ulist.append(c_basic[0]) 79 | elif nc == 0: 80 | ulist.append(c_basic[int(ch)]) 81 | elif nc == 1 and ch == "1" and unit[1] == "0": 82 | # special case for tens 83 | # edit the 'elif' if you don't like 84 | # 十四, 三千零十四, 三千三百一十四 85 | ulist.append(c_unit1[0]) 86 | elif nc > 1 and ch == "2": 87 | ulist.append(c_twoalt + c_unit1[nc - 1]) 88 | else: 89 | ulist.append(c_basic[int(ch)] + c_unit1[nc - 1]) 90 | ustr = revuniq(ulist) 91 | if nu == 0: 92 | intresult.append(ustr) 93 | else: 94 | intresult.append(ustr + c_unit2[nu - 1]) 95 | result.append(revuniq(intresult).strip(c_basic[0])) 96 | else: 97 | result.append(c_basic[0]) 98 | if remainder: 99 | result.append(c_symbol[2]) 100 | result.append("".join(c_basic[int(ch)] for ch in remainder)) 101 | return "".join(result) 102 | 103 | 104 | def _number_replace(match) -> str: 105 | """function to apply in a match, transform all numbers in a match by chinese characters 106 | 107 | Args: 108 | match (re.Match): numbers regex matches 109 | 110 | Returns: 111 | str: replaced characters for the numbers 112 | """ 113 | match_str: str = match.group() 114 | return _num2chinese(match_str) 115 | 116 | 117 | def replace_numbers_to_characters_in_text(text: str) -> str: 118 | """Replace all arabic numbers in a text by their equivalent in chinese characters (simplified) 119 | 120 | Args: 121 | text (str): input text to transform 122 | 123 | Returns: 124 | str: output text 125 | """ 126 | text = re.sub(r"[0-9]+", _number_replace, text) 127 | return text 128 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/chinese_mandarin/phonemizer.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | import jieba 4 | import pypinyin 5 | 6 | from .pinyinToPhonemes import PINYIN_DICT 7 | 8 | 9 | def _chinese_character_to_pinyin(text: str) -> List[str]: 10 | pinyins = pypinyin.pinyin(text, style=pypinyin.Style.TONE3, heteronym=False, neutral_tone_with_five=True) 11 | pinyins_flat_list = [item for sublist in pinyins for item in sublist] 12 | return pinyins_flat_list 13 | 14 | 15 | def _chinese_pinyin_to_phoneme(pinyin: str) -> str: 16 | segment = pinyin[:-1] 17 | tone = pinyin[-1] 18 | phoneme = PINYIN_DICT.get(segment, [""])[0] 19 | return phoneme + tone 20 | 21 | 22 | def chinese_text_to_phonemes(text: str, seperator: str = "|") -> str: 23 | tokenized_text = jieba.cut(text, HMM=False) 24 | tokenized_text = " ".join(tokenized_text) 25 | pinyined_text: List[str] = _chinese_character_to_pinyin(tokenized_text) 26 | 27 | results: List[str] = [] 28 | 29 | for token in pinyined_text: 30 | if token[-1] in "12345": # TODO transform to is_pinyin() 31 | pinyin_phonemes = _chinese_pinyin_to_phoneme(token) 32 | 33 | results += list(pinyin_phonemes) 34 | else: # is ponctuation or other 35 | results += list(token) 36 | 37 | return seperator.join(results) 38 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/cmudict.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | 3 | import re 4 | 5 | VALID_SYMBOLS = [ 6 | "AA", 7 | "AA0", 8 | "AA1", 9 | "AA2", 10 | "AE", 11 | "AE0", 12 | "AE1", 13 | "AE2", 14 | "AH", 15 | "AH0", 16 | "AH1", 17 | "AH2", 18 | "AO", 19 | "AO0", 20 | "AO1", 21 | "AO2", 22 | "AW", 23 | "AW0", 24 | "AW1", 25 | "AW2", 26 | "AY", 27 | "AY0", 28 | "AY1", 29 | "AY2", 30 | "B", 31 | "CH", 32 | "D", 33 | "DH", 34 | "EH", 35 | "EH0", 36 | "EH1", 37 | "EH2", 38 | "ER", 39 | "ER0", 40 | "ER1", 41 | "ER2", 42 | "EY", 43 | "EY0", 44 | "EY1", 45 | "EY2", 46 | "F", 47 | "G", 48 | "HH", 49 | "IH", 50 | "IH0", 51 | "IH1", 52 | "IH2", 53 | "IY", 54 | "IY0", 55 | "IY1", 56 | "IY2", 57 | "JH", 58 | "K", 59 | "L", 60 | "M", 61 | "N", 62 | "NG", 63 | "OW", 64 | "OW0", 65 | "OW1", 66 | "OW2", 67 | "OY", 68 | "OY0", 69 | "OY1", 70 | "OY2", 71 | "P", 72 | "R", 73 | "S", 74 | "SH", 75 | "T", 76 | "TH", 77 | "UH", 78 | "UH0", 79 | "UH1", 80 | "UH2", 81 | "UW", 82 | "UW0", 83 | "UW1", 84 | "UW2", 85 | "V", 86 | "W", 87 | "Y", 88 | "Z", 89 | "ZH", 90 | ] 91 | 92 | 93 | class CMUDict: 94 | """Thin wrapper around CMUDict data. http://www.speech.cs.cmu.edu/cgi-bin/cmudict""" 95 | 96 | def __init__(self, file_or_path, keep_ambiguous=True): 97 | if isinstance(file_or_path, str): 98 | with open(file_or_path, encoding="latin-1") as f: 99 | entries = _parse_cmudict(f) 100 | else: 101 | entries = _parse_cmudict(file_or_path) 102 | if not keep_ambiguous: 103 | entries = {word: pron for word, pron in entries.items() if len(pron) == 1} 104 | self._entries = entries 105 | 106 | def __len__(self): 107 | return len(self._entries) 108 | 109 | def lookup(self, word): 110 | """Returns list of ARPAbet pronunciations of the given word.""" 111 | return self._entries.get(word.upper()) 112 | 113 | @staticmethod 114 | def get_arpabet(word, cmudict, punctuation_symbols): 115 | first_symbol, last_symbol = "", "" 116 | if word and word[0] in punctuation_symbols: 117 | first_symbol = word[0] 118 | word = word[1:] 119 | if word and word[-1] in punctuation_symbols: 120 | last_symbol = word[-1] 121 | word = word[:-1] 122 | arpabet = cmudict.lookup(word) 123 | if arpabet is not None: 124 | return first_symbol + "{%s}" % arpabet[0] + last_symbol 125 | return first_symbol + word + last_symbol 126 | 127 | 128 | _alt_re = re.compile(r"\([0-9]+\)") 129 | 130 | 131 | def _parse_cmudict(file): 132 | cmudict = {} 133 | for line in file: 134 | if line and (line[0] >= "A" and line[0] <= "Z" or line[0] == "'"): 135 | parts = line.split(" ") 136 | word = re.sub(_alt_re, "", parts[0]) 137 | pronunciation = _get_pronunciation(parts[1]) 138 | if pronunciation: 139 | if word in cmudict: 140 | cmudict[word].append(pronunciation) 141 | else: 142 | cmudict[word] = [pronunciation] 143 | return cmudict 144 | 145 | 146 | def _get_pronunciation(s): 147 | parts = s.strip().split(" ") 148 | for part in parts: 149 | if part not in VALID_SYMBOLS: 150 | return None 151 | return " ".join(parts) 152 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/english/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/text/english/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/text/english/abbreviations.py: -------------------------------------------------------------------------------- 1 | import re 2 | 3 | # List of (regular expression, replacement) pairs for abbreviations in english: 4 | abbreviations_en = [ 5 | (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1]) 6 | for x in [ 7 | ("mrs", "misess"), 8 | ("mr", "mister"), 9 | ("dr", "doctor"), 10 | ("st", "saint"), 11 | ("co", "company"), 12 | ("jr", "junior"), 13 | ("maj", "major"), 14 | ("gen", "general"), 15 | ("drs", "doctors"), 16 | ("rev", "reverend"), 17 | ("lt", "lieutenant"), 18 | ("hon", "honorable"), 19 | ("sgt", "sergeant"), 20 | ("capt", "captain"), 21 | ("esq", "esquire"), 22 | ("ltd", "limited"), 23 | ("col", "colonel"), 24 | ("ft", "fort"), 25 | ] 26 | ] 27 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/english/number_norm.py: -------------------------------------------------------------------------------- 1 | """ from https://github.com/keithito/tacotron """ 2 | 3 | import re 4 | from typing import Dict 5 | 6 | import inflect 7 | 8 | _inflect = inflect.engine() 9 | _comma_number_re = re.compile(r"([0-9][0-9\,]+[0-9])") 10 | _decimal_number_re = re.compile(r"([0-9]+\.[0-9]+)") 11 | _currency_re = re.compile(r"(£|\$|¥)([0-9\,\.]*[0-9]+)") 12 | _ordinal_re = re.compile(r"[0-9]+(st|nd|rd|th)") 13 | _number_re = re.compile(r"-?[0-9]+") 14 | 15 | 16 | def _remove_commas(m): 17 | return m.group(1).replace(",", "") 18 | 19 | 20 | def _expand_decimal_point(m): 21 | return m.group(1).replace(".", " point ") 22 | 23 | 24 | def __expand_currency(value: str, inflection: Dict[float, str]) -> str: 25 | parts = value.replace(",", "").split(".") 26 | if len(parts) > 2: 27 | return f"{value} {inflection[2]}" # Unexpected format 28 | text = [] 29 | integer = int(parts[0]) if parts[0] else 0 30 | if integer > 0: 31 | integer_unit = inflection.get(integer, inflection[2]) 32 | text.append(f"{integer} {integer_unit}") 33 | fraction = int(parts[1]) if len(parts) > 1 and parts[1] else 0 34 | if fraction > 0: 35 | fraction_unit = inflection.get(fraction / 100, inflection[0.02]) 36 | text.append(f"{fraction} {fraction_unit}") 37 | if len(text) == 0: 38 | return f"zero {inflection[2]}" 39 | return " ".join(text) 40 | 41 | 42 | def _expand_currency(m: "re.Match") -> str: 43 | currencies = { 44 | "$": { 45 | 0.01: "cent", 46 | 0.02: "cents", 47 | 1: "dollar", 48 | 2: "dollars", 49 | }, 50 | "€": { 51 | 0.01: "cent", 52 | 0.02: "cents", 53 | 1: "euro", 54 | 2: "euros", 55 | }, 56 | "£": { 57 | 0.01: "penny", 58 | 0.02: "pence", 59 | 1: "pound sterling", 60 | 2: "pounds sterling", 61 | }, 62 | "¥": { 63 | # TODO rin 64 | 0.02: "sen", 65 | 2: "yen", 66 | }, 67 | } 68 | unit = m.group(1) 69 | currency = currencies[unit] 70 | value = m.group(2) 71 | return __expand_currency(value, currency) 72 | 73 | 74 | def _expand_ordinal(m): 75 | return _inflect.number_to_words(m.group(0)) 76 | 77 | 78 | def _expand_number(m): 79 | num = int(m.group(0)) 80 | if 1000 < num < 3000: 81 | if num == 2000: 82 | return "two thousand" 83 | if 2000 < num < 2010: 84 | return "two thousand " + _inflect.number_to_words(num % 100) 85 | if num % 100 == 0: 86 | return _inflect.number_to_words(num // 100) + " hundred" 87 | return _inflect.number_to_words(num, andword="", zero="oh", group=2).replace(", ", " ") 88 | return _inflect.number_to_words(num, andword="") 89 | 90 | 91 | def normalize_numbers(text): 92 | text = re.sub(_comma_number_re, _remove_commas, text) 93 | text = re.sub(_currency_re, _expand_currency, text) 94 | text = re.sub(_decimal_number_re, _expand_decimal_point, text) 95 | text = re.sub(_ordinal_re, _expand_ordinal, text) 96 | text = re.sub(_number_re, _expand_number, text) 97 | return text 98 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/english/time_norm.py: -------------------------------------------------------------------------------- 1 | import re 2 | 3 | import inflect 4 | 5 | _inflect = inflect.engine() 6 | 7 | _time_re = re.compile( 8 | r"""\b 9 | ((0?[0-9])|(1[0-1])|(1[2-9])|(2[0-3])) # hours 10 | : 11 | ([0-5][0-9]) # minutes 12 | \s*(a\\.m\\.|am|pm|p\\.m\\.|a\\.m|p\\.m)? # am/pm 13 | \b""", 14 | re.IGNORECASE | re.X, 15 | ) 16 | 17 | 18 | def _expand_num(n: int) -> str: 19 | return _inflect.number_to_words(n) 20 | 21 | 22 | def _expand_time_english(match: "re.Match") -> str: 23 | hour = int(match.group(1)) 24 | past_noon = hour >= 12 25 | time = [] 26 | if hour > 12: 27 | hour -= 12 28 | elif hour == 0: 29 | hour = 12 30 | past_noon = True 31 | time.append(_expand_num(hour)) 32 | 33 | minute = int(match.group(6)) 34 | if minute > 0: 35 | if minute < 10: 36 | time.append("oh") 37 | time.append(_expand_num(minute)) 38 | am_pm = match.group(7) 39 | if am_pm is None: 40 | time.append("p m" if past_noon else "a m") 41 | else: 42 | time.extend(list(am_pm.replace(".", ""))) 43 | return " ".join(time) 44 | 45 | 46 | def expand_time_english(text: str) -> str: 47 | return re.sub(_time_re, _expand_time_english, text) 48 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/french/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/text/french/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/text/french/abbreviations.py: -------------------------------------------------------------------------------- 1 | import re 2 | 3 | # List of (regular expression, replacement) pairs for abbreviations in french: 4 | abbreviations_fr = [ 5 | (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1]) 6 | for x in [ 7 | ("M", "monsieur"), 8 | ("Mlle", "mademoiselle"), 9 | ("Mlles", "mesdemoiselles"), 10 | ("Mme", "Madame"), 11 | ("Mmes", "Mesdames"), 12 | ("N.B", "nota bene"), 13 | ("M", "monsieur"), 14 | ("p.c.q", "parce que"), 15 | ("Pr", "professeur"), 16 | ("qqch", "quelque chose"), 17 | ("rdv", "rendez-vous"), 18 | ("max", "maximum"), 19 | ("min", "minimum"), 20 | ("no", "numéro"), 21 | ("adr", "adresse"), 22 | ("dr", "docteur"), 23 | ("st", "saint"), 24 | ("co", "companie"), 25 | ("jr", "junior"), 26 | ("sgt", "sergent"), 27 | ("capt", "capitain"), 28 | ("col", "colonel"), 29 | ("av", "avenue"), 30 | ("av. J.