├── .gitignore ├── ChatTTS ├── __init__.py ├── core.py ├── experimental │ └── llm.py ├── infer │ └── api.py ├── model │ ├── dvae.py │ └── gpt.py └── utils │ ├── gpu_utils.py │ ├── infer_utils.py │ └── io_utils.py ├── Dockerfile ├── LICENSE ├── Makefile ├── README.md ├── README_CN.md ├── api.py ├── client.py ├── example.ipynb ├── requirements-api-ui.txt ├── requirements-chat-tts.txt ├── requirements.txt ├── webui.py └── webui_api.py /.gitignore: -------------------------------------------------------------------------------- 1 | output.mp3 2 | tts_model/ 3 | # Byte-compiled / optimized / DLL files 4 | __pycache__/ 5 | *.py[cod] 6 | *$py.class 7 | *.ckpt 8 | # C extensions 9 | *.so 10 | *.pt 11 | 12 | # Distribution / packaging 13 | .Python 14 | outputs/ 15 | build/ 16 | develop-eggs/ 17 | dist/ 18 | downloads/ 19 | eggs/ 20 | .eggs/ 21 | lib/ 22 | lib64/ 23 | parts/ 24 | sdist/ 25 | var/ 26 | wheels/ 27 | share/python-wheels/ 28 | *.egg-info/ 29 | asset/* 30 | .installed.cfg 31 | *.egg 32 | MANIFEST 33 | 34 | # PyInstaller 35 | # Usually these files are written by a python script from a template 36 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 37 | *.manifest 38 | *.spec 39 | 40 | # Installer logs 41 | pip-log.txt 42 | pip-delete-this-directory.txt 43 | 44 | # Unit test / coverage reports 45 | htmlcov/ 46 | .tox/ 47 | .nox/ 48 | .coverage 49 | .coverage.* 50 | .cache 51 | nosetests.xml 52 | coverage.xml 53 | *.cover 54 | *.py,cover 55 | .hypothesis/ 56 | .pytest_cache/ 57 | cover/ 58 | 59 | # Translations 60 | *.mo 61 | *.pot 62 | 63 | # Django stuff: 64 | *.log 65 | local_settings.py 66 | db.sqlite3 67 | db.sqlite3-journal 68 | 69 | # Flask stuff: 70 | instance/ 71 | .webassets-cache 72 | 73 | # Scrapy stuff: 74 | .scrapy 75 | 76 | # Sphinx documentation 77 | docs/_build/ 78 | 79 | # PyBuilder 80 | .pybuilder/ 81 | target/ 82 | 83 | # Jupyter Notebook 84 | .ipynb_checkpoints 85 | 86 | # IPython 87 | profile_default/ 88 | ipython_config.py 89 | 90 | # pyenv 91 | # For a library or package, you might want to ignore these files since the code is 92 | # intended to run in multiple environments; otherwise, check them in: 93 | # .python-version 94 | 95 | # pipenv 96 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 97 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 98 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 99 | # install all needed dependencies. 100 | #Pipfile.lock 101 | 102 | # poetry 103 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 104 | # This is especially recommended for binary packages to ensure reproducibility, and is more 105 | # commonly ignored for libraries. 106 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 107 | #poetry.lock 108 | 109 | # pdm 110 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 111 | #pdm.lock 112 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 113 | # in version control. 114 | # https://pdm.fming.dev/#use-with-ide 115 | .pdm.toml 116 | 117 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 118 | __pypackages__/ 119 | 120 | # Celery stuff 121 | celerybeat-schedule 122 | celerybeat.pid 123 | 124 | # SageMath parsed files 125 | *.sage.py 126 | 127 | # Environments 128 | .env 129 | .venv 130 | env/ 131 | venv/ 132 | ENV/ 133 | env.bak/ 134 | venv.bak/ 135 | 136 | # Spyder project settings 137 | .spyderproject 138 | .spyproject 139 | 140 | # Rope project settings 141 | .ropeproject 142 | 143 | # mkdocs documentation 144 | /site 145 | 146 | # mypy 147 | .mypy_cache/ 148 | .dmypy.json 149 | dmypy.json 150 | 151 | # Pyre type checker 152 | .pyre/ 153 | 154 | # pytype static type analyzer 155 | .pytype/ 156 | 157 | # Cython debug symbols 158 | cython_debug/ 159 | 160 | # PyCharm 161 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 162 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 163 | # and can be added to the global gitignore or merged into this file. For a more nuclear 164 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 165 | #.idea/ 166 | -------------------------------------------------------------------------------- /ChatTTS/__init__.py: -------------------------------------------------------------------------------- 1 | from .core import Chat -------------------------------------------------------------------------------- /ChatTTS/core.py: -------------------------------------------------------------------------------- 1 | 2 | import os 3 | import logging 4 | from functools import partial 5 | from omegaconf import OmegaConf 6 | 7 | import torch 8 | from vocos import Vocos 9 | from .model.dvae import DVAE 10 | from .model.gpt import GPT_warpper 11 | from .utils.gpu_utils import select_device 12 | from .utils.infer_utils import count_invalid_characters, detect_language, apply_character_map, apply_half2full_map 13 | from .utils.io_utils import get_latest_modified_file 14 | from .infer.api import refine_text, infer_code 15 | 16 | from huggingface_hub import snapshot_download 17 | 18 | logging.basicConfig(level = logging.INFO) 19 | 20 | 21 | class Chat: 22 | def __init__(self, ): 23 | self.pretrain_models = {} 24 | self.normalizer = {} 25 | self.logger = logging.getLogger(__name__) 26 | 27 | def check_model(self, level = logging.INFO, use_decoder = False): 28 | not_finish = False 29 | check_list = ['vocos', 'gpt', 'tokenizer'] 30 | 31 | if use_decoder: 32 | check_list.append('decoder') 33 | else: 34 | check_list.append('dvae') 35 | 36 | for module in check_list: 37 | if module not in self.pretrain_models: 38 | self.logger.log(logging.WARNING, f'{module} not initialized.') 39 | not_finish = True 40 | 41 | if not not_finish: 42 | self.logger.log(level, f'All initialized.') 