-C", "avant Jésus-Christ"), 31 | ("apr. J.-C", "après Jésus-Christ"), 32 | ("art", "article"), 33 | ("boul", "boulevard"), 34 | ("c.-à-d", "c’est-à-dire"), 35 | ("etc", "et cetera"), 36 | ("ex", "exemple"), 37 | ("excl", "exclusivement"), 38 | ("boul", "boulevard"), 39 | ] 40 | ] + [ 41 | (re.compile("\\b%s" % x[0]), x[1]) 42 | for x in [ 43 | ("Mlle", "mademoiselle"), 44 | ("Mlles", "mesdemoiselles"), 45 | ("Mme", "Madame"), 46 | ("Mmes", "Mesdames"), 47 | ] 48 | ] 49 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/japanese/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/text/japanese/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/text/korean/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/tts/utils/text/korean/__init__.py -------------------------------------------------------------------------------- /TTS/tts/utils/text/korean/ko_dictionary.py: -------------------------------------------------------------------------------- 1 | # coding: utf-8 2 | # Add the word you want to the dictionary. 3 | etc_dictionary = {"1+1": "원플러스원", "2+1": "투플러스원"} 4 | 5 | 6 | english_dictionary = { 7 | "KOREA": "코리아", 8 | "IDOL": "아이돌", 9 | "IT": "아이티", 10 | "IQ": "아이큐", 11 | "UP": "업", 12 | "DOWN": "다운", 13 | "PC": "피씨", 14 | "CCTV": "씨씨티비", 15 | "SNS": "에스엔에스", 16 | "AI": "에이아이", 17 | "CEO": "씨이오", 18 | "A": "에이", 19 | "B": "비", 20 | "C": "씨", 21 | "D": "디", 22 | "E": "이", 23 | "F": "에프", 24 | "G": "지", 25 | "H": "에이치", 26 | "I": "아이", 27 | "J": "제이", 28 | "K": "케이", 29 | "L": "엘", 30 | "M": "엠", 31 | "N": "엔", 32 | "O": "오", 33 | "P": "피", 34 | "Q": "큐", 35 | "R": "알", 36 | "S": "에스", 37 | "T": "티", 38 | "U": "유", 39 | "V": "브이", 40 | "W": "더블유", 41 | "X": "엑스", 42 | "Y": "와이", 43 | "Z": "제트", 44 | } 45 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/korean/korean.py: -------------------------------------------------------------------------------- 1 | # coding: utf-8 2 | # Code based on https://github.com/carpedm20/multi-speaker-tacotron-tensorflow/blob/master/text/korean.py 3 | import re 4 | 5 | from TTS.tts.utils.text.korean.ko_dictionary import english_dictionary, etc_dictionary 6 | 7 | 8 | def normalize(text): 9 | text = text.strip() 10 | text = re.sub("[⺀-⺙⺛-⻳⼀-⿕々〇〡-〩〸-〺〻㐀-䶵一-鿃豈-鶴侮-頻並-龎]", "", text) 11 | text = normalize_with_dictionary(text, etc_dictionary) 12 | text = normalize_english(text) 13 | text = text.lower() 14 | return text 15 | 16 | 17 | def normalize_with_dictionary(text, dic): 18 | if any(key in text for key in dic.keys()): 19 | pattern = re.compile("|".join(re.escape(key) for key in dic.keys())) 20 | return pattern.sub(lambda x: dic[x.group()], text) 21 | return text 22 | 23 | 24 | def normalize_english(text): 25 | def fn(m): 26 | word = m.group() 27 | if word in english_dictionary: 28 | return english_dictionary.get(word) 29 | return word 30 | 31 | text = re.sub("([A-Za-z]+)", fn, text) 32 | return text 33 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/korean/phonemizer.py: -------------------------------------------------------------------------------- 1 | from jamo import hangul_to_jamo 2 | 3 | from TTS.tts.utils.text.korean.korean import normalize 4 | 5 | g2p = None 6 | 7 | 8 | def korean_text_to_phonemes(text, character: str = "hangeul") -> str: 9 | """ 10 | 11 | The input and output values look the same, but they are different in Unicode. 12 | 13 | example : 14 | 15 | input = '하늘' (Unicode : \ud558\ub298), (하 + 늘) 16 | output = '하늘' (Unicode :\u1112\u1161\u1102\u1173\u11af), (ᄒ + ᅡ + ᄂ + ᅳ + ᆯ) 17 | 18 | """ 19 | global g2p # pylint: disable=global-statement 20 | if g2p is None: 21 | from g2pkk import G2p 22 | 23 | g2p = G2p() 24 | 25 | if character == "english": 26 | from anyascii import anyascii 27 | 28 | text = normalize(text) 29 | text = g2p(text) 30 | text = anyascii(text) 31 | return text 32 | 33 | text = normalize(text) 34 | text = g2p(text) 35 | text = list(hangul_to_jamo(text)) # '하늘' --> ['ᄒ', 'ᅡ', 'ᄂ', 'ᅳ', 'ᆯ'] 36 | return "".join(text) 37 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/phonemizers/__init__.py: -------------------------------------------------------------------------------- 1 | from TTS.tts.utils.text.phonemizers.base import BasePhonemizer 2 | from TTS.tts.utils.text.phonemizers.espeak_wrapper import ESpeak 3 | from TTS.tts.utils.text.phonemizers.gruut_wrapper import Gruut 4 | from TTS.tts.utils.text.phonemizers.ja_jp_phonemizer import JA_JP_Phonemizer 5 | from TTS.tts.utils.text.phonemizers.ko_kr_phonemizer import KO_KR_Phonemizer 6 | from TTS.tts.utils.text.phonemizers.zh_cn_phonemizer import ZH_CN_Phonemizer 7 | 8 | PHONEMIZERS = {b.name(): b for b in (ESpeak, Gruut, JA_JP_Phonemizer)} 9 | 10 | 11 | ESPEAK_LANGS = list(ESpeak.supported_languages().keys()) 12 | GRUUT_LANGS = list(Gruut.supported_languages()) 13 | 14 | 15 | # Dict setting default phonemizers for each language 16 | # Add Gruut languages 17 | _ = [Gruut.name()] * len(GRUUT_LANGS) 18 | DEF_LANG_TO_PHONEMIZER = dict(list(zip(GRUUT_LANGS, _))) 19 | 20 | 21 | # Add ESpeak languages and override any existing ones 22 | _ = [ESpeak.name()] * len(ESPEAK_LANGS) 23 | _new_dict = dict(list(zip(list(ESPEAK_LANGS), _))) 24 | DEF_LANG_TO_PHONEMIZER.update(_new_dict) 25 | 26 | # Force default for some languages 27 | DEF_LANG_TO_PHONEMIZER["en"] = DEF_LANG_TO_PHONEMIZER["en-us"] 28 | DEF_LANG_TO_PHONEMIZER["ja-jp"] = JA_JP_Phonemizer.name() 29 | DEF_LANG_TO_PHONEMIZER["zh-cn"] = ZH_CN_Phonemizer.name() 30 | DEF_LANG_TO_PHONEMIZER["ko-kr"] = KO_KR_Phonemizer.name() 31 | 32 | 33 | def get_phonemizer_by_name(name: str, **kwargs) -> BasePhonemizer: 34 | """Initiate a phonemizer by name 35 | 36 | Args: 37 | name (str): 38 | Name of the phonemizer that should match `phonemizer.name()`. 39 | 40 | kwargs (dict): 41 | Extra keyword arguments that should be passed to the phonemizer. 42 | """ 43 | if name == "espeak": 44 | return ESpeak(**kwargs) 45 | if name == "gruut": 46 | return Gruut(**kwargs) 47 | if name == "zh_cn_phonemizer": 48 | return ZH_CN_Phonemizer(**kwargs) 49 | if name == "ja_jp_phonemizer": 50 | return JA_JP_Phonemizer(**kwargs) 51 | if name == "ko_kr_phonemizer": 52 | return KO_KR_Phonemizer(**kwargs) 53 | raise ValueError(f"Phonemizer {name} not found") 54 | 55 | 56 | if __name__ == "__main__": 57 | print(DEF_LANG_TO_PHONEMIZER) 58 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/phonemizers/ja_jp_phonemizer.py: -------------------------------------------------------------------------------- 1 | from typing import Dict 2 | 3 | from TTS.tts.utils.text.japanese.phonemizer import japanese_text_to_phonemes 4 | from TTS.tts.utils.text.phonemizers.base import BasePhonemizer 5 | 6 | _DEF_JA_PUNCS = "、.,[]()?!〽~『』「」【】" 7 | 8 | _TRANS_TABLE = {"、": ","} 9 | 10 | 11 | def trans(text): 12 | for i, j in _TRANS_TABLE.items(): 13 | text = text.replace(i, j) 14 | return text 15 | 16 | 17 | class JA_JP_Phonemizer(BasePhonemizer): 18 | """🐸TTS Ja-Jp phonemizer using functions in `TTS.tts.utils.text.japanese.phonemizer` 19 | 20 | TODO: someone with JA knowledge should check this implementation 21 | 22 | Example: 23 | 24 | >>> from TTS.tts.utils.text.phonemizers import JA_JP_Phonemizer 25 | >>> phonemizer = JA_JP_Phonemizer() 26 | >>> phonemizer.phonemize("どちらに行きますか？", separator="|") 27 | 'd|o|c|h|i|r|a|n|i|i|k|i|m|a|s|u|k|a|?' 28 | 29 | """ 30 | 31 | language = "ja-jp" 32 | 33 | def __init__(self, punctuations=_DEF_JA_PUNCS, keep_puncs=True, **kwargs): # pylint: disable=unused-argument 34 | super().__init__(self.language, punctuations=punctuations, keep_puncs=keep_puncs) 35 | 36 | @staticmethod 37 | def name(): 38 | return "ja_jp_phonemizer" 39 | 40 | def _phonemize(self, text: str, separator: str = "|") -> str: 41 | ph = japanese_text_to_phonemes(text) 42 | if separator is not None or separator != "": 43 | return separator.join(ph) 44 | return ph 45 | 46 | def phonemize(self, text: str, separator="|", language=None) -> str: 47 | """Custom phonemize for JP_JA 48 | 49 | Skip pre-post processing steps used by the other phonemizers. 50 | """ 51 | return self._phonemize(text, separator) 52 | 53 | @staticmethod 54 | def supported_languages() -> Dict: 55 | return {"ja-jp": "Japanese (Japan)"} 56 | 57 | def version(self) -> str: 58 | return "0.0.1" 59 | 60 | def is_available(self) -> bool: 61 | return True 62 | 63 | 64 | # if __name__ == "__main__": 65 | # text = "これは、電話をかけるための私の日本語の例のテキストです。" 66 | # e = JA_JP_Phonemizer() 67 | # print(e.supported_languages()) 68 | # print(e.version()) 69 | # print(e.language) 70 | # print(e.name()) 71 | # print(e.is_available()) 72 | # print("`" + e.phonemize(text) + "`") 73 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/phonemizers/ko_kr_phonemizer.py: -------------------------------------------------------------------------------- 1 | from typing import Dict 2 | 3 | from TTS.tts.utils.text.korean.phonemizer import korean_text_to_phonemes 4 | from TTS.tts.utils.text.phonemizers.base import BasePhonemizer 5 | 6 | _DEF_KO_PUNCS = "、.,[]()?!〽~『』「」【】" 7 | 8 | 9 | class KO_KR_Phonemizer(BasePhonemizer): 10 | """🐸TTS ko_kr_phonemizer using functions in `TTS.tts.utils.text.korean.phonemizer` 11 | 12 | TODO: Add Korean to character (ᄀᄁᄂᄃᄄᄅᄆᄇᄈᄉᄊᄋᄌᄍᄎᄏᄐᄑ하ᅢᅣᅤᅥᅦᅧᅨᅩᅪᅫᅬᅭᅮᅯᅰᅱᅲᅳᅴᅵᆨᆩᆪᆫᆬᆭᆮᆯᆰᆱᆲᆳᆴᆵᆶᆷᆸᆹᆺᆻᆼᆽᆾᆿᇀᇁᇂ) 13 | 14 | Example: 15 | 16 | >>> from TTS.tts.utils.text.phonemizers import KO_KR_Phonemizer 17 | >>> phonemizer = KO_KR_Phonemizer() 18 | >>> phonemizer.phonemize("이 문장은 음성합성 테스트를 위한 문장입니다.", separator="|") 19 | 'ᄋ|ᅵ| |ᄆ|ᅮ|ᆫ|ᄌ|ᅡ|ᆼ|ᄋ|ᅳ| |ᄂ|ᅳ|ᆷ|ᄉ|ᅥ|ᆼ|ᄒ|ᅡ|ᆸ|ᄊ|ᅥ|ᆼ| |ᄐ|ᅦ|ᄉ|ᅳ|ᄐ|ᅳ|ᄅ|ᅳ| |ᄅ|ᅱ|ᄒ|ᅡ|ᆫ| |ᄆ|ᅮ|ᆫ|ᄌ|ᅡ|ᆼ|ᄋ|ᅵ|ᆷ|ᄂ|ᅵ|ᄃ|ᅡ|.' 20 | 21 | >>> from TTS.tts.utils.text.phonemizers import KO_KR_Phonemizer 22 | >>> phonemizer = KO_KR_Phonemizer() 23 | >>> phonemizer.phonemize("이 문장은 음성합성 테스트를 위한 문장입니다.", separator="|", character='english') 24 | 'I| |M|u|n|J|a|n|g|E|u| |N|e|u|m|S|e|o|n|g|H|a|b|S|s|e|o|n|g| |T|e|S|e|u|T|e|u|L|e|u| |L|w|i|H|a|n| |M|u|n|J|a|n|g|I|m|N|i|D|a|.' 25 | 26 | """ 27 | 28 | language = "ko-kr" 29 | 30 | def __init__(self, punctuations=_DEF_KO_PUNCS, keep_puncs=True, **kwargs): # pylint: disable=unused-argument 31 | super().__init__(self.language, punctuations=punctuations, keep_puncs=keep_puncs) 32 | 33 | @staticmethod 34 | def name(): 35 | return "ko_kr_phonemizer" 36 | 37 | def _phonemize(self, text: str, separator: str = "", character: str = "hangeul") -> str: 38 | ph = korean_text_to_phonemes(text, character=character) 39 | if separator is not None or separator != "": 40 | return separator.join(ph) 41 | return ph 42 | 43 | def phonemize(self, text: str, separator: str = "", character: str = "hangeul", language=None) -> str: 44 | return self._phonemize(text, separator, character) 45 | 46 | @staticmethod 47 | def supported_languages() -> Dict: 48 | return {"ko-kr": "hangeul(korean)"} 49 | 50 | def version(self) -> str: 51 | return "0.0.2" 52 | 53 | def is_available(self) -> bool: 54 | return True 55 | 56 | 57 | if __name__ == "__main__": 58 | texts = "이 문장은 음성합성 테스트를 위한 문장입니다." 59 | e = KO_KR_Phonemizer() 60 | print(e.supported_languages()) 61 | print(e.version()) 62 | print(e.language) 63 | print(e.name()) 64 | print(e.is_available()) 65 | print(e.phonemize(texts)) 66 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/phonemizers/multi_phonemizer.py: -------------------------------------------------------------------------------- 1 | from typing import Dict, List 2 | 3 | from TTS.tts.utils.text.phonemizers import DEF_LANG_TO_PHONEMIZER, get_phonemizer_by_name 4 | 5 | 6 | class MultiPhonemizer: 7 | """🐸TTS multi-phonemizer that operates phonemizers for multiple langugages 8 | 9 | Args: 10 | custom_lang_to_phonemizer (Dict): 11 | Custom phonemizer mapping if you want to change the defaults. In the format of 12 | `{"lang_code", "phonemizer_name"}`. When it is None, `DEF_LANG_TO_PHONEMIZER` is used. Defaults to `{}`. 13 | 14 | TODO: find a way to pass custom kwargs to the phonemizers 15 | """ 16 | 17 | lang_to_phonemizer = {} 18 | 19 | def __init__(self, lang_to_phonemizer_name: Dict = {}) -> None: # pylint: disable=dangerous-default-value 20 | for k, v in lang_to_phonemizer_name.items(): 21 | if v == "" and k in DEF_LANG_TO_PHONEMIZER.keys(): 22 | lang_to_phonemizer_name[k] = DEF_LANG_TO_PHONEMIZER[k] 23 | elif v == "": 24 | raise ValueError(f"Phonemizer wasn't set for language {k} and doesn't have a default.") 25 | self.lang_to_phonemizer_name = lang_to_phonemizer_name 26 | self.lang_to_phonemizer = self.init_phonemizers(self.lang_to_phonemizer_name) 27 | 28 | @staticmethod 29 | def init_phonemizers(lang_to_phonemizer_name: Dict) -> Dict: 30 | lang_to_phonemizer = {} 31 | for k, v in lang_to_phonemizer_name.items(): 32 | lang_to_phonemizer[k] = get_phonemizer_by_name(v, language=k) 33 | return lang_to_phonemizer 34 | 35 | @staticmethod 36 | def name(): 37 | return "multi-phonemizer" 38 | 39 | def phonemize(self, text, separator="|", language=""): 40 | if language == "": 41 | raise ValueError("Language must be set for multi-phonemizer to phonemize.") 42 | return self.lang_to_phonemizer[language].phonemize(text, separator) 43 | 44 | def supported_languages(self) -> List: 45 | return list(self.lang_to_phonemizer.keys()) 46 | 47 | def print_logs(self, level: int = 0): 48 | indent = "\t" * level 49 | print(f"{indent}| > phoneme language: {self.supported_languages()}") 50 | print(f"{indent}| > phoneme backend: {self.name()}") 51 | 52 | 53 | # if __name__ == "__main__": 54 | # texts = { 55 | # "tr": "Merhaba, bu Türkçe bit örnek!", 56 | # "en-us": "Hello, this is English example!", 57 | # "de": "Hallo, das ist ein Deutches Beipiel!", 58 | # "zh-cn": "这是中国的例子", 59 | # } 60 | # phonemes = {} 61 | # ph = MultiPhonemizer({"tr": "espeak", "en-us": "", "de": "gruut", "zh-cn": ""}) 62 | # for lang, text in texts.items(): 63 | # phoneme = ph.phonemize(text, lang) 64 | # phonemes[lang] = phoneme 65 | # print(phonemes) 66 | -------------------------------------------------------------------------------- /TTS/tts/utils/text/phonemizers/zh_cn_phonemizer.py: -------------------------------------------------------------------------------- 1 | from typing import Dict 2 | 3 | from TTS.tts.utils.text.chinese_mandarin.phonemizer import chinese_text_to_phonemes 4 | from TTS.tts.utils.text.phonemizers.base import BasePhonemizer 5 | 6 | _DEF_ZH_PUNCS = "、.,[]()?!〽~『』「」【】" 7 | 8 | 9 | class ZH_CN_Phonemizer(BasePhonemizer): 10 | """🐸TTS Zh-Cn phonemizer using functions in `TTS.tts.utils.text.chinese_mandarin.phonemizer` 11 | 12 | Args: 13 | punctuations (str): 14 | Set of characters to be treated as punctuation. Defaults to `_DEF_ZH_PUNCS`. 15 | 16 | keep_puncs (bool): 17 | If True, keep the punctuations after phonemization. Defaults to False. 18 | 19 | Example :: 20 | 21 | "这是，样本中文。" -> `d|ʒ|ø|4| |ʂ|ʏ|4| |，| |i|ɑ|ŋ|4|b|œ|n|3| |d|ʒ|o|ŋ|1|w|œ|n|2| |。` 22 | 23 | TODO: someone with Mandarin knowledge should check this implementation 24 | """ 25 | 26 | language = "zh-cn" 27 | 28 | def __init__(self, punctuations=_DEF_ZH_PUNCS, keep_puncs=False, **kwargs): # pylint: disable=unused-argument 29 | super().__init__(self.language, punctuations=punctuations, keep_puncs=keep_puncs) 30 | 31 | @staticmethod 32 | def name(): 33 | return "zh_cn_phonemizer" 34 | 35 | @staticmethod 36 | def phonemize_zh_cn(text: str, separator: str = "|") -> str: 37 | ph = chinese_text_to_phonemes(text, separator) 38 | return ph 39 | 40 | def _phonemize(self, text, separator): 41 | return self.phonemize_zh_cn(text, separator) 42 | 43 | @staticmethod 44 | def supported_languages() -> Dict: 45 | return {"zh-cn": "Chinese (China)"} 46 | 47 | def version(self) -> str: 48 | return "0.0.1" 49 | 50 | def is_available(self) -> bool: 51 | return True 52 | 53 | 54 | # if __name__ == "__main__": 55 | # text = "这是，样本中文。" 56 | # e = ZH_CN_Phonemizer() 57 | # print(e.supported_languages()) 58 | # print(e.version()) 59 | # print(e.language) 60 | # print(e.name()) 61 | # print(e.is_available()) 62 | # print("`" + e.phonemize(text) + "`") 63 | -------------------------------------------------------------------------------- /TTS/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/utils/__init__.py -------------------------------------------------------------------------------- /TTS/utils/audio/__init__.py: -------------------------------------------------------------------------------- 1 | from TTS.utils.audio.processor import AudioProcessor 2 | -------------------------------------------------------------------------------- /TTS/utils/callbacks.py: -------------------------------------------------------------------------------- 1 | class TrainerCallback: 2 | @staticmethod 3 | def on_init_start(trainer) -> None: 4 | if hasattr(trainer.model, "module"): 5 | if hasattr(trainer.model.module, "on_init_start"): 6 | trainer.model.module.on_init_start(trainer) 7 | else: 8 | if hasattr(trainer.model, "on_init_start"): 9 | trainer.model.on_init_start(trainer) 10 | 11 | if hasattr(trainer.criterion, "on_init_start"): 12 | trainer.criterion.on_init_start(trainer) 13 | 14 | if hasattr(trainer.optimizer, "on_init_start"): 15 | trainer.optimizer.on_init_start(trainer) 16 | 17 | @staticmethod 18 | def on_init_end(trainer) -> None: 19 | if hasattr(trainer.model, "module"): 20 | if hasattr(trainer.model.module, "on_init_end"): 21 | trainer.model.module.on_init_end(trainer) 22 | else: 23 | if hasattr(trainer.model, "on_init_end"): 24 | trainer.model.on_init_end(trainer) 25 | 26 | if hasattr(trainer.criterion, "on_init_end"): 27 | trainer.criterion.on_init_end(trainer) 28 | 29 | if hasattr(trainer.optimizer, "on_init_end"): 30 | trainer.optimizer.on_init_end(trainer) 31 | 32 | @staticmethod 33 | def on_epoch_start(trainer) -> None: 34 | if hasattr(trainer.model, "module"): 35 | if hasattr(trainer.model.module, "on_epoch_start"): 36 | trainer.model.module.on_epoch_start(trainer) 37 | else: 38 | if hasattr(trainer.model, "on_epoch_start"): 39 | trainer.model.on_epoch_start(trainer) 40 | 41 | if hasattr(trainer.criterion, "on_epoch_start"): 42 | trainer.criterion.on_epoch_start(trainer) 43 | 44 | if hasattr(trainer.optimizer, "on_epoch_start"): 45 | trainer.optimizer.on_epoch_start(trainer) 46 | 47 | @staticmethod 48 | def on_epoch_end(trainer) -> None: 49 | if hasattr(trainer.model, "module"): 50 | if hasattr(trainer.model.module, "on_epoch_end"): 51 | trainer.model.module.on_epoch_end(trainer) 52 | else: 53 | if hasattr(trainer.model, "on_epoch_end"): 54 | trainer.model.on_epoch_end(trainer) 55 | 56 | if hasattr(trainer.criterion, "on_epoch_end"): 57 | trainer.criterion.on_epoch_end(trainer) 58 | 59 | if hasattr(trainer.optimizer, "on_epoch_end"): 60 | trainer.optimizer.on_epoch_end(trainer) 61 | 62 | @staticmethod 63 | def on_train_step_start(trainer) -> None: 64 | if hasattr(trainer.model, "module"): 65 | if hasattr(trainer.model.module, "on_train_step_start"): 66 | trainer.model.module.on_train_step_start(trainer) 67 | else: 68 | if hasattr(trainer.model, "on_train_step_start"): 69 | trainer.model.on_train_step_start(trainer) 70 | 71 | if hasattr(trainer.criterion, "on_train_step_start"): 72 | trainer.criterion.on_train_step_start(trainer) 73 | 74 | if hasattr(trainer.optimizer, "on_train_step_start"): 75 | trainer.optimizer.on_train_step_start(trainer) 76 | 77 | @staticmethod 78 | def on_train_step_end(trainer) -> None: 79 | if hasattr(trainer.model, "module"): 80 | if hasattr(trainer.model.module, "on_train_step_end"): 81 | trainer.model.module.on_train_step_end(trainer) 82 | else: 83 | if hasattr(trainer.model, "on_train_step_end"): 84 | trainer.model.on_train_step_end(trainer) 85 | 86 | if hasattr(trainer.criterion, "on_train_step_end"): 87 | trainer.criterion.on_train_step_end(trainer) 88 | 89 | if hasattr(trainer.optimizer, "on_train_step_end"): 90 | trainer.optimizer.on_train_step_end(trainer) 91 | 92 | @staticmethod 93 | def on_keyboard_interrupt(trainer) -> None: 94 | if hasattr(trainer.model, "module"): 95 | if hasattr(trainer.model.module, "on_keyboard_interrupt"): 96 | trainer.model.module.on_keyboard_interrupt(trainer) 97 | else: 98 | if hasattr(trainer.model, "on_keyboard_interrupt"): 99 | trainer.model.on_keyboard_interrupt(trainer) 100 | 101 | if hasattr(trainer.criterion, "on_keyboard_interrupt"): 102 | trainer.criterion.on_keyboard_interrupt(trainer) 103 | 104 | if hasattr(trainer.optimizer, "on_keyboard_interrupt"): 105 | trainer.optimizer.on_keyboard_interrupt(trainer) 106 | -------------------------------------------------------------------------------- /TTS/utils/capacitron_optimizer.py: -------------------------------------------------------------------------------- 1 | from typing import Generator 2 | 3 | from trainer.trainer_utils import get_optimizer 4 | 5 | 6 | class CapacitronOptimizer: 7 | """Double optimizer class for the Capacitron model.""" 8 | 9 | def __init__(self, config: dict, model_params: Generator) -> None: 10 | self.primary_params, self.secondary_params = self.split_model_parameters(model_params) 11 | 12 | optimizer_names = list(config.optimizer_params.keys()) 13 | optimizer_parameters = list(config.optimizer_params.values()) 14 | 15 | self.primary_optimizer = get_optimizer( 16 | optimizer_names[0], 17 | optimizer_parameters[0], 18 | config.lr, 19 | parameters=self.primary_params, 20 | ) 21 | 22 | self.secondary_optimizer = get_optimizer( 23 | optimizer_names[1], 24 | self.extract_optimizer_parameters(optimizer_parameters[1]), 25 | optimizer_parameters[1]["lr"], 26 | parameters=self.secondary_params, 27 | ) 28 | 29 | self.param_groups = self.primary_optimizer.param_groups 30 | 31 | def first_step(self): 32 | self.secondary_optimizer.step() 33 | self.secondary_optimizer.zero_grad() 34 | self.primary_optimizer.zero_grad() 35 | 36 | def step(self): 37 | # Update param groups to display the correct learning rate 38 | self.param_groups = self.primary_optimizer.param_groups 39 | self.primary_optimizer.step() 40 | 41 | def zero_grad(self, set_to_none=False): 42 | self.primary_optimizer.zero_grad(set_to_none) 43 | self.secondary_optimizer.zero_grad(set_to_none) 44 | 45 | def load_state_dict(self, state_dict): 46 | self.primary_optimizer.load_state_dict(state_dict[0]) 47 | self.secondary_optimizer.load_state_dict(state_dict[1]) 48 | 49 | def state_dict(self): 50 | return [self.primary_optimizer.state_dict(), self.secondary_optimizer.state_dict()] 51 | 52 | @staticmethod 53 | def split_model_parameters(model_params: Generator) -> list: 54 | primary_params = [] 55 | secondary_params = [] 56 | for name, param in model_params: 57 | if param.requires_grad: 58 | if name == "capacitron_vae_layer.beta": 59 | secondary_params.append(param) 60 | else: 61 | primary_params.append(param) 62 | return [iter(primary_params), iter(secondary_params)] 63 | 64 | @staticmethod 65 | def extract_optimizer_parameters(params: dict) -> dict: 66 | """Extract parameters that are not the learning rate""" 67 | return {k: v for k, v in params.items() if k != "lr"} 68 | -------------------------------------------------------------------------------- /TTS/utils/distribute.py: -------------------------------------------------------------------------------- 1 | # edited from https://github.com/fastai/imagenet-fast/blob/master/imagenet_nv/distributed.py 2 | import torch 3 | import torch.distributed as dist 4 | 5 | 6 | def reduce_tensor(tensor, num_gpus): 7 | rt = tensor.clone() 8 | dist.all_reduce(rt, op=dist.reduce_op.SUM) 9 | rt /= num_gpus 10 | return rt 11 | 12 | 13 | def init_distributed(rank, num_gpus, group_name, dist_backend, dist_url): 14 | assert torch.cuda.is_available(), "Distributed mode requires CUDA." 15 | 16 | # Set cuda device so everything is done on the right GPU. 17 | torch.cuda.set_device(rank % torch.cuda.device_count()) 18 | 19 | # Initialize distributed communication 20 | dist.init_process_group(dist_backend, init_method=dist_url, world_size=num_gpus, rank=rank, group_name=group_name) 21 | -------------------------------------------------------------------------------- /TTS/utils/training.