43 | 44 | return not not_finish 45 | 46 | def load_models(self, source='huggingface', force_redownload=False, local_path='', **kwargs): 47 | if source == 'huggingface': 48 | hf_home = os.getenv('HF_HOME', os.path.expanduser("~/.cache/huggingface")) 49 | try: 50 | download_path = get_latest_modified_file(os.path.join(hf_home, 'hub/models--2Noise--ChatTTS/snapshots')) 51 | except: 52 | download_path = None 53 | if download_path is None or force_redownload: 54 | self.logger.log(logging.INFO, f'Download from HF: https://huggingface.co/2Noise/ChatTTS') 55 | download_path = snapshot_download(repo_id="2Noise/ChatTTS", allow_patterns=["*.pt", "*.yaml"]) 56 | else: 57 | self.logger.log(logging.INFO, f'Load from cache: {download_path}') 58 | elif source == 'local': 59 | self.logger.log(logging.INFO, f'Load from local: {local_path}') 60 | download_path = local_path 61 | 62 | self._load(**{k: os.path.join(download_path, v) for k, v in OmegaConf.load(os.path.join(download_path, 'config', 'path.yaml')).items()}, **kwargs) 63 | 64 | def _load( 65 | self, 66 | vocos_config_path: str = None, 67 | vocos_ckpt_path: str = None, 68 | dvae_config_path: str = None, 69 | dvae_ckpt_path: str = None, 70 | gpt_config_path: str = None, 71 | gpt_ckpt_path: str = None, 72 | decoder_config_path: str = None, 73 | decoder_ckpt_path: str = None, 74 | tokenizer_path: str = None, 75 | device: str = None, 76 | compile: bool = True, 77 | ): 78 | if not device: 79 | device = select_device(4096) 80 | self.logger.log(logging.INFO, f'use {device}') 81 | 82 | if vocos_config_path: 83 | vocos = Vocos.from_hparams(vocos_config_path).to(device).eval() 84 | assert vocos_ckpt_path, 'vocos_ckpt_path should not be None' 85 | vocos.load_state_dict(torch.load(vocos_ckpt_path)) 86 | self.pretrain_models['vocos'] = vocos 87 | self.logger.log(logging.INFO, 'vocos loaded.') 88 | 89 | if dvae_config_path: 90 | cfg = OmegaConf.load(dvae_config_path) 91 | dvae = DVAE(**cfg).to(device).eval() 92 | assert dvae_ckpt_path, 'dvae_ckpt_path should not be None' 93 | dvae.load_state_dict(torch.load(dvae_ckpt_path, map_location='cpu')) 94 | self.pretrain_models['dvae'] = dvae 95 | self.logger.log(logging.INFO, 'dvae loaded.') 96 | 97 | if gpt_config_path: 98 | cfg = OmegaConf.load(gpt_config_path) 99 | gpt = GPT_warpper(**cfg).to(device).eval() 100 | assert gpt_ckpt_path, 'gpt_ckpt_path should not be None' 101 | gpt.load_state_dict(torch.load(gpt_ckpt_path, map_location='cpu')) 102 | if compile and 'cuda' in str(device): 103 | gpt.gpt.forward = torch.compile(gpt.gpt.forward, backend='inductor', dynamic=True) 104 | self.pretrain_models['gpt'] = gpt 105 | spk_stat_path = os.path.join(os.path.dirname(gpt_ckpt_path), 'spk_stat.pt') 106 | assert os.path.exists(spk_stat_path), f'Missing spk_stat.pt: {spk_stat_path}' 107 | self.pretrain_models['spk_stat'] = torch.load(spk_stat_path).to(device) 108 | self.logger.log(logging.INFO, 'gpt loaded.') 109 | 110 | if decoder_config_path: 111 | cfg = OmegaConf.load(decoder_config_path) 112 | decoder = DVAE(**cfg).to(device).eval() 113 | assert decoder_ckpt_path, 'decoder_ckpt_path should not be None' 114 | decoder.load_state_dict(torch.load(decoder_ckpt_path, map_location='cpu')) 115 | self.pretrain_models['decoder'] = decoder 116 | self.logger.log(logging.INFO, 'decoder loaded.') 117 | 118 | if tokenizer_path: 119 | tokenizer = torch.load(tokenizer_path, map_location='cpu') 120 | tokenizer.padding_side = 'left' 121 | self.pretrain_models['tokenizer'] = tokenizer 122 | self.logger.log(logging.INFO, 'tokenizer loaded.') 123 | 124 | self.check_model() 125 | 126 | def infer( 127 | self, 128 | text, 129 | skip_refine_text=False, 130 | refine_text_only=False, 131 | params_refine_text={}, 132 | params_infer_code={'prompt':'[speed_5]'}, 133 | use_decoder=True, 134 | do_text_normalization=True, 135 | lang=None, 136 | ): 137 | 138 | assert self.check_model(use_decoder=use_decoder) 139 | 140 | if not isinstance(text, list): 141 | text = [text] 142 | 143 | if do_text_normalization: 144 | for i, t in enumerate(text): 145 | _lang = detect_language(t) if lang is None else lang 146 | self.init_normalizer(_lang) 147 | text[i] = self.normalizer[_lang](t) 148 | if _lang == 'zh': 149 | text[i] = apply_half2full_map(text[i]) 150 | 151 | for i, t in enumerate(text): 152 | invalid_characters = count_invalid_characters(t) 153 | if len(invalid_characters): 154 | self.logger.log(logging.WARNING, f'Invalid characters found! : {invalid_characters}') 155 | text[i] = apply_character_map(t) 156 | 157 | if not skip_refine_text: 158 | text_tokens = refine_text(self.pretrain_models, text, **params_refine_text)['ids'] 159 | text_tokens = [i[i < self.pretrain_models['tokenizer'].convert_tokens_to_ids('[break_0]')] for i in text_tokens] 160 | text = self.pretrain_models['tokenizer'].batch_decode(text_tokens) 161 | if refine_text_only: 162 | return text 163 | 164 | text = [params_infer_code.get('prompt', '') + i for i in text] 165 | params_infer_code.pop('prompt', '') 166 | result = infer_code(self.pretrain_models, text, **params_infer_code, return_hidden=use_decoder) 167 | 168 | if use_decoder: 169 | mel_spec = [self.pretrain_models['decoder'](i[None].permute(0,2,1)) for i in result['hiddens']] 170 | else: 171 | mel_spec = [self.pretrain_models['dvae'](i[None].permute(0,2,1)) for i in result['ids']] 172 | 173 | wav = [self.pretrain_models['vocos'].decode(i).cpu().numpy() for i in mel_spec] 174 | 175 | return wav 176 | 177 | def sample_random_speaker(self, ): 178 | 179 | dim = self.pretrain_models['gpt'].gpt.layers[0].mlp.gate_proj.in_features 180 | std, mean = self.pretrain_models['spk_stat'].chunk(2) 