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | 5 | def check_update(model, grad_clip, ignore_stopnet=False, amp_opt_params=None): 6 | r"""Check model gradient against unexpected jumps and failures""" 7 | skip_flag = False 8 | if ignore_stopnet: 9 | if not amp_opt_params: 10 | grad_norm = torch.nn.utils.clip_grad_norm_( 11 | [param for name, param in model.named_parameters() if "stopnet" not in name], grad_clip 12 | ) 13 | else: 14 | grad_norm = torch.nn.utils.clip_grad_norm_(amp_opt_params, grad_clip) 15 | else: 16 | if not amp_opt_params: 17 | grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip) 18 | else: 19 | grad_norm = torch.nn.utils.clip_grad_norm_(amp_opt_params, grad_clip) 20 | 21 | # compatibility with different torch versions 22 | if isinstance(grad_norm, float): 23 | if np.isinf(grad_norm): 24 | print(" | > Gradient is INF !!") 25 | skip_flag = True 26 | else: 27 | if torch.isinf(grad_norm): 28 | print(" | > Gradient is INF !!") 29 | skip_flag = True 30 | return grad_norm, skip_flag 31 | 32 | 33 | def gradual_training_scheduler(global_step, config): 34 | """Setup the gradual training schedule wrt number 35 | of active GPUs""" 36 | num_gpus = torch.cuda.device_count() 37 | if num_gpus == 0: 38 | num_gpus = 1 39 | new_values = None 40 | # we set the scheduling wrt num_gpus 41 | for values in config.gradual_training: 42 | if global_step * num_gpus >= values[0]: 43 | new_values = values 44 | return new_values[1], new_values[2] 45 | -------------------------------------------------------------------------------- /TTS/utils/vad.py: -------------------------------------------------------------------------------- 1 | import soundfile as sf 2 | import torch 3 | import torchaudio 4 | 5 | 6 | def read_audio(path): 7 | wav, sr = torchaudio.load(path) 8 | 9 | if wav.size(0) > 1: 10 | wav = wav.mean(dim=0, keepdim=True) 11 | 12 | return wav.squeeze(0), sr 13 | 14 | 15 | def resample_wav(wav, sr, new_sr): 16 | wav = wav.unsqueeze(0) 17 | transform = torchaudio.transforms.Resample(orig_freq=sr, new_freq=new_sr) 18 | wav = transform(wav) 19 | return wav.squeeze(0) 20 | 21 | 22 | def map_timestamps_to_new_sr(vad_sr, new_sr, timestamps, just_begging_end=False): 23 | factor = new_sr / vad_sr 24 | new_timestamps = [] 25 | if just_begging_end and timestamps: 26 | # get just the start and end timestamps 27 | new_dict = {"start": int(timestamps[0]["start"] * factor), "end": int(timestamps[-1]["end"] * factor)} 28 | new_timestamps.append(new_dict) 29 | else: 30 | for ts in timestamps: 31 | # map to the new SR 32 | new_dict = {"start": int(ts["start"] * factor), "end": int(ts["end"] * factor)} 33 | new_timestamps.append(new_dict) 34 | 35 | return new_timestamps 36 | 37 | 38 | def get_vad_model_and_utils(use_cuda=False): 39 | model, utils = torch.hub.load(repo_or_dir="snakers4/silero-vad", model="silero_vad", force_reload=True, onnx=False) 40 | if use_cuda: 41 | model = model.cuda() 42 | 43 | get_speech_timestamps, save_audio, _, _, collect_chunks = utils 44 | return model, get_speech_timestamps, save_audio, collect_chunks 45 | 46 | 47 | def remove_silence( 48 | model_and_utils, audio_path, out_path, vad_sample_rate=8000, trim_just_beginning_and_end=True, use_cuda=False 49 | ): 50 | # get the VAD model and utils functions 51 | model, get_speech_timestamps, _, collect_chunks = model_and_utils 52 | 53 | # read ground truth wav and resample the audio for the VAD 54 | wav, gt_sample_rate = read_audio(audio_path) 55 | 56 | # if needed, resample the audio for the VAD model 57 | if gt_sample_rate != vad_sample_rate: 58 | wav_vad = resample_wav(wav, gt_sample_rate, vad_sample_rate) 59 | else: 60 | wav_vad = wav 61 | 62 | if use_cuda: 63 | wav_vad = wav_vad.cuda() 64 | 65 | # get speech timestamps from full audio file 66 | speech_timestamps = get_speech_timestamps(wav_vad, model, sampling_rate=vad_sample_rate, window_size_samples=768) 67 | 68 | # map the current speech_timestamps to the sample rate of the ground truth audio 69 | new_speech_timestamps = map_timestamps_to_new_sr( 70 | vad_sample_rate, gt_sample_rate, speech_timestamps, trim_just_beginning_and_end 71 | ) 72 | 73 | # if have speech timestamps else save the wav 74 | if new_speech_timestamps: 75 | wav = collect_chunks(new_speech_timestamps, wav) 76 | is_speech = True 77 | else: 78 | print(f"> The file {audio_path} probably does not have speech please check it !!") 79 | is_speech = False 80 | 81 | # save audio 82 | sf.write(out_path, wav, gt_sample_rate, subtype="PCM_16") 83 | return out_path, is_speech 84 | -------------------------------------------------------------------------------- /TTS/vocoder/README.md: -------------------------------------------------------------------------------- 1 | # Mozilla TTS Vocoders (Experimental) 2 | 3 | Here there are vocoder model implementations which can be combined with the other TTS models. 4 | 5 | Currently, following models are implemented: 6 | 7 | - Melgan 8 | - MultiBand-Melgan 9 | - ParallelWaveGAN 10 | - GAN-TTS (Discriminator Only) 11 | 12 | It is also very easy to adapt different vocoder models as we provide a flexible and modular (but not too modular) framework. 13 | 14 | ## Training a model 15 | 16 | You can see here an example (Soon)[Colab Notebook]() training MelGAN with LJSpeech dataset. 17 | 18 | In order to train a new model, you need to gather all wav files into a folder and give this folder to `data_path` in '''config.json''' 19 | 20 | You need to define other relevant parameters in your ```config.json``` and then start traning with the following command. 21 | 22 | ```CUDA_VISIBLE_DEVICES='0' python tts/bin/train_vocoder.py --config_path path/to/config.json``` 23 | 24 | Example config files can be found under `tts/vocoder/configs/` folder. 25 | 26 | You can continue a previous training run by the following command. 27 | 28 | ```CUDA_VISIBLE_DEVICES='0' python tts/bin/train_vocoder.py --continue_path path/to/your/model/folder``` 29 | 30 | You can fine-tune a pre-trained model by the following command. 31 | 32 | ```CUDA_VISIBLE_DEVICES='0' python tts/bin/train_vocoder.py --restore_path path/to/your/model.pth``` 33 | 34 | Restoring a model starts a new training in a different folder. It only restores model weights with the given checkpoint file. However, continuing a training starts from the same directory where the previous training run left off. 35 | 36 | You can also follow your training runs on Tensorboard as you do with our TTS models. 37 | 38 | ## Acknowledgement 39 | Thanks to @kan-bayashi for his [repository](https://github.com/kan-bayashi/ParallelWaveGAN) being the start point of our work. 40 | -------------------------------------------------------------------------------- /TTS/vocoder/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/vocoder/__init__.py -------------------------------------------------------------------------------- /TTS/vocoder/configs/__init__.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import os 3 | from inspect import isclass 4 | 5 | # import all files under configs/ 6 | configs_dir = os.path.dirname(__file__) 7 | for file in os.listdir(configs_dir): 8 | path = os.path.join(configs_dir, file) 9 | if not file.startswith("_") and not file.startswith(".") and (file.endswith(".py") or os.path.isdir(path)): 10 | config_name = file[: file.find(".py")] if file.endswith(".py") else file 11 | module = importlib.import_module("TTS.vocoder.configs." + config_name) 12 | for attribute_name in dir(module): 13 | attribute = getattr(module, attribute_name) 14 | 15 | if isclass(attribute): 16 | # Add the class to this package's variables 17 | globals()[attribute_name] = attribute 18 | -------------------------------------------------------------------------------- /TTS/vocoder/configs/wavegrad_config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import dataclass, field 2 | 3 | from TTS.vocoder.configs.shared_configs import BaseVocoderConfig 4 | from TTS.vocoder.models.wavegrad import WavegradArgs 5 | 6 | 7 | @dataclass 8 | class WavegradConfig(BaseVocoderConfig): 9 | """Defines parameters for WaveGrad vocoder. 10 | Example: 11 | 12 | >>> from TTS.vocoder.configs import WavegradConfig 13 | >>> config = WavegradConfig() 14 | 15 | Args: 16 | model (str): 17 | Model name used for selecting the right model at initialization. Defaults to `wavegrad`. 18 | generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is 19 | considered as a generator too. Defaults to `wavegrad`. 20 | model_params (WavegradArgs): Model parameters. Check `WavegradArgs` for default values. 21 | target_loss (str): 22 | Target loss name that defines the quality of the model. Defaults to `avg_wavegrad_loss`. 23 | epochs (int): 24 | Number of epochs to traing the model. Defaults to 10000. 25 | batch_size (int): 26 | Batch size used at training. Larger values use more memory. Defaults to 96. 27 | seq_len (int): 28 | Audio segment length used at training. Larger values use more memory. Defaults to 6144. 29 | use_cache (bool): 30 | enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is 31 | not large enough. Defaults to True. 32 | mixed_precision (bool): 33 | enable / disable mixed precision training. Default is True. 34 | eval_split_size (int): 35 | Number of samples used for evalutaion. Defaults to 50. 36 | train_noise_schedule (dict): 37 | Training noise schedule. Defaults to 38 | `{"min_val": 1e-6, "max_val": 1e-2, "num_steps": 1000}` 39 | test_noise_schedule (dict): 40 | Inference noise schedule. For a better performance, you may need to use `bin/tune_wavegrad.py` to find a 41 | better schedule. Defaults to 42 | ` 43 | { 44 | "min_val": 1e-6, 45 | "max_val": 1e-2, 46 | "num_steps": 50, 47 | } 48 | ` 49 | grad_clip (float): 50 | Gradient clipping threshold. If <= 0.0, no clipping is applied. Defaults to 1.0 51 | lr (float): 52 | Initila leraning rate. Defaults to 1e-4. 53 | lr_scheduler (str): 54 | One of the learning rate schedulers from `torch.optim.scheduler.*`. Defaults to `MultiStepLR`. 55 | lr_scheduler_params (dict): 56 | kwargs for the scheduler. Defaults to `{"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}` 57 | """ 58 | 59 | model: str = "wavegrad" 60 | # Model specific params 61 | generator_model: str = "wavegrad" 62 | model_params: WavegradArgs = field(default_factory=WavegradArgs) 63 | target_loss: str = "loss" # loss value to pick the best model to save after each epoch 64 | 65 | # Training - overrides 66 | epochs: int = 10000 67 | batch_size: int = 96 68 | seq_len: int = 6144 69 | use_cache: bool = True 70 | mixed_precision: bool = True 71 | eval_split_size: int = 50 72 | 73 | # NOISE SCHEDULE PARAMS 74 | train_noise_schedule: dict = field(default_factory=lambda: {"min_val": 1e-6, "max_val": 1e-2, "num_steps": 1000}) 75 | 76 | test_noise_schedule: dict = field( 77 | default_factory=lambda: { # inference noise schedule. Try TTS/bin/tune_wavegrad.py to find the optimal values. 78 | "min_val": 1e-6, 79 | "max_val": 1e-2, 80 | "num_steps": 50, 81 | } 82 | ) 83 | 84 | # optimizer overrides 85 | grad_clip: float = 1.0 86 | lr: float = 1e-4 # Initial learning rate. 87 | lr_scheduler: str = "MultiStepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html 88 | lr_scheduler_params: dict = field( 89 | default_factory=lambda: {"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]} 90 | ) 91 | -------------------------------------------------------------------------------- /TTS/vocoder/datasets/__init__.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from coqpit import Coqpit 4 | from torch.utils.data import Dataset 5 | 6 | from TTS.utils.audio import AudioProcessor 7 | from TTS.vocoder.datasets.gan_dataset import GANDataset 8 | from TTS.vocoder.datasets.preprocess import load_wav_data, load_wav_feat_data 9 | from TTS.vocoder.datasets.wavegrad_dataset import WaveGradDataset 10 | from TTS.vocoder.datasets.wavernn_dataset import WaveRNNDataset 11 | 12 | 13 | def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: List, verbose: bool) -> Dataset: 14 | if config.model.lower() in "gan": 15 | dataset = GANDataset( 16 | ap=ap, 17 | items=data_items, 18 | seq_len=config.seq_len, 19 | hop_len=ap.hop_length, 20 | pad_short=config.pad_short, 21 | conv_pad=config.conv_pad, 22 | return_pairs=config.diff_samples_for_G_and_D if "diff_samples_for_G_and_D" in config else False, 23 | is_training=not is_eval, 24 | return_segments=not is_eval, 25 | use_noise_augment=config.use_noise_augment, 26 | use_cache=config.use_cache, 27 | verbose=verbose, 28 | ) 29 | dataset.shuffle_mapping() 30 | elif config.model.lower() == "wavegrad": 31 | dataset = WaveGradDataset( 32 | ap=ap, 33 | items=data_items, 34 | seq_len=config.seq_len, 35 | hop_len=ap.hop_length, 36 | pad_short=config.pad_short, 37 | conv_pad=config.conv_pad, 38 | is_training=not is_eval, 39 | return_segments=True, 40 | use_noise_augment=False, 41 | use_cache=config.use_cache, 42 | verbose=verbose, 43 | ) 44 | elif config.model.lower() == "wavernn": 45 | dataset = WaveRNNDataset( 46 | ap=ap, 47 | items=data_items, 48 | seq_len=config.seq_len, 49 | hop_len=ap.hop_length, 50 | pad=config.model_params.pad, 51 | mode=config.model_params.mode, 52 | mulaw=config.model_params.mulaw, 53 | is_training=not is_eval, 54 | verbose=verbose, 55 | ) 56 | else: 57 | raise ValueError(f" [!] Dataset for model {config.model.lower()} cannot be found.") 58 | return dataset 59 | -------------------------------------------------------------------------------- /TTS/vocoder/datasets/preprocess.py: -------------------------------------------------------------------------------- 1 | import glob 2 | import os 3 | from pathlib import Path 4 | 5 | import numpy as np 6 | from coqpit import Coqpit 7 | from tqdm import tqdm 8 | 9 | from TTS.utils.audio import AudioProcessor 10 | 11 | 12 | def preprocess_wav_files(out_path: str, config: Coqpit, ap: AudioProcessor): 13 | """Process wav and compute mel and quantized wave signal. 14 | It is mainly used by WaveRNN dataloader. 15 | 16 | Args: 17 | out_path (str): Parent folder path to save the files. 18 | config (Coqpit): Model config. 19 | ap (AudioProcessor): Audio processor. 20 | """ 21 | os.makedirs(os.path.join(out_path, "quant"), exist_ok=True) 22 | os.makedirs(os.path.join(out_path, "mel"), exist_ok=True) 23 | wav_files = find_wav_files(config.data_path) 24 | for path in tqdm(wav_files): 25 | wav_name = Path(path).stem 26 | quant_path = os.path.join(out_path, "quant", wav_name + ".npy") 27 | mel_path = os.path.join(out_path, "mel", wav_name + ".npy") 28 | y = ap.load_wav(path) 29 | mel = ap.melspectrogram(y) 30 | np.save(mel_path, mel) 31 | if isinstance(config.mode, int): 32 | quant = ap.mulaw_encode(y, qc=config.mode) if config.model_args.mulaw else ap.quantize(y, bits=config.mode) 33 | np.save(quant_path, quant) 34 | 35 | 36 | def find_wav_files(data_path, file_ext="wav"): 37 | wav_paths = glob.glob(os.path.join(data_path, "**", f"*.{file_ext}"), recursive=True) 38 | return wav_paths 39 | 40 | 41 | def find_feat_files(data_path): 42 | feat_paths = glob.glob(os.path.join(data_path, "**", "*.npy"), recursive=True) 43 | return feat_paths 44 | 45 | 46 | def load_wav_data(data_path, eval_split_size, file_ext="wav"): 47 | wav_paths = find_wav_files(data_path, file_ext=file_ext) 48 | assert len(wav_paths) > 0, f" [!] {data_path} is empty." 49 | np.random.seed(0) 50 | np.random.shuffle(wav_paths) 51 | return wav_paths[:eval_split_size], wav_paths[eval_split_size:] 52 | 53 | 54 | def load_wav_feat_data(data_path, feat_path, eval_split_size): 55 | wav_paths = find_wav_files(data_path) 56 | feat_paths = find_feat_files(feat_path) 57 | 58 | wav_paths.sort(key=lambda x: Path(x).stem) 59 | feat_paths.sort(key=lambda x: Path(x).stem) 60 | 61 | assert len(wav_paths) == len(feat_paths), f" [!] {len(wav_paths)} vs {feat_paths}" 62 | for wav, feat in zip(wav_paths, feat_paths): 63 | wav_name = Path(wav).stem 64 | feat_name = Path(feat).stem 65 | assert wav_name == feat_name 66 | 67 | items = list(zip(wav_paths, feat_paths)) 68 | np.random.seed(0) 69 | np.random.shuffle(items) 70 | return items[:eval_split_size], items[eval_split_size:] 71 | -------------------------------------------------------------------------------- /TTS/vocoder/datasets/wavernn_dataset.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from torch.utils.data import Dataset 4 | 5 | 6 | class WaveRNNDataset(Dataset): 7 | """ 8 | WaveRNN Dataset searchs for all the wav files under root path 9 | and converts them to acoustic features on the fly. 10 | """ 11 | 12 | def __init__( 13 | self, ap, items, seq_len, hop_len, pad, mode, mulaw, is_training=True, verbose=False, return_segments=True 14 | ): 15 | super().__init__() 16 | self.ap = ap 17 | self.compute_feat = not isinstance(items[0], (tuple, list)) 18 | self.item_list = items 19 | self.seq_len = seq_len 20 | self.hop_len = hop_len 21 | self.mel_len = seq_len // hop_len 22 | self.pad = pad 23 | self.mode = mode 24 | self.mulaw = mulaw 25 | self.is_training = is_training 26 | self.verbose = verbose 27 | self.return_segments = return_segments 28 | 29 | assert self.seq_len % self.hop_len == 0 30 | 31 | def __len__(self): 32 | return len(self.item_list) 33 | 34 | def __getitem__(self, index): 35 | item = self.load_item(index) 36 | return item 37 | 38 | def load_test_samples(self, num_samples): 39 | samples = [] 40 | return_segments = self.return_segments 41 | self.return_segments = False 42 | for idx in range(num_samples): 43 | mel, audio, _ = self.load_item(idx) 44 | samples.append([mel, audio]) 45 | self.return_segments = return_segments 46 | return samples 47 | 48 | def load_item(self, index): 49 | """ 50 | load (audio, feat) couple if feature_path is set 51 | else compute it on the fly 52 | """ 53 | if self.compute_feat: 54 | wavpath = self.item_list[index] 55 | audio = self.ap.load_wav(wavpath) 56 | if self.return_segments: 57 | min_audio_len = 2 * self.seq_len + (2 * self.pad * self.hop_len) 58 | else: 59 | min_audio_len = audio.shape[0] + (2 * self.pad * self.hop_len) 60 | if audio.shape[0] < min_audio_len: 61 | print(" [!] Instance is too short! : {}".format(wavpath)) 62 | audio = np.pad(audio, [0, min_audio_len - audio.shape[0] + self.hop_len]) 63 | mel = self.ap.melspectrogram(audio) 64 | 65 | if self.mode in ["gauss", "mold"]: 66 | x_input = audio 67 | elif isinstance(self.mode, int): 68 | x_input = ( 69 | self.ap.mulaw_encode(audio, qc=self.mode) if self.mulaw else self.ap.quantize(audio, bits=self.mode) 70 | ) 71 | else: 72 | raise RuntimeError("Unknown dataset mode - ", self.mode) 73 | 74 | else: 75 | wavpath, feat_path = self.item_list[index] 76 | mel = np.load(feat_path.replace("/quant/", "/mel/")) 77 | 78 | if mel.shape[-1] < self.mel_len + 2 * self.pad: 79 | print(" [!] Instance is too short! : {}".format(wavpath)) 80 | self.item_list[index] = self.item_list[index + 1] 81 | feat_path = self.item_list[index] 82 | mel = np.load(feat_path.replace("/quant/", "/mel/")) 83 | if self.mode in ["gauss", "mold"]: 84 | x_input = self.ap.load_wav(wavpath) 85 | elif isinstance(self.mode, int): 86 | x_input = np.load(feat_path.replace("/mel/", "/quant/")) 87 | else: 88 | raise RuntimeError("Unknown dataset mode - ", self.mode) 89 | 90 | return mel, x_input, wavpath 91 | 92 | def collate(self, batch): 93 | mel_win = self.seq_len // self.hop_len + 2 * self.pad 94 | max_offsets = [x[0].shape[-1] - (mel_win + 2 * self.pad) for x in batch] 95 | 96 | mel_offsets = [np.random.randint(0, offset) for offset in max_offsets] 97 | sig_offsets = [(offset + self.pad) * self.hop_len for offset in mel_offsets] 98 | 99 | mels = [x[0][:, mel_offsets[i] : mel_offsets[i] + mel_win] for i, x in enumerate(batch)] 100 | 101 | coarse = [x[1][sig_offsets[i] : sig_offsets[i] + self.seq_len + 1] for i, x in enumerate(batch)] 102 | 103 | mels = np.stack(mels).astype(np.float32) 104 | if self.mode in ["gauss", "mold"]: 105 | coarse = np.stack(coarse).astype(np.float32) 106 | coarse = torch.FloatTensor(coarse) 107 | x_input = coarse[:, : self.seq_len] 108 | elif isinstance(self.mode, int): 109 | coarse = np.stack(coarse).astype(np.int64) 110 | coarse = torch.LongTensor(coarse) 111 | x_input = 2 * coarse[:, : self.seq_len].float() / (2**self.mode - 1.0) - 1.0 112 | y_coarse = coarse[:, 1:] 113 | mels = torch.FloatTensor(mels) 114 | return x_input, mels, y_coarse 115 | -------------------------------------------------------------------------------- /TTS/vocoder/layers/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/vocoder/layers/__init__.py -------------------------------------------------------------------------------- /TTS/vocoder/layers/hifigan.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | 4 | # pylint: disable=dangerous-default-value 5 | class ResStack(nn.Module): 6 | def __init__(self, kernel, channel, padding, dilations=[1, 3, 5]): 7 | super().__init__() 8 | resstack = [] 9 | for dilation in dilations: 10 | resstack += [ 11 | nn.LeakyReLU(0.2), 12 | nn.ReflectionPad1d(dilation), 13 | nn.utils.weight_norm(nn.Conv1d(channel, channel, kernel_size=kernel, dilation=dilation)), 14 | nn.LeakyReLU(0.2), 15 | nn.ReflectionPad1d(padding), 16 | nn.utils.weight_norm(nn.Conv1d(channel, channel, kernel_size=1)), 17 | ] 18 | self.resstack = nn.Sequential(*resstack) 19 | 20 | self.shortcut = nn.utils.weight_norm(nn.Conv1d(channel, channel, kernel_size=1)) 21 | 22 | def forward(self, x): 23 | x1 = self.shortcut(x) 24 | x2 = self.resstack(x) 25 | return x1 + x2 26 | 27 | def remove_weight_norm(self): 28 | nn.utils.remove_weight_norm(self.shortcut) 29 | nn.utils.remove_weight_norm(self.resstack[2]) 30 | nn.utils.remove_weight_norm(self.resstack[5]) 31 | nn.utils.remove_weight_norm(self.resstack[8]) 32 | nn.utils.remove_weight_norm(self.resstack[11]) 33 | nn.utils.remove_weight_norm(self.resstack[14]) 34 | nn.utils.remove_weight_norm(self.resstack[17]) 35 | 36 | 37 | class MRF(nn.Module): 38 | def __init__(self, kernels, channel, dilations=[1, 3, 5]): # # pylint: disable=dangerous-default-value 39 | super().