181 | return torch.randn(dim, device=std.device) * std + mean 182 | 183 | def init_normalizer(self, lang): 184 | 185 | if lang not in self.normalizer: 186 | if lang == 'zh': 187 | try: 188 | from tn.chinese.normalizer import Normalizer 189 | except: 190 | self.logger.log(logging.WARNING, f'Package WeTextProcessing not found! \ 191 | Run: conda install -c conda-forge pynini=2.1.5 && pip install WeTextProcessing') 192 | self.normalizer[lang] = Normalizer().normalize 193 | else: 194 | try: 195 | from nemo_text_processing.text_normalization.normalize import Normalizer 196 | except: 197 | self.logger.log(logging.WARNING, f'Package nemo_text_processing not found! \ 198 | Run: conda install -c conda-forge pynini=2.1.5 && pip install nemo_text_processing') 199 | self.normalizer[lang] = partial(Normalizer(input_case='cased', lang=lang).normalize, verbose=False, punct_post_process=True) 200 | 201 | -------------------------------------------------------------------------------- /ChatTTS/experimental/llm.py: -------------------------------------------------------------------------------- 1 | 2 | from openai import OpenAI 3 | 4 | prompt_dict = { 5 | 'kimi': [ {"role": "system", "content": "你是 Kimi，由 Moonshot AI 提供的人工智能助手，你更擅长中文和英文的对话。"}, 6 | {"role": "user", "content": "你好，请注意你现在生成的文字要按照人日常生活的口吻，你的回复将会后续用TTS模型转为语音，并且请把回答控制在100字以内。并且标点符号仅包含逗号和句号，将数字等转为文字回答。"}, 7 | {"role": "assistant", "content": "好的，我现在生成的文字将按照人日常生活的口吻， 并且我会把回答控制在一百字以内, 标点符号仅包含逗号和句号，将阿拉伯数字等转为中文文字回答。下面请开始对话。"},], 8 | 'deepseek': [ 9 | {"role": "system", "content": "You are a helpful assistant"}, 10 | {"role": "user", "content": "你好，请注意你现在生成的文字要按照人日常生活的口吻，你的回复将会后续用TTS模型转为语音，并且请把回答控制在100字以内。并且标点符号仅包含逗号和句号，将数字等转为文字回答。"}, 11 | {"role": "assistant", "content": "好的，我现在生成的文字将按照人日常生活的口吻， 并且我会把回答控制在一百字以内, 标点符号仅包含逗号和句号，将阿拉伯数字等转为中文文字回答。下面请开始对话。"},], 12 | 'deepseek_TN': [ 13 | {"role": "system", "content": "You are a helpful assistant"}, 14 | {"role": "user", "content": "你好，现在我们在处理TTS的文本输入，下面将会给你输入一段文本，请你将其中的阿拉伯数字等等转为文字表达，并且输出的文本里仅包含逗号和句号这两个标点符号"}, 15 | {"role": "assistant", "content": "好的，我现在对TTS的文本输入进行处理。这一般叫做text normalization。下面请输入"}, 16 | {"role": "user", "content": "We paid $123 for this desk."}, 17 | {"role": "assistant", "content": "We paid one hundred and twenty three dollars for this desk."}, 18 | {"role": "user", "content": "详询请拨打010-724654"}, 19 | {"role": "assistant", "content": "详询请拨打零幺零，七二四六五四"}, 20 | {"role": "user", "content": "罗森宣布将于7月24日退市，在华门店超6000家！"}, 21 | {"role": "assistant", "content": "罗森宣布将于七月二十四日退市，在华门店超过六千家。"}, 22 | ], 23 | } 24 | 25 | class llm_api: 26 | def __init__(self, api_key, base_url, model): 27 | self.client = OpenAI( 28 | api_key = api_key, 29 | base_url = base_url, 30 | ) 31 | self.model = model 32 | def call(self, user_question, temperature = 0.3, prompt_version='kimi', **kwargs): 33 | 34 | completion = self.client.chat.completions.create( 35 | model = self.model, 36 | messages = prompt_dict[prompt_version]+[{"role": "user", "content": user_question},], 37 | temperature = temperature, 38 | **kwargs 39 | ) 40 | return completion.choices[0].message.content 41 | -------------------------------------------------------------------------------- /ChatTTS/infer/api.py: -------------------------------------------------------------------------------- 1 | 2 | import torch 3 | import torch.nn.functional as F 4 | from transformers.generation import TopKLogitsWarper, TopPLogitsWarper 5 | from ..utils.infer_utils import CustomRepetitionPenaltyLogitsProcessorRepeat 6 | 7 | def infer_code( 8 | models, 9 | text, 10 | spk_emb = None, 11 | top_P = 0.7, 12 | top_K = 20, 13 | temperature = 0.3, 14 | repetition_penalty = 1.05, 15 | max_new_token = 2048, 16 | **kwargs 17 | ): 18 | 19 | device = next(models['gpt'].parameters()).device 20 | 21 | if not isinstance(text, list): 22 | text = [text] 23 | 24 | if not isinstance(temperature, list): 25 | temperature = [temperature] * models['gpt'].num_vq 26 | 27 | if spk_emb is not None: 28 | text = [f'[Stts][spk_emb]{i}[Ptts]' for i in text] 29 | else: 30 | text = [f'[Stts][empty_spk]{i}[Ptts]' for i in text] 31 | 32 | text_token = models['tokenizer'](text, return_tensors='pt', add_special_tokens=False, padding=True).to(device) 33 | input_ids = text_token['input_ids'][...,None].expand(-1, -1, models['gpt'].num_vq) 34 | text_mask = torch.ones(text_token['input_ids'].shape, dtype=bool, device=device) 35 | 36 | inputs = { 37 | 'input_ids': input_ids, 38 | 'text_mask': text_mask, 39 | 'attention_mask': text_token['attention_mask'], 40 | } 41 | 42 | emb = models['gpt'].get_emb(**inputs) 43 | if spk_emb is not None: 44 | emb[inputs['input_ids'][..., 0] == models['tokenizer'].convert_tokens_to_ids('[spk_emb]')] = \ 45 | F.normalize(spk_emb.to(device).to(emb.dtype)[None].expand(len(text), -1), p=2.0, dim=1, eps=1e-12) 46 | 47 | num_code = models['gpt'].emb_code[0].num_embeddings - 1 48 | 49 | LogitsWarpers = [] 50 | if top_P is not None: 51 | LogitsWarpers.append(TopPLogitsWarper(top_P, min_tokens_to_keep=3)) 52 | if top_K is not None: 53 | LogitsWarpers.append(TopKLogitsWarper(top_K, min_tokens_to_keep=3)) 54 | 55 | LogitsProcessors = [] 56 | if repetition_penalty is not None and repetition_penalty != 1: 57 | LogitsProcessors.append(CustomRepetitionPenaltyLogitsProcessorRepeat(\ 58 | repetition_penalty, num_code, 16)) 59 | 60 | result = models['gpt'].generate( 61 | emb, inputs['input_ids'], 62 | temperature = torch.tensor(temperature, device=device), 63 | attention_mask = inputs['attention_mask'], 64 | LogitsWarpers = LogitsWarpers, 65 | LogitsProcessors = LogitsProcessors, 66 | eos_token = num_code, 67 | max_new_token = max_new_token, 68 | infer_text = False, 69 | **kwargs 70 | ) 71 | 72 | return result 73 | 74 | 75 | def refine_text( 76 | models, 77 | text, 78 | top_P = 0.7, 79 | top_K = 20, 80 | temperature = 0.7, 81 | repetition_penalty = 1.0, 82 | max_new_token = 384, 83 | prompt = '', 84 | **kwargs 85 | ): 86 | 87 | device = next(models['gpt'].parameters()).device 88 | 89 | if not isinstance(text, list): 90 | text = [text] 91 | 92 | assert len(text), 'text should not be empty' 93 | 94 | text = [f"[Sbreak]{i}[Pbreak]{prompt}" for i in text] 95 | text_token = models['tokenizer'](text, return_tensors='pt', add_special_tokens=False, padding=True).to(device) 96 | text_mask = torch.ones(text_token['input_ids'].shape, dtype=bool, device=device) 97 | 98 | inputs = { 99 | 'input_ids': text_token['input_ids'][...,None].expand(-1, -1, models['gpt'].num_vq), 100 | 'text_mask': text_mask, 101 | 'attention_mask': text_token['attention_mask'], 102 | } 103 | 104 | LogitsWarpers = [] 105 | if top_P is not None: 106 | LogitsWarpers.append(TopPLogitsWarper(top_P, min_tokens_to_keep=3)) 107 | if top_K is not None: 108 | LogitsWarpers.append(TopKLogitsWarper(top_K, min_tokens_to_keep=3)) 109 | 110 | LogitsProcessors = [] 111 | if repetition_penalty is not None and repetition_penalty != 1: 112 | LogitsProcessors.append(CustomRepetitionPenaltyLogitsProcessorRepeat(repetition_penalty, len(models['tokenizer']), 16)) 113 | 114 | result = models['gpt'].generate( 115 | models['gpt'].get_emb(**inputs), inputs['input_ids'], 116 | temperature = torch.tensor([temperature,], device=device), 117 | attention_mask = inputs['attention_mask'], 118 | LogitsWarpers = LogitsWarpers, 119 | LogitsProcessors = LogitsProcessors, 120 | eos_token = torch.tensor(models['tokenizer'].convert_tokens_to_ids('[Ebreak]'), device=device)[None], 121 | max_new_token = max_new_token, 122 | infer_text = True, 123 | **kwargs 124 | ) 125 | return result -------------------------------------------------------------------------------- /ChatTTS/model/dvae.py: -------------------------------------------------------------------------------- 1 | import math 2 | from einops import rearrange 3 | from vector_quantize_pytorch import GroupedResidualFSQ 4 | 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | class ConvNeXtBlock(nn.Module): 10 | def __init__( 11 | self, 12 | dim: int, 13 | intermediate_dim: int, 14 | kernel, dilation, 15 | layer_scale_init_value: float = 1e-6, 16 | ): 17 | # ConvNeXt Block copied from Vocos. 18 | super().__init__() 19 | self.dwconv = nn.Conv1d(dim, dim, 20 | kernel_size=kernel, padding=dilation*(kernel//2), 21 | dilation=dilation, groups=dim 22 | ) # depthwise conv 23 | 24 | self.norm = nn.LayerNorm(dim, eps=1e-6) 25 | self.pwconv1 = nn.Linear(dim, intermediate_dim) # pointwise/1x1 convs, implemented with linear layers 26 | self.act = nn.GELU() 27 | self.pwconv2 = nn.Linear(intermediate_dim, dim) 28 | self.gamma = ( 29 | nn.Parameter(layer_scale_init_value * torch.ones(dim), requires_grad=True) 30 | if layer_scale_init_value > 0 31 | else None 32 | ) 33 | 34 | def forward(self, x: torch.Tensor, cond = None) -> torch.Tensor: 35 | residual = x 36 | x = self.dwconv(x) 37 | x = x.transpose(1, 2) # (B, C, T) -> (B, T, C) 38 | x = self.norm(x) 39 | x = self.pwconv1(x) 40 | x = self.act(x) 41 | x = self.pwconv2(x) 42 | if self.gamma is not None: 43 | x = self.gamma * x 44 | x = x.transpose(1, 2) # (B, T, C) -> (B, C, T) 45 | 46 | x = residual + x 47 | return x 48 | 49 | 50 | 51 | class GFSQ(nn.Module): 52 | 53 | def __init__(self, 54 | dim, levels, G, R, eps=1e-5, transpose = True 55 | ): 56 | super(GFSQ, self).__init__() 57 | self.quantizer = GroupedResidualFSQ( 58 | dim=dim, 59 | levels=levels, 60 | num_quantizers=R, 61 | groups=G, 62 | ) 63 | self.n_ind = math.prod(levels) 64 | self.eps = eps 65 | self.transpose = transpose 66 | self.G = G 67 | self.R = R 68 | 69 | def _embed(self, x): 70 | if self.transpose: 71 | x = x.transpose(1,2) 72 | x = rearrange( 73 | x, "b t (g r) -> g b t r", g = self.G, r = self.R, 74 | ) 75 | feat = self.quantizer.get_output_from_indices(x) 76 | return feat.transpose(1,2) if self.transpose else feat 77 | 78 | def forward(self, x,): 79 | if self.transpose: 80 | x = x.transpose(1,2) 81 | feat, ind = self.quantizer(x) 82 | ind = rearrange( 83 | ind, "g b t r ->b t (g r)", 84 | ) 85 | embed_onehot = F.one_hot(ind.long(), self.n_ind).to(x.dtype) 86 | e_mean = torch.mean(embed_onehot, dim=[0,1]) 87 | e_mean = e_mean / (e_mean.sum(dim=1) + self.eps).unsqueeze(1) 88 | perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + self.eps), dim=1)) 89 | 90 | return ( 91 | torch.zeros(perplexity.shape, dtype=x.dtype, device=x.device), 92 | feat.transpose(1,2) if self.transpose else feat, 93 | perplexity, 94 | None, 95 | ind.transpose(1,2) if self.transpose else ind, 96 | ) 97 | 98 | class DVAEDecoder(nn.Module): 99 | def __init__(self, idim, odim, 100 | n_layer = 12, bn_dim = 64, hidden = 256, 101 | kernel = 7, dilation = 2, up = False 102 | ): 103 | super().__init__() 104 | self.up = up 105 | self.conv_in = nn.Sequential( 106 | nn.Conv1d(idim, bn_dim, 3, 1, 1), nn.GELU(), 107 | nn.Conv1d(bn_dim, hidden, 3, 1, 1) 108 | ) 109 | self.decoder_block = nn.ModuleList([ 110 | ConvNeXtBlock(hidden, hidden* 4, kernel, dilation,) 111 | for _ in range(n_layer)]) 112 | self.conv_out = nn.Conv1d(hidden, odim, kernel_size=1, bias=False) 113 | 114 | def forward(self, input, conditioning=None): 115 | # B, T, C 116 | x = input.transpose(1, 2) 117 | x = self.conv_in(x) 118 | for f in self.decoder_block: 119 | x = f(x, conditioning) 120 | 121 | x = self.conv_out(x) 122 | return x.transpose(1, 2) 123 | 124 | 125 | class DVAE(nn.Module): 126 | def __init__( 127 | self, decoder_config, vq_config, dim=512 128 | ): 129 | super().