__init__() 40 | self.resblock1 = ResStack(kernels[0], channel, 0, dilations) 41 | self.resblock2 = ResStack(kernels[1], channel, 6, dilations) 42 | self.resblock3 = ResStack(kernels[2], channel, 12, dilations) 43 | 44 | def forward(self, x): 45 | x1 = self.resblock1(x) 46 | x2 = self.resblock2(x) 47 | x3 = self.resblock3(x) 48 | return x1 + x2 + x3 49 | 50 | def remove_weight_norm(self): 51 | self.resblock1.remove_weight_norm() 52 | self.resblock2.remove_weight_norm() 53 | self.resblock3.remove_weight_norm() 54 | -------------------------------------------------------------------------------- /TTS/vocoder/layers/melgan.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | from torch.nn.utils import weight_norm 3 | 4 | 5 | class ResidualStack(nn.Module): 6 | def __init__(self, channels, num_res_blocks, kernel_size): 7 | super().__init__() 8 | 9 | assert (kernel_size - 1) % 2 == 0, " [!] kernel_size has to be odd." 10 | base_padding = (kernel_size - 1) // 2 11 | 12 | self.blocks = nn.ModuleList() 13 | for idx in range(num_res_blocks): 14 | layer_kernel_size = kernel_size 15 | layer_dilation = layer_kernel_size**idx 16 | layer_padding = base_padding * layer_dilation 17 | self.blocks += [ 18 | nn.Sequential( 19 | nn.LeakyReLU(0.2), 20 | nn.ReflectionPad1d(layer_padding), 21 | weight_norm( 22 | nn.Conv1d(channels, channels, kernel_size=kernel_size, dilation=layer_dilation, bias=True) 23 | ), 24 | nn.LeakyReLU(0.2), 25 | weight_norm(nn.Conv1d(channels, channels, kernel_size=1, bias=True)), 26 | ) 27 | ] 28 | 29 | self.shortcuts = nn.ModuleList( 30 | [weight_norm(nn.Conv1d(channels, channels, kernel_size=1, bias=True)) for i in range(num_res_blocks)] 31 | ) 32 | 33 | def forward(self, x): 34 | for block, shortcut in zip(self.blocks, self.shortcuts): 35 | x = shortcut(x) + block(x) 36 | return x 37 | 38 | def remove_weight_norm(self): 39 | for block, shortcut in zip(self.blocks, self.shortcuts): 40 | nn.utils.remove_weight_norm(block[2]) 41 | nn.utils.remove_weight_norm(block[4]) 42 | nn.utils.remove_weight_norm(shortcut) 43 | -------------------------------------------------------------------------------- /TTS/vocoder/layers/parallel_wavegan.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.nn import functional as F 3 | 4 | 5 | class ResidualBlock(torch.nn.Module): 6 | """Residual block module in WaveNet.""" 7 | 8 | def __init__( 9 | self, 10 | kernel_size=3, 11 | res_channels=64, 12 | gate_channels=128, 13 | skip_channels=64, 14 | aux_channels=80, 15 | dropout=0.0, 16 | dilation=1, 17 | bias=True, 18 | use_causal_conv=False, 19 | ): 20 | super().__init__() 21 | self.dropout = dropout 22 | # no future time stamps available 23 | if use_causal_conv: 24 | padding = (kernel_size - 1) * dilation 25 | else: 26 | assert (kernel_size - 1) % 2 == 0, "Not support even number kernel size." 27 | padding = (kernel_size - 1) // 2 * dilation 28 | self.use_causal_conv = use_causal_conv 29 | 30 | # dilation conv 31 | self.conv = torch.nn.Conv1d( 32 | res_channels, gate_channels, kernel_size, padding=padding, dilation=dilation, bias=bias 33 | ) 34 | 35 | # local conditioning 36 | if aux_channels > 0: 37 | self.conv1x1_aux = torch.nn.Conv1d(aux_channels, gate_channels, 1, bias=False) 38 | else: 39 | self.conv1x1_aux = None 40 | 41 | # conv output is split into two groups 42 | gate_out_channels = gate_channels // 2 43 | self.conv1x1_out = torch.nn.Conv1d(gate_out_channels, res_channels, 1, bias=bias) 44 | self.conv1x1_skip = torch.nn.Conv1d(gate_out_channels, skip_channels, 1, bias=bias) 45 | 46 | def forward(self, x, c): 47 | """ 48 | x: B x D_res x T 49 | c: B x D_aux x T 50 | """ 51 | residual = x 52 | x = F.dropout(x, p=self.dropout, training=self.training) 53 | x = self.conv(x) 54 | 55 | # remove future time steps if use_causal_conv conv 56 | x = x[:, :, : residual.size(-1)] if self.use_causal_conv else x 57 | 58 | # split into two part for gated activation 59 | splitdim = 1 60 | xa, xb = x.split(x.size(splitdim) // 2, dim=splitdim) 61 | 62 | # local conditioning 63 | if c is not None: 64 | assert self.conv1x1_aux is not None 65 | c = self.conv1x1_aux(c) 66 | ca, cb = c.split(c.size(splitdim) // 2, dim=splitdim) 67 | xa, xb = xa + ca, xb + cb 68 | 69 | x = torch.tanh(xa) * torch.sigmoid(xb) 70 | 71 | # for skip connection 72 | s = self.conv1x1_skip(x) 73 | 74 | # for residual connection 75 | x = (self.conv1x1_out(x) + residual) * (0.5**2) 76 | 77 | return x, s 78 | -------------------------------------------------------------------------------- /TTS/vocoder/layers/pqmf.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn.functional as F 4 | from scipy import signal as sig 5 | 6 | 7 | # adapted from 8 | # https://github.com/kan-bayashi/ParallelWaveGAN/tree/master/parallel_wavegan 9 | class PQMF(torch.nn.Module): 10 | def __init__(self, N=4, taps=62, cutoff=0.15, beta=9.0): 11 | super().__init__() 12 | 13 | self.N = N 14 | self.taps = taps 15 | self.cutoff = cutoff 16 | self.beta = beta 17 | 18 | QMF = sig.firwin(taps + 1, cutoff, window=("kaiser", beta)) 19 | H = np.zeros((N, len(QMF))) 20 | G = np.zeros((N, len(QMF))) 21 | for k in range(N): 22 | constant_factor = ( 23 | (2 * k + 1) * (np.pi / (2 * N)) * (np.arange(taps + 1) - ((taps - 1) / 2)) 24 | ) # TODO: (taps - 1) -> taps 25 | phase = (-1) ** k * np.pi / 4 26 | H[k] = 2 * QMF * np.cos(constant_factor + phase) 27 | 28 | G[k] = 2 * QMF * np.cos(constant_factor - phase) 29 | 30 | H = torch.from_numpy(H[:, None, :]).float() 31 | G = torch.from_numpy(G[None, :, :]).float() 32 | 33 | self.register_buffer("H", H) 34 | self.register_buffer("G", G) 35 | 36 | updown_filter = torch.zeros((N, N, N)).float() 37 | for k in range(N): 38 | updown_filter[k, k, 0] = 1.0 39 | self.register_buffer("updown_filter", updown_filter) 40 | self.N = N 41 | 42 | self.pad_fn = torch.nn.ConstantPad1d(taps // 2, 0.0) 43 | 44 | def forward(self, x): 45 | return self.analysis(x) 46 | 47 | def analysis(self, x): 48 | return F.conv1d(x, self.H, padding=self.taps // 2, stride=self.N) 49 | 50 | def synthesis(self, x): 51 | x = F.conv_transpose1d(x, self.updown_filter * self.N, stride=self.N) 52 | x = F.conv1d(x, self.G, padding=self.taps // 2) 53 | return x 54 | -------------------------------------------------------------------------------- /TTS/vocoder/layers/upsample.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.nn import functional as F 3 | 4 | 5 | class Stretch2d(torch.nn.Module): 6 | def __init__(self, x_scale, y_scale, mode="nearest"): 7 | super().__init__() 8 | self.x_scale = x_scale 9 | self.y_scale = y_scale 10 | self.mode = mode 11 | 12 | def forward(self, x): 13 | """ 14 | x (Tensor): Input tensor (B, C, F, T). 15 | Tensor: Interpolated tensor (B, C, F * y_scale, T * x_scale), 16 | """ 17 | return F.interpolate(x, scale_factor=(self.y_scale, self.x_scale), mode=self.mode) 18 | 19 | 20 | class UpsampleNetwork(torch.nn.Module): 21 | # pylint: disable=dangerous-default-value 22 | def __init__( 23 | self, 24 | upsample_factors, 25 | nonlinear_activation=None, 26 | nonlinear_activation_params={}, 27 | interpolate_mode="nearest", 28 | freq_axis_kernel_size=1, 29 | use_causal_conv=False, 30 | ): 31 | super().__init__() 32 | self.use_causal_conv = use_causal_conv 33 | self.up_layers = torch.nn.ModuleList() 34 | for scale in upsample_factors: 35 | # interpolation layer 36 | stretch = Stretch2d(scale, 1, interpolate_mode) 37 | self.up_layers += [stretch] 38 | 39 | # conv layer 40 | assert (freq_axis_kernel_size - 1) % 2 == 0, "Not support even number freq axis kernel size." 41 | freq_axis_padding = (freq_axis_kernel_size - 1) // 2 42 | kernel_size = (freq_axis_kernel_size, scale * 2 + 1) 43 | if use_causal_conv: 44 | padding = (freq_axis_padding, scale * 2) 45 | else: 46 | padding = (freq_axis_padding, scale) 47 | conv = torch.nn.Conv2d(1, 1, kernel_size=kernel_size, padding=padding, bias=False) 48 | self.up_layers += [conv] 49 | 50 | # nonlinear 51 | if nonlinear_activation is not None: 52 | nonlinear = getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params) 53 | self.up_layers += [nonlinear] 54 | 55 | def forward(self, c): 56 | """ 57 | c : (B, C, T_in). 58 | Tensor: (B, C, T_upsample) 59 | """ 60 | c = c.unsqueeze(1) # (B, 1, C, T) 61 | for f in self.up_layers: 62 | c = f(c) 63 | return c.squeeze(1) # (B, C, T') 64 | 65 | 66 | class ConvUpsample(torch.nn.Module): 67 | # pylint: disable=dangerous-default-value 68 | def __init__( 69 | self, 70 | upsample_factors, 71 | nonlinear_activation=None, 72 | nonlinear_activation_params={}, 73 | interpolate_mode="nearest", 74 | freq_axis_kernel_size=1, 75 | aux_channels=80, 76 | aux_context_window=0, 77 | use_causal_conv=False, 78 | ): 79 | super().__init__() 80 | self.aux_context_window = aux_context_window 81 | self.use_causal_conv = use_causal_conv and aux_context_window > 0 82 | # To capture wide-context information in conditional features 83 | kernel_size = aux_context_window + 1 if use_causal_conv else 2 * aux_context_window + 1 84 | # NOTE(kan-bayashi): Here do not use padding because the input is already padded 85 | self.conv_in = torch.nn.Conv1d(aux_channels, aux_channels, kernel_size=kernel_size, bias=False) 86 | self.upsample = UpsampleNetwork( 87 | upsample_factors=upsample_factors, 88 | nonlinear_activation=nonlinear_activation, 89 | nonlinear_activation_params=nonlinear_activation_params, 90 | interpolate_mode=interpolate_mode, 91 | freq_axis_kernel_size=freq_axis_kernel_size, 92 | use_causal_conv=use_causal_conv, 93 | ) 94 | 95 | def forward(self, c): 96 | """ 97 | c : (B, C, T_in). 98 | Tensor: (B, C, T_upsampled), 99 | """ 100 | c_ = self.conv_in(c) 101 | c = c_[:, :, : -self.aux_context_window] if self.use_causal_conv else c_ 102 | return self.upsample(c) 103 | -------------------------------------------------------------------------------- /TTS/vocoder/models/base_vocoder.py: -------------------------------------------------------------------------------- 1 | from coqpit import Coqpit 2 | 3 | from TTS.model import BaseTrainerModel 4 | 5 | # pylint: skip-file 6 | 7 | 8 | class BaseVocoder(BaseTrainerModel): 9 | """Base `vocoder` class. Every new `vocoder` model must inherit this. 10 | 11 | It defines `vocoder` specific functions on top of `Model`. 12 | 13 | Notes on input/output tensor shapes: 14 | Any input or output tensor of the model must be shaped as 15 | 16 | - 3D tensors `batch x time x channels` 17 | - 2D tensors `batch x channels` 18 | - 1D tensors `batch x 1` 19 | """ 20 | 21 | def __init__(self, config): 22 | super().__init__() 23 | self._set_model_args(config) 24 | 25 | def _set_model_args(self, config: Coqpit): 26 | """Setup model args based on the config type. 27 | 28 | If the config is for training with a name like "*Config", then the model args are embeded in the 29 | config.model_args 30 | 31 | If the config is for the model with a name like "*Args", then we assign the directly. 32 | """ 33 | # don't use isintance not to import recursively 34 | if "Config" in config.__class__.__name__: 35 | if "characters" in config: 36 | _, self.config, num_chars = self.get_characters(config) 37 | self.config.num_chars = num_chars 38 | if hasattr(self.config, "model_args"): 39 | config.model_args.num_chars = num_chars 40 | if "model_args" in config: 41 | self.args = self.config.model_args 42 | # This is for backward compatibility 43 | if "model_params" in config: 44 | self.args = self.config.model_params 45 | else: 46 | self.config = config 47 | if "model_args" in config: 48 | self.args = self.config.model_args 49 | # This is for backward compatibility 50 | if "model_params" in config: 51 | self.args = self.config.model_params 52 | else: 53 | raise ValueError("config must be either a *Config or *Args") 54 | -------------------------------------------------------------------------------- /TTS/vocoder/models/fullband_melgan_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from TTS.vocoder.models.melgan_generator import MelganGenerator 4 | 5 | 6 | class FullbandMelganGenerator(MelganGenerator): 7 | def __init__( 8 | self, 9 | in_channels=80, 10 | out_channels=1, 11 | proj_kernel=7, 12 | base_channels=512, 13 | upsample_factors=(2, 8, 2, 2), 14 | res_kernel=3, 15 | num_res_blocks=4, 16 | ): 17 | super().__init__( 18 | in_channels=in_channels, 19 | out_channels=out_channels, 20 | proj_kernel=proj_kernel, 21 | base_channels=base_channels, 22 | upsample_factors=upsample_factors, 23 | res_kernel=res_kernel, 24 | num_res_blocks=num_res_blocks, 25 | ) 26 | 27 | @torch.no_grad() 28 | def inference(self, cond_features): 29 | cond_features = cond_features.to(self.layers[1].weight.device) 30 | cond_features = torch.nn.functional.pad( 31 | cond_features, (self.inference_padding, self.inference_padding), "replicate" 32 | ) 33 | return self.layers(cond_features) 34 | -------------------------------------------------------------------------------- /TTS/vocoder/models/melgan_discriminator.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from torch import nn 3 | from torch.nn.utils import weight_norm 4 | 5 | 6 | class MelganDiscriminator(nn.Module): 7 | def __init__( 8 | self, 9 | in_channels=1, 10 | out_channels=1, 11 | kernel_sizes=(5, 3), 12 | base_channels=16, 13 | max_channels=1024, 14 | downsample_factors=(4, 4, 4, 4), 15 | groups_denominator=4, 16 | ): 17 | super().__init__() 18 | self.layers = nn.ModuleList() 19 | 20 | layer_kernel_size = np.prod(kernel_sizes) 21 | layer_padding = (layer_kernel_size - 1) // 2 22 | 23 | # initial layer 24 | self.layers += [ 25 | nn.Sequential( 26 | nn.ReflectionPad1d(layer_padding), 27 | weight_norm(nn.Conv1d(in_channels, base_channels, layer_kernel_size, stride=1)), 28 | nn.LeakyReLU(0.2, inplace=True), 29 | ) 30 | ] 31 | 32 | # downsampling layers 33 | layer_in_channels = base_channels 34 | for downsample_factor in downsample_factors: 35 | layer_out_channels = min(layer_in_channels * downsample_factor, max_channels) 36 | layer_kernel_size = downsample_factor * 10 + 1 37 | layer_padding = (layer_kernel_size - 1) // 2 38 | layer_groups = layer_in_channels // groups_denominator 39 | self.layers += [ 40 | nn.Sequential( 41 | weight_norm( 42 | nn.Conv1d( 43 | layer_in_channels, 44 | layer_out_channels, 45 | kernel_size=layer_kernel_size, 46 | stride=downsample_factor, 47 | padding=layer_padding, 48 | groups=layer_groups, 49 | ) 50 | ), 51 | nn.LeakyReLU(0.2, inplace=True), 52 | ) 53 | ] 54 | layer_in_channels = layer_out_channels 55 | 56 | # last 2 layers 57 | layer_padding1 = (kernel_sizes[0] - 1) // 2 58 | layer_padding2 = (kernel_sizes[1] - 1) // 2 59 | self.layers += [ 60 | nn.Sequential( 61 | weight_norm( 62 | nn.Conv1d( 63 | layer_out_channels, 64 | layer_out_channels, 65 | kernel_size=kernel_sizes[0], 66 | stride=1, 67 | padding=layer_padding1, 68 | ) 69 | ), 70 | nn.LeakyReLU(0.2, inplace=True), 71 | ), 72 | weight_norm( 73 | nn.Conv1d( 74 | layer_out_channels, out_channels, kernel_size=kernel_sizes[1], stride=1, padding=layer_padding2 75 | ) 76 | ), 77 | ] 78 | 79 | def forward(self, x): 80 | feats = [] 81 | for layer in self.layers: 82 | x = layer(x) 83 | feats.append(x) 84 | return x, feats 85 | -------------------------------------------------------------------------------- /TTS/vocoder/models/melgan_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from torch.nn.utils import weight_norm 4 | 5 | from TTS.utils.io import load_fsspec 6 | from TTS.vocoder.layers.melgan import ResidualStack 7 | 8 | 9 | class MelganGenerator(nn.Module): 10 | def __init__( 11 | self, 12 | in_channels=80, 13 | out_channels=1, 14 | proj_kernel=7, 15 | base_channels=512, 16 | upsample_factors=(8, 8, 2, 2), 17 | res_kernel=3, 18 | num_res_blocks=3, 19 | ): 20 | super().__init__() 21 | 22 | # assert model parameters 23 | assert (proj_kernel - 1) % 2 == 0, " [!] proj_kernel should be an odd number." 24 | 25 | # setup additional model parameters 26 | base_padding = (proj_kernel - 1) // 2 27 | act_slope = 0.2 28 | self.inference_padding = 2 29 | 30 | # initial layer 31 | layers = [] 32 | layers += [ 33 | nn.ReflectionPad1d(base_padding), 34 | weight_norm(nn.Conv1d(in_channels, base_channels, kernel_size=proj_kernel, stride=1, bias=True)), 35 | ] 36 | 37 | # upsampling layers and residual stacks 38 | for idx, upsample_factor in enumerate(upsample_factors): 39 | layer_in_channels = base_channels // (2**idx) 40 | layer_out_channels = base_channels // (2 ** (idx + 1)) 41 | layer_filter_size = upsample_factor * 2 42 | layer_stride = upsample_factor 43 | layer_output_padding = upsample_factor % 2 44 | layer_padding = upsample_factor // 2 + layer_output_padding 45 | layers += [ 46 | nn.LeakyReLU(act_slope), 47 | weight_norm( 48 | nn.ConvTranspose1d( 49 | layer_in_channels, 50 | layer_out_channels, 51 | layer_filter_size, 52 | stride=layer_stride, 53 | padding=layer_padding, 54 | output_padding=layer_output_padding, 55 | bias=True, 56 | ) 57 | ), 58 | ResidualStack(channels=layer_out_channels, num_res_blocks=num_res_blocks, kernel_size=res_kernel), 59 | ] 60 | 61 | layers += [nn.LeakyReLU(act_slope)] 62 | 63 | # final layer 64 | layers += [ 65 | nn.ReflectionPad1d(base_padding), 66 | weight_norm(nn.Conv1d(layer_out_channels, out_channels, proj_kernel, stride=1, bias=True)), 67 | nn.Tanh(), 68 | ] 69 | self.layers = nn.Sequential(*layers) 70 | 71 | def forward(self, c): 72 | return self.layers(c) 73 | 74 | def inference(self, c): 75 | c = c.to(self.layers[1].weight.device) 76 | c = torch.nn.functional.pad(c, (self.inference_padding, self.inference_padding), "replicate") 77 | return self.layers(c) 78 | 79 | def remove_weight_norm(self): 80 | for _, layer in enumerate(self.layers): 81 | if len(layer.state_dict()) != 0: 82 | try: 83 | nn.utils.remove_weight_norm(layer) 84 | except ValueError: 85 | layer.remove_weight_norm() 86 | 87 | def load_checkpoint( 88 | self, config, checkpoint_path, eval=False, cache=False 89 | ): # pylint: disable=unused-argument, redefined-builtin 90 | state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache) 91 | self.load_state_dict(state["model"]) 92 | if eval: 93 | self.eval() 94 | assert not self.training 95 | self.remove_weight_norm() 96 | -------------------------------------------------------------------------------- /TTS/vocoder/models/melgan_multiscale_discriminator.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from TTS.vocoder.models.melgan_discriminator import MelganDiscriminator 4 | 5 | 6 | class MelganMultiscaleDiscriminator(nn.Module): 7 | def __init__( 8 | self, 9 | in_channels=1, 10 | out_channels=1, 11 | num_scales=3, 12 | kernel_sizes=(5, 3), 13 | base_channels=16, 14 | max_channels=1024, 15 | downsample_factors=(4, 4, 4), 16 | pooling_kernel_size=4, 17 | pooling_stride=2, 18 | pooling_padding=2, 19 | groups_denominator=4, 20 | ): 21 | super().__init__() 22 | 23 | self.discriminators = nn.ModuleList( 24 | [ 25 | MelganDiscriminator( 26 | in_channels=in_channels, 27 | out_channels=out_channels, 28 | kernel_sizes=kernel_sizes, 29 | base_channels=base_channels, 30 | max_channels=max_channels, 31 | downsample_factors=downsample_factors, 32 | groups_denominator=groups_denominator, 33 | ) 34 | for _ in range(num_scales) 35 | ] 36 | ) 37 | 38 | self.pooling = nn.AvgPool1d( 39 | kernel_size=pooling_kernel_size, stride=pooling_stride, padding=pooling_padding, count_include_pad=False 40 | ) 41 | 42 | def forward(self, x): 43 | scores = [] 44 | feats = [] 45 | for disc in self.discriminators: 46 | score, feat = disc(x) 47 | scores.append(score) 48 | feats.append(feat) 49 | x = self.pooling(x) 50 | return scores, feats 51 | -------------------------------------------------------------------------------- /TTS/vocoder/models/multiband_melgan_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from TTS.vocoder.layers.pqmf import PQMF 4 | from TTS.vocoder.models.melgan_generator import MelganGenerator 5 | 6 | 7 | class MultibandMelganGenerator(MelganGenerator): 8 | def __init__( 9 | self, 10 | in_channels=80, 11 | out_channels=4, 12 | proj_kernel=7, 13 | base_channels=384, 14 | upsample_factors=(2, 8, 2, 2), 15 | res_kernel=3, 16 | num_res_blocks=3, 17 | ): 18 | super().__init__( 19 | in_channels=in_channels, 20 | out_channels=out_channels, 21 | proj_kernel=proj_kernel, 22 | base_channels=base_channels, 23 | upsample_factors=upsample_factors, 24 | res_kernel=res_kernel, 25 | num_res_blocks=num_res_blocks, 26 | ) 27 | self.pqmf_layer = PQMF(N=4, taps=62, cutoff=0.15, beta=9.0) 28 | 29 | def pqmf_analysis(self, x): 30 | return self.pqmf_layer.analysis(x) 31 | 32 | def pqmf_synthesis(self, x): 33 | return self.pqmf_layer.synthesis(x) 34 | 35 | @torch.no_grad() 36 | def inference(self, cond_features): 37 | cond_features = cond_features.to(self.layers[1].weight.device) 38 | cond_features = torch.nn.functional.pad( 39 | cond_features, (self.inference_padding, self.inference_padding), "replicate" 40 | ) 41 | return self.pqmf_synthesis(self.layers(cond_features)) 42 | -------------------------------------------------------------------------------- /TTS/vocoder/models/univnet_discriminator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | from torch import nn 4 | from torch.nn.utils import spectral_norm, weight_norm 5 | 6 | from TTS.utils.audio.torch_transforms import TorchSTFT 7 | from TTS.vocoder.models.hifigan_discriminator import MultiPeriodDiscriminator 8 | 9 | LRELU_SLOPE = 0.1 10 | 11 | 12 | class SpecDiscriminator(nn.Module): 13 | """docstring for Discriminator.""" 14 | 15 | def __init__(self, fft_size=1024, hop_length=120, win_length=600, use_spectral_norm=False): 16 | super().__init__() 17 | norm_f = weight_norm if use_spectral_norm is False else spectral_norm 18 | self.fft_size = fft_size 19 | self.hop_length = hop_length 20 | self.win_length = win_length 21 | self.stft = TorchSTFT(fft_size, hop_length, win_length) 22 | self.discriminators = nn.ModuleList( 23 | [ 24 | norm_f(nn.Conv2d(1, 32, kernel_size=(3, 9), padding=(1, 4))), 25 | norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1, 2), padding=(1, 4))), 26 | norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1, 2), padding=(1, 4))), 27 | norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1, 2), padding=(1, 4))), 28 | norm_f(nn.Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))), 29 | ] 30 | ) 31 | 32 | self.out = norm_f(nn.Conv2d(32, 1, 3, 1, 1)) 33 | 34 | def forward(self, y): 35 | fmap = [] 36 | with torch.no_grad(): 37 | y = y.squeeze(1) 38 | y = self.stft(y) 39 | y = y.unsqueeze(1) 40 | for _, d in enumerate(self.discriminators): 41 | y = d(y) 42 | y = F.leaky_relu(y, LRELU_SLOPE) 43 | fmap.append(y) 44 | 45 | y = self.out(y) 46 | fmap.append(y) 47 | 48 | return torch.flatten(y, 1, -1), fmap 49 | 50 | 51 | class MultiResSpecDiscriminator(torch.nn.Module): 52 | def __init__( # pylint: disable=dangerous-default-value 53 | self, fft_sizes=[1024, 2048, 512], hop_sizes=[120, 240, 50], win_lengths=[600, 1200, 240], window="hann_window" 54 | ): 55 | super().__init__() 56 | self.discriminators = nn.ModuleList( 57 | [ 58 | SpecDiscriminator(fft_sizes[0], hop_sizes[0], win_lengths[0], window), 59 | SpecDiscriminator(fft_sizes[1], hop_sizes[1], win_lengths[1], window), 60 | SpecDiscriminator(fft_sizes[2], hop_sizes[2], win_lengths[2], window), 61 | ] 62 | ) 63 | 64 | def forward(self, x): 65 | scores = [] 66 | feats = [] 67 | for d in self.discriminators: 68 | score, feat = d(x) 69 | scores.append(score) 70 | feats.append(feat) 71 | 72 | return scores, feats 73 | 74 | 75 | class UnivnetDiscriminator(nn.Module): 76 | """Univnet discriminator wrapping MPD and MSD.""" 77 | 78 | def __init__(self): 79 | super().