__init__() 130 | self.register_buffer('coef', torch.randn(1, 100, 1)) 131 | 132 | self.decoder = DVAEDecoder(**decoder_config) 133 | self.out_conv = nn.Conv1d(dim, 100, 3, 1, 1, bias=False) 134 | if vq_config is not None: 135 | self.vq_layer = GFSQ(**vq_config) 136 | else: 137 | self.vq_layer = None 138 | 139 | def forward(self, inp): 140 | 141 | if self.vq_layer is not None: 142 | vq_feats = self.vq_layer._embed(inp) 143 | else: 144 | vq_feats = inp.detach().clone() 145 | 146 | temp = torch.chunk(vq_feats, 2, dim=1) # flatten trick :) 147 | temp = torch.stack(temp, -1) 148 | vq_feats = temp.reshape(*temp.shape[:2], -1) 149 | 150 | vq_feats = vq_feats.transpose(1, 2) 151 | dec_out = self.decoder(input=vq_feats) 152 | dec_out = self.out_conv(dec_out.transpose(1, 2)) 153 | mel = dec_out * self.coef 154 | 155 | return mel 156 | -------------------------------------------------------------------------------- /ChatTTS/model/gpt.py: -------------------------------------------------------------------------------- 1 | import os 2 | os.environ["TOKENIZERS_PARALLELISM"] = "false" 3 | 4 | import logging 5 | from tqdm import tqdm 6 | from einops import rearrange 7 | from transformers.cache_utils import Cache 8 | 9 | import torch 10 | import torch.nn as nn 11 | import torch.nn.functional as F 12 | import torch.nn.utils.parametrize as P 13 | from torch.nn.utils.parametrizations import weight_norm 14 | from transformers import LlamaModel, LlamaConfig 15 | 16 | 17 | class LlamaMLP(nn.Module): 18 | def __init__(self, hidden_size, intermediate_size): 19 | super().__init__() 20 | self.hidden_size = hidden_size 21 | self.intermediate_size = intermediate_size 22 | self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) 23 | self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) 24 | self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) 25 | self.act_fn = F.silu 26 | 27 | def forward(self, x): 28 | down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) 29 | return down_proj 30 | 31 | 32 | class GPT_warpper(nn.Module): 33 | def __init__( 34 | self, 35 | gpt_config, 36 | num_audio_tokens, 37 | num_text_tokens, 38 | num_vq=4, 39 | **kwargs, 40 | ): 41 | super().__init__() 42 | 43 | self.logger = logging.getLogger(__name__) 44 | self.gpt = self.build_model(gpt_config) 45 | self.model_dim = self.gpt.config.hidden_size 46 | 47 | self.num_vq = num_vq 48 | self.emb_code = nn.ModuleList([nn.Embedding(num_audio_tokens, self.model_dim) for i in range(self.num_vq)]) 49 | self.emb_text = nn.Embedding(num_text_tokens, self.model_dim) 50 | self.head_text = weight_norm(nn.Linear(self.model_dim, num_text_tokens, bias=False), name='weight') 51 | self.head_code = nn.ModuleList([weight_norm(nn.Linear(self.model_dim, num_audio_tokens, bias=False), name='weight') for i in range(self.num_vq)]) 52 | 53 | def build_model(self, config): 54 | 55 | configuration = LlamaConfig(**config) 56 | model = LlamaModel(configuration) 57 | del model.embed_tokens 58 | 59 | return model 60 | 61 | def get_emb(self, input_ids, text_mask, **kwargs): 62 | 63 | emb_text = self.emb_text(input_ids[text_mask][:, 0]) 64 | 65 | emb_code = [self.emb_code[i](input_ids[~text_mask][:, i]) for i in range(self.num_vq)] 66 | emb_code = torch.stack(emb_code, 2).sum(2) 67 | 68 | emb = torch.zeros((input_ids.shape[:-1])+(emb_text.shape[-1],), device=emb_text.device, dtype=emb_text.dtype) 69 | emb[text_mask] = emb_text 70 | emb[~text_mask] = emb_code.to(emb.dtype) 71 | 72 | return emb 73 | 74 | def prepare_inputs_for_generation( 75 | self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, **kwargs 76 | ): 77 | # With static cache, the `past_key_values` is None 78 | # TODO joao: standardize interface for the different Cache classes and remove of this if 79 | has_static_cache = False 80 | if past_key_values is None: 81 | past_key_values = getattr(self.gpt.layers[0].self_attn, "past_key_value", None) 82 | has_static_cache = past_key_values is not None 83 | 84 | past_length = 0 85 | if past_key_values is not None: 86 | if isinstance(past_key_values, Cache): 87 | past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() 88 | max_cache_length = ( 89 | torch.tensor(past_key_values.get_max_length(), device=input_ids.device) 90 | if past_key_values.get_max_length() is not None 91 | else None 92 | ) 93 | cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) 94 | # TODO joao: remove this `else` after `generate` prioritizes `Cache` objects 95 | else: 96 | cache_length = past_length = past_key_values[0][0].shape[2] 97 | max_cache_length = None 98 | 99 | # Keep only the unprocessed tokens: 100 | # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where 101 | # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as 102 | # input) 103 | if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: 104 | input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] 105 | # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard 106 | # input_ids based on the past_length. 107 | elif past_length < input_ids.shape[1]: 108 | input_ids = input_ids[:, past_length:] 109 | # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. 110 | 111 | # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. 