__init__() 80 | self.mpd = MultiPeriodDiscriminator() 81 | self.msd = MultiResSpecDiscriminator() 82 | 83 | def forward(self, x): 84 | """ 85 | Args: 86 | x (Tensor): input waveform. 87 | 88 | Returns: 89 | List[Tensor]: discriminator scores. 90 | List[List[Tensor]]: list of list of features from each layers of each discriminator. 91 | """ 92 | scores, feats = self.mpd(x) 93 | scores_, feats_ = self.msd(x) 94 | return scores + scores_, feats + feats_ 95 | -------------------------------------------------------------------------------- /TTS/vocoder/pqmf_output.wav: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/vocoder/pqmf_output.wav -------------------------------------------------------------------------------- /TTS/vocoder/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/TTS/vocoder/utils/__init__.py -------------------------------------------------------------------------------- /TTS/vocoder/utils/generic_utils.py: -------------------------------------------------------------------------------- 1 | from typing import Dict 2 | 3 | import numpy as np 4 | import torch 5 | from matplotlib import pyplot as plt 6 | 7 | from TTS.tts.utils.visual import plot_spectrogram 8 | from TTS.utils.audio import AudioProcessor 9 | 10 | 11 | def interpolate_vocoder_input(scale_factor, spec): 12 | """Interpolate spectrogram by the scale factor. 13 | It is mainly used to match the sampling rates of 14 | the tts and vocoder models. 15 | 16 | Args: 17 | scale_factor (float): scale factor to interpolate the spectrogram 18 | spec (np.array): spectrogram to be interpolated 19 | 20 | Returns: 21 | torch.tensor: interpolated spectrogram. 22 | """ 23 | print(" > before interpolation :", spec.shape) 24 | spec = torch.tensor(spec).unsqueeze(0).unsqueeze(0) # pylint: disable=not-callable 25 | spec = torch.nn.functional.interpolate( 26 | spec, scale_factor=scale_factor, recompute_scale_factor=True, mode="bilinear", align_corners=False 27 | ).squeeze(0) 28 | print(" > after interpolation :", spec.shape) 29 | return spec 30 | 31 | 32 | def plot_results(y_hat: torch.tensor, y: torch.tensor, ap: AudioProcessor, name_prefix: str = None) -> Dict: 33 | """Plot the predicted and the real waveform and their spectrograms. 34 | 35 | Args: 36 | y_hat (torch.tensor): Predicted waveform. 37 | y (torch.tensor): Real waveform. 38 | ap (AudioProcessor): Audio processor used to process the waveform. 39 | name_prefix (str, optional): Name prefix used to name the figures. Defaults to None. 40 | 41 | Returns: 42 | Dict: output figures keyed by the name of the figures. 43 | """ """Plot vocoder model results""" 44 | if name_prefix is None: 45 | name_prefix = "" 46 | 47 | # select an instance from batch 48 | y_hat = y_hat[0].squeeze().detach().cpu().numpy() 49 | y = y[0].squeeze().detach().cpu().numpy() 50 | 51 | spec_fake = ap.melspectrogram(y_hat).T 52 | spec_real = ap.melspectrogram(y).T 53 | spec_diff = np.abs(spec_fake - spec_real) 54 | 55 | # plot figure and save it 56 | fig_wave = plt.figure() 57 | plt.subplot(2, 1, 1) 58 | plt.plot(y) 59 | plt.title("groundtruth speech") 60 | plt.subplot(2, 1, 2) 61 | plt.plot(y_hat) 62 | plt.title("generated speech") 63 | plt.tight_layout() 64 | plt.close() 65 | 66 | figures = { 67 | name_prefix + "spectrogram/fake": plot_spectrogram(spec_fake), 68 | name_prefix + "spectrogram/real": plot_spectrogram(spec_real), 69 | name_prefix + "spectrogram/diff": plot_spectrogram(spec_diff), 70 | name_prefix + "speech_comparison": fig_wave, 71 | } 72 | return figures 73 | -------------------------------------------------------------------------------- /app.py: -------------------------------------------------------------------------------- 1 | import streamlit as st 2 | from helpers import ProcessModelList, ConvertTextToSpeech, AudioClonning, download_audio_file 3 | import io 4 | 5 | 6 | st.sidebar.title("Welcome to :blue[MimicMania]") 7 | sidebar_options = ["Text To Speech", "Voice Clonning"] 8 | choice = st.sidebar.selectbox(label="Select Your Usecase: ", options=sidebar_options) 9 | 10 | 11 | if choice == sidebar_options[0]: 12 | st.title("Convert your Text to Speech") 13 | model_list = ProcessModelList() 14 | 15 | selected_language = st.selectbox("Select The Language: ", options=model_list.get_langauge_labels()) 16 | selected_model = st.selectbox("Select The Model: ", options=model_list.get_model_name(selected_language=selected_language)) 17 | selected_model_path = model_list.get_model_path(selected_language=selected_language, selected_model=selected_model) 18 | multi_speaker, multi_speaker_list = model_list.get_multi_speaker_model(model_path=selected_model_path) 19 | 20 | if selected_language == "Multi Language": 21 | speakers, languages = model_list.multi_language_selected(model_path=selected_model_path) 22 | selected_speaker = st.selectbox("Select the voice: ", options=speakers) 23 | selected_speaker_language = st.selectbox("Select the Speaker: ", options=languages) 24 | 25 | 26 | elif multi_speaker == True: 27 | selected_speaker = st.selectbox("Select the voice: ", options=multi_speaker_list) 28 | 29 | else: 30 | selected_speaker = None 31 | selected_speaker_language = None 32 | 33 | text = st.text_area("Enter Your Text which you want to convert to audio.") 34 | 35 | if st.button("Convert"): 36 | 37 | text_to_speech = ConvertTextToSpeech(model_name=selected_model, model_path=selected_model_path, text=text) 38 | 39 | if selected_language == "Multi Language" or selected_language == "Popular Person": 40 | text_to_speech.convert_text_to_speech_multi_langauge(speaker=selected_speaker, language=selected_speaker_language, model_name=selected_model, selected_langauge=selected_language) 41 | else: 42 | text_to_speech.convert_text_to_speech(speaker_id=selected_speaker) 43 | 44 | read_audio = text_to_speech.read_audio_file() 45 | st.audio(read_audio, format='audio/wav') 46 | download_audio_file(audio=read_audio, file_name="TTS") 47 | 48 | 49 | 50 | elif choice == sidebar_options[1]: 51 | st.title("Clone Anyone's Voice") 52 | st.subheader("The Better the quality and duration of the data the more realistic the sound will be.") 53 | uploaded_music = st.file_uploader(label="Upload Your Audio File: ", type=["mp3", "wav"]) 54 | text = st.text_area(label="Enter The text you want to convert: ") 55 | emotion = "Neutral"#st.selectbox(label="Select What will the voice emotion: ", options=["Neutral", "Happy", "Sad", "Angry", "Surprise", "Dull"]) 56 | 57 | if st.button("Start Clonning"): 58 | if uploaded_music is not None: 59 | 60 | audio_filename = uploaded_music.name 61 | audio = io.BytesIO(uploaded_music.read()) 62 | 63 | audio_clonning = AudioClonning(audio=audio, audio_filename=audio_filename, text=text, emotion=emotion) 64 | 65 | cloned_voice = audio_clonning.convert_text_to_speech() 66 | #cloned_voice = audio_clonning.emotion_modification() 67 | 68 | st.audio(cloned_voice, format="audio/wav") 69 | download_audio_file(audio=cloned_voice, file_name="Voice-Cloned") 70 | -------------------------------------------------------------------------------- /clonner_output/sample.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/clonner_output/sample.txt -------------------------------------------------------------------------------- /language_model/sample.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/language_model/sample.txt -------------------------------------------------------------------------------- /model_list.json: -------------------------------------------------------------------------------- 1 | { 2 | "labels": ["Multi Language", "American English", "United Kingdom", "Japanese", "Italian"], 3 | "custom_voice": ["Chales Hoskinson"], 4 | "data": { 5 | "Multi Language": [ 6 | { 7 | "name": "Multi Language", 8 | "model_path": "tts_models--multilingual--multi-dataset--your_tts" 9 | } 10 | ], 11 | 12 | "American English": [ 13 | { 14 | "name": "c50", 15 | "model_path": "tts_models--en--blizzard2013--capacitron-t2-c50" 16 | }, 17 | { 18 | "name": "c150", 19 | "model_path": "tts_models--en--blizzard2013--capacitron-t2-c150_v2" 20 | }, 21 | { 22 | "name": "tacotron2", 23 | "model_path": "tts_models--en--ek1--tacotron2" 24 | }, 25 | { 26 | "name": "fast_pitch", 27 | "model_path": "tts_models--en--ljspeech--fast_pitch" 28 | }, 29 | { 30 | "name": "glow_tts", 31 | "model_path": "tts_models--en--ljspeech--glow-tts" 32 | }, 33 | { 34 | "name": "neutral-hmm", 35 | "model_path": "tts_models--en--ljspeech--neural_hmm" 36 | }, 37 | { 38 | "name": "overflow", 39 | "model_path": "tts_models--en--ljspeech--overflow" 40 | }, 41 | { 42 | "name": "speedy-speech", 43 | "model_path": "tts_models--en--ljspeech--speedy-speech" 44 | }, 45 | { 46 | "name": "DCA", 47 | "model_path": "tts_models--en--ljspeech--tacotron2-DCA" 48 | }, 49 | { 50 | "name": "DDC", 51 | "model_path": "tts_models--en--ljspeech--tacotron2-DDC" 52 | }, 53 | { 54 | "name": "DDS-PH", 55 | "model_path": "tts_models--en--ljspeech--tacotron2-DDC_ph" 56 | }, 57 | { 58 | "name": "Vits", 59 | "model_path": "tts_models--en--ljspeech--vits" 60 | }, 61 | { 62 | "name": "DDCv2", 63 | "model_path": "tts_models--en--sam--tacotron-DDC" 64 | }, 65 | { 66 | "name": "fast_pitchv2", 67 | "model_path": "tts_models--en--vctk--fast_pitch" 68 | }, 69 | { 70 | "name": "Vitsv2", 71 | "model_path": "tts_models--en--vctk--vits" 72 | } 73 | ], 74 | 75 | "United Kingdom": [ 76 | { 77 | "name": "Mai-Glow", 78 | "model_path": "tts_models--uk--mai--glow-tts" 79 | } 80 | ], 81 | "Japanese": [ 82 | { 83 | "name": "DDC", 84 | "model_path": "tts_models--ja--kokoro--tacotron2-DDC" 85 | } 86 | ], 87 | "Italian": [ 88 | { 89 | "name": "Male TTS", 90 | "model_path": "tts_models--it--mai_male--glow-tts" 91 | }, 92 | { 93 | "name": "Female TTS", 94 | "model_path": "tts_models--it--mai_female--glow-tts" 95 | } 96 | ], 97 | "Popular Person": [ 98 | { 99 | "name": "Chales Hoskinson", 100 | "model_path": "popular_person/popular_person_model", 101 | "model_voice": "popular_person/popular_person_voice/chales_hoskinson.wav" 102 | } 103 | ] 104 | } 105 | } 106 | -------------------------------------------------------------------------------- /output/sample.ttx: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/everydaycodings/MimicMania/eecbaae2c430785145afe48bc92a7634ab9fe660/output/sample.ttx -------------------------------------------------------------------------------- /packages.txt: -------------------------------------------------------------------------------- 1 | espeak-ng 2 | espeak 3 | ffmpeg -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # core deps 2 | numpy==1.21.6;python_version<"3.10" 3 | numpy;python_version=="3.10" 4 | cython==0.29.28 5 | scipy>=1.4.0 6 | torch>=1.7 7 | torchaudio 8 | soundfile 9 | librosa==0.8.0 10 | numba==0.55.1;python_version<"3.10" 11 | numba==0.55.2;python_version=="3.10" 12 | inflect==5.6.0 13 | tqdm 14 | anyascii 15 | pyyaml 16 | fsspec>=2021.04.0 17 | packaging 18 | streamlit 19 | # deps for examples 20 | flask 21 | # deps for inference 22 | pysbd 23 | # deps for notebooks 24 | umap-learn==0.5.1 25 | pandas 26 | # deps for training 27 | matplotlib 28 | # coqui stack 29 | trainer==0.0.20 30 | # config management 31 | coqpit>=0.0.16 32 | # chinese g2p deps 33 | jieba 34 | pypinyin 35 | # japanese g2p deps 36 | mecab-python3==1.0.5 37 | unidic-lite==1.0.8 38 | # gruut+supported langs 39 | gruut[de]==2.2.3 40 | # deps for korean 41 | jamo 42 | nltk 43 | g2pkk>=0.1.1 44 | streamlit 45 | pydub -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | from helpers import DownloadModels 2 | 3 | DownloadModels().download_models() --------------------------------------------------------------------------------