112 | if ( 113 | max_cache_length is not None 114 | and attention_mask is not None 115 | and cache_length + input_ids.shape[1] > max_cache_length 116 | ): 117 | attention_mask = attention_mask[:, -max_cache_length:] 118 | 119 | position_ids = kwargs.get("position_ids", None) 120 | if attention_mask is not None and position_ids is None: 121 | # create position_ids on the fly for batch generation 122 | position_ids = attention_mask.long().cumsum(-1) - 1 123 | position_ids.masked_fill_(attention_mask == 0, 1) 124 | if past_key_values: 125 | position_ids = position_ids[:, -input_ids.shape[1] :] 126 | 127 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step 128 | if inputs_embeds is not None and past_key_values is None: 129 | model_inputs = {"inputs_embeds": inputs_embeds} 130 | else: 131 | # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise 132 | # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114 133 | # TODO: use `next_tokens` directly instead. 134 | model_inputs = {"input_ids": input_ids.contiguous()} 135 | 136 | input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] 137 | if cache_position is None: 138 | cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) 139 | else: 140 | cache_position = cache_position[-input_length:] 141 | 142 | if has_static_cache: 143 | past_key_values = None 144 | 145 | model_inputs.update( 146 | { 147 | "position_ids": position_ids, 148 | "cache_position": cache_position, 149 | "past_key_values": past_key_values, 150 | "use_cache": kwargs.get("use_cache"), 151 | "attention_mask": attention_mask, 152 | } 153 | ) 154 | return model_inputs 155 | 156 | def generate( 157 | self, 158 | emb, 159 | inputs_ids, 160 | temperature, 161 | eos_token, 162 | attention_mask = None, 163 | max_new_token = 2048, 164 | min_new_token = 0, 165 | LogitsWarpers = [], 166 | LogitsProcessors = [], 167 | infer_text=False, 168 | return_attn=False, 169 | return_hidden=False, 170 | ): 171 | 172 | with torch.no_grad(): 173 | 174 | attentions = [] 175 | hiddens = [] 176 | 177 | start_idx, end_idx = inputs_ids.shape[1], torch.zeros(inputs_ids.shape[0], device=inputs_ids.device, dtype=torch.long) 178 | finish = torch.zeros(inputs_ids.shape[0], device=inputs_ids.device).bool() 179 | 180 | temperature = temperature[None].expand(inputs_ids.shape[0], -1) 181 | temperature = rearrange(temperature, "b n -> (b n) 1") 182 | 183 | attention_mask_cache = torch.ones((inputs_ids.shape[0], inputs_ids.shape[1]+max_new_token,), dtype=torch.bool, device=inputs_ids.device) 184 | if attention_mask is not None: 185 | attention_mask_cache[:, :attention_mask.shape[1]] = attention_mask 186 | 187 | for i in tqdm(range(max_new_token)): 188 | 189 | model_input = self.prepare_inputs_for_generation(inputs_ids, 190 | outputs.past_key_values if i!=0 else None, 191 | attention_mask_cache[:, :inputs_ids.shape[1]], use_cache=True) 192 | 193 | if i == 0: 194 | model_input['inputs_embeds'] = emb 195 | else: 196 | if infer_text: 197 | model_input['inputs_embeds'] = self.emb_text(model_input['input_ids'][:,:,0]) 198 | else: 199 | code_emb = [self.emb_code[i](model_input['input_ids'][:,:,i]) for i in range(self.num_vq)] 200 | model_input['inputs_embeds'] = torch.stack(code_emb, 3).sum(3) 201 | 202 | model_input['input_ids'] = None 203 | outputs = self.gpt.forward(**model_input, output_attentions=return_attn) 204 | attentions.append(outputs.attentions) 205 | hidden_states = outputs[0] # 🐻 206 | if return_hidden: 207 | hiddens.append(hidden_states[:, -1]) 208 | 209 | with P.cached(): 210 | if infer_text: 211 | logits = self.head_text(hidden_states) 212 | else: 213 | logits = torch.stack([self.head_code[i](hidden_states) for i in range(self.num_vq)], 3) 214 | 215 | logits = logits[:, -1].float() 216 | 217 | if not infer_text: 218 | logits = rearrange(logits, "b c n -> (b n) c") 219 | logits_token = rearrange(inputs_ids[:, start_idx:], "b c n -> (b n) c") 220 | else: 221 | logits_token = inputs_ids[:, start_idx:, 0] 222 | 223 | logits = logits / temperature 224 | 225 | for logitsProcessors in LogitsProcessors: 226 | logits = logitsProcessors(logits_token, logits) 227 | 228 | for logitsWarpers in LogitsWarpers: 229 | logits = logitsWarpers(logits_token, logits) 230 | 231 | if i < min_new_token: 232 | logits[:, eos_token] = -torch.inf 233 | 234 | scores = F.softmax(logits, dim=-1) 235 | 236 | idx_next = torch.multinomial(scores, num_samples=1) 237 | 238 | if not infer_text: 239 | idx_next = rearrange(idx_next, "(b n) 1 -> b n", n=self.num_vq) 240 | finish = finish | (idx_next == eos_token).any(1) 241 | inputs_ids = torch.cat([inputs_ids, idx_next.unsqueeze(1)], 1) 242 | else: 243 | finish = finish | (idx_next == eos_token).any(1) 244 | inputs_ids = torch.cat([inputs_ids, idx_next.unsqueeze(-1).expand(-1, -1, self.num_vq)], 1) 245 | 246 | end_idx = end_idx + (~finish).int() 247 | 248 | if finish.all(): 249 | break 250 | 251 | inputs_ids = [inputs_ids[idx, start_idx: start_idx+i] for idx, i in enumerate(end_idx.int())] 252 | inputs_ids = [i[:, 0] for i in inputs_ids] if infer_text else inputs_ids 253 | 254 | if return_hidden: 255 | hiddens = torch.stack(hiddens, 1) 256 | hiddens = [hiddens[idx, :i] for idx, i in enumerate(end_idx.int())] 257 | 258 | if not finish.all(): 259 | self.logger.warn(f'Incomplete result. hit max_new_token: {max_new_token}') 260 | 261 | return { 262 | 'ids': inputs_ids, 263 | 'attentions': attentions, 264 | 'hiddens':hiddens, 265 | } -------------------------------------------------------------------------------- /ChatTTS/utils/gpu_utils.py: -------------------------------------------------------------------------------- 1 | 2 | import torch 3 | import logging 4 | 5 | def select_device(min_memory = 2048): 6 | logger = logging.getLogger(__name__) 7 | if torch.cuda.is_available(): 8 | available_gpus = [] 9 | for i in range(torch.cuda.device_count()): 10 | props = torch.cuda.get_device_properties(i) 11 | free_memory = props.total_memory - torch.cuda.memory_reserved(i) 12 | available_gpus.append((i, free_memory)) 13 | selected_gpu, max_free_memory = max(available_gpus, key=lambda x: x[1]) 14 | device = torch.device(f'cuda:{selected_gpu}') 15 | free_memory_mb = max_free_memory / (1024 * 1024) 16 | if free_memory_mb < min_memory: 17 | logger.log(logging.WARNING, f'GPU {selected_gpu} has {round(free_memory_mb, 2)} MB memory left.') 18 | device = torch.device('cpu') 19 | else: 20 | logger.log(logging.WARNING, f'No GPU found, use CPU instead') 21 | device = torch.device('cpu') 22 | 23 | return device 24 | -------------------------------------------------------------------------------- /ChatTTS/utils/infer_utils.py: -------------------------------------------------------------------------------- 1 | 2 | import re 3 | import torch 4 | import torch.nn.functional as F 5 | 6 | 7 | class CustomRepetitionPenaltyLogitsProcessorRepeat(): 8 | 9 | def __init__(self, penalty: float, max_input_ids, past_window): 10 | if not isinstance(penalty, float) or not (penalty > 0): 11 | raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}") 12 | 13 | self.penalty = penalty 14 | self.max_input_ids = max_input_ids 15 | self.past_window = past_window 16 | 17 | def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor: 18 | 19 | input_ids = input_ids[:, -self.past_window:] 20 | freq = F.one_hot(input_ids, scores.size(1)).sum(1) 21 | freq[self.max_input_ids:] = 0 22 | alpha = self.penalty**freq 23 | scores = torch.where(scores < 0, scores*alpha, scores/alpha) 24 | 25 | return scores 26 | 27 | class CustomRepetitionPenaltyLogitsProcessor(): 28 | 29 | def __init__(self, penalty: float, max_input_ids, past_window): 30 | if not isinstance(penalty, float) or not (penalty > 0): 31 | raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}") 32 | 33 | self.penalty = penalty 34 | self.max_input_ids = max_input_ids 35 | self.past_window = past_window 36 | 37 | def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor: 38 | 39 | input_ids = input_ids[:, -self.past_window:] 40 | score = torch.gather(scores, 1, input_ids) 41 | _score = score.detach().clone() 42 | score = torch.where(score < 0, score * self.penalty, score / self.penalty) 43 | score[input_ids>=self.max_input_ids] = _score[input_ids>=self.max_input_ids] 44 | scores.scatter_(1, input_ids, score) 45 | 46 | return scores 47 | 48 | def count_invalid_characters(s): 49 | 50 | s = re.sub(r'\[uv_break\]|\[laugh\]|\[lbreak\]', '', s) 51 | pattern = re.compile(r'[^\u4e00-\u9fffA-Za-z，。、,\. ]') 52 | non_alphabetic_chinese_chars = pattern.findall(s) 53 | return set(non_alphabetic_chinese_chars) 54 | 55 | def detect_language(sentence): 56 | 57 | chinese_char_pattern = re.compile(r'[\u4e00-\u9fff]') 58 | english_word_pattern = re.compile(r'\b[A-Za-z]+\b') 59 | 60 | chinese_chars = chinese_char_pattern.findall(sentence) 61 | english_words = english_word_pattern.findall(sentence) 62 | 63 | if len(chinese_chars) > len(english_words): 64 | return "zh" 65 | else: 66 | return "en" 67 | 68 | 69 | character_map = { 70 | '：': '，', 71 | '；': '，', 72 | '！': '。', 73 | '（': '，', 74 | '）': '，', 75 | '【': '，', 76 | '】': '，', 77 | '『': '，', 78 | '』': '，', 79 | '「': '，', 80 | '」': '，', 81 | '《': '，', 82 | '》': '，', 83 | '－': '，', 84 | '‘': '', 85 | '“': '', 86 | '’': '', 87 | '”': '', 88 | ':': ',', 89 | ';': ',', 90 | '!': '.', 91 | '(': ',', 92 | ')': ',', 93 | '[': ',', 94 | ']': ',', 95 | '>': ',', 96 | '<': ',', 97 | '-': ',', 98 | } 99 | 100 | halfwidth_2_fullwidth_map = { 101 | '!': '！', 102 | '"': '“', 103 | "'": '‘', 104 | '#': '＃', 105 | '$': '＄', 106 | '%': '％', 107 | '&': '＆', 108 | '(': '（', 109 | ')': '）', 110 | ',': '，', 111 | '-': '－', 112 | '*': '＊', 113 | '+': '＋', 114 | '.': '。', 115 | '/': '／', 116 | ':': '：', 117 | ';': '；', 118 | '<': '＜', 119 | '=': '＝', 120 | '>': '＞', 121 | '?': '？', 122 | '@': '＠', 123 | # '[': '［', 124 | '\\': '＼', 125 | # ']': '］', 126 | '^': '＾', 127 | # '_': '＿', 128 | '`': '｀', 129 | '{': '｛', 130 | '|': '｜', 131 | '}': '｝', 132 | '~': '～' 133 | } 134 | 135 | def apply_half2full_map(text): 136 | translation_table = str.maketrans(halfwidth_2_fullwidth_map) 137 | return text.translate(translation_table) 138 | 139 | def apply_character_map(text): 140 | translation_table = str.maketrans(character_map) 141 | return text.translate(translation_table) -------------------------------------------------------------------------------- /ChatTTS/utils/io_utils.py: -------------------------------------------------------------------------------- 1 | 2 | import os 3 | import logging 4 | 5 | def get_latest_modified_file(directory): 6 | logger = logging.getLogger(__name__) 7 | 8 | files = [os.path.join(directory, f) for f in os.listdir(directory)] 9 | if not files: 10 | logger.log(logging.WARNING, f'No files found in the directory: {directory}') 11 | return None 12 | latest_file = max(files, key=os.path.getmtime) 13 | 14 | return latest_file -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | FROM pytorch/pytorch:2.3.0-cuda11.8-cudnn8-runtime 2 | 3 | ENV DEBIAN_FRONTEND=noninteractive 4 | 5 | RUN apt-get update && apt-get install -y \ 6 | vim \ 7 | curl \ 8 | ca-certificates \ 9 | curl \ 10 | git \ 11 | bzip2 \ 12 | libsndfile1 \ 13 | gcc \ 14 | cmake \ 15 | make \ 16 | g++ \ 17 | vim \ 18 | wget \ 19 | tzdata \ 20 | && apt-get clean \ 21 | && rm -rf /var/lib/apt/lists/* 22 | 23 | 24 | WORKDIR /app 25 | 26 | COPY tts_model /app/tts_model 27 | # 避免跟官方仓库的 requirements 冲突 28 | COPY ./requirements-chat-tts.txt /app/requirements.txt 29 | 30 | RUN pip config set global.index-url https://mirrors.cloud.tencent.com/pypi/simple && \ 31 | pip install --no-cache-dir -r /app/requirements.txt 32 | 33 | COPY ./ChatTTS /app/ChatTTS 34 | 35 | COPY ./requirements-api-ui.txt /app/requirements-api-ui.txt 36 | RUN pip config set global.index-url https://mirrors.cloud.tencent.com/pypi/simple && \ 37 | pip install --no-cache-dir -r /app/requirements-api-ui.txt 38 | 39 | COPY ./api.py /app 40 | COPY ./webui_api.py /app 41 | 42 | ENV LOG_LEVEL=INFO 43 | ENV LOG_FILE=/app/logs/service.log 44 | ENV WORKERS=1 45 | ENV PYTHONIOENCODING=UTF-8 46 | ENV TZ=Asia/Shanghai 47 | ENV LANG=C.UTF-8 48 | ENV API_PORT=8080 49 | 50 | CMD streamlit run webui_api.py --browser.gatherUsageStats False & \ 51 | gunicorn api:app -w ${WORKERS} -k uvicorn.workers.UvicornWorker -b :${API_PORT} 52 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Attribution-NonCommercial 4.0 International 2 | 3 | ======================================================================= 4 | 5 | Creative Commons Corporation ("Creative Commons") is not a law firm and 6 | does not provide legal services or legal advice. Distribution of 7 | Creative Commons public licenses does not create a lawyer-client or 8 | other relationship. Creative Commons makes its licenses and related 9 | information available on an "as-is" basis. Creative Commons gives no 10 | warranties regarding its licenses, any material licensed under their 11 | terms and conditions, or any related information. Creative Commons 12 | disclaims all liability for damages resulting from their use to the 13 | fullest extent possible. 14 | 15 | Using Creative Commons Public Licenses 16 | 17 | Creative Commons public licenses provide a standard set of terms and 18 | conditions that creators and other rights holders may use to share 19 | original works of authorship and other material subject to copyright 20 | and certain other rights specified in the public license below. The 21 | following considerations are for informational purposes only, are not 22 | exhaustive, and do not form part of our licenses. 23 | 24 | Considerations for licensors: Our public licenses are 25 | intended for use by those authorized to give the public 26 | permission to use material in ways otherwise restricted by 27 | copyright and certain other rights. Our licenses are 28 | irrevocable. Licensors should read and understand the terms 29 | and conditions of the license they choose before applying it. 30 | Licensors should also secure all rights necessary before 31 | applying our licenses so that the public can reuse the 32 | material as expected. Licensors should clearly mark any 33 | material not subject to the license. This includes other CC- 34 | licensed material, or material used under an exception or 35 | limitation to copyright. More considerations for licensors: 36 | wiki.creativecommons.org/Considerations_for_licensors 37 | 38 | Considerations for the public: By using one of our public 39 | licenses, a licensor grants the public permission to use the 40 | licensed material under specified terms and conditions. If 41 | the licensor's permission is not necessary for any reason--for 42 | example, because of any applicable exception or limitation to 43 | copyright--then that use is not regulated by the license. Our 44 | licenses grant only permissions under copyright and certain 45 | other rights that a licensor has authority to grant. Use of 46 | the licensed material may still be restricted for other 47 | reasons, including because others have copyright or other 48 | rights in the material. A licensor may make special requests, 49 | such as asking that all changes be marked or described. 50 | Although not required by our licenses, you are encouraged to 51 | respect those requests where reasonable. More considerations 52 | for the public: 53 | wiki.creativecommons.org/Considerations_for_licensees 54 | 55 | ======================================================================= 56 | 57 | Creative Commons Attribution-NonCommercial 4.0 International Public 58 | License 59 | 60 | By exercising the Licensed Rights (defined below), You accept and agree 61 | to be bound by the terms and conditions of this Creative Commons 62 | Attribution-NonCommercial 4.0 International Public License ("Public 63 | License"). To the extent this Public License may be interpreted as a 64 | contract, You are granted the Licensed Rights in consideration of Your 65 | acceptance of these terms and conditions, and the Licensor grants You 66 | such rights in consideration of benefits the Licensor receives from 67 | making the Licensed Material available under these terms and 68 | conditions. 69 | 70 | 71 | Section 1 -- Definitions. 72 | 73 | a. Adapted Material means material subject to Copyright and Similar 74 | Rights that is derived from or based upon the Licensed Material 75 | and in which the Licensed Material is translated, altered, 76 | arranged, transformed, or otherwise modified in a manner requiring 77 | permission under the Copyright and Similar Rights held by the 78 | Licensor. For purposes of this Public License, where the Licensed 79 | Material is a musical work, performance, or sound recording, 80 | Adapted Material is always produced where the Licensed Material is 81 | synched in timed relation with a moving image. 82 | 83 | b. Adapter's License means the license You apply to Your Copyright 84 | and Similar Rights in Your contributions to Adapted Material in 85 | accordance with the terms and conditions of this Public License. 86 | 87 | c. Copyright and Similar Rights means copyright and/or similar rights 88 | closely related to copyright including, without limitation, 89 | performance, broadcast, sound recording, and Sui Generis Database 90 | Rights, without regard to how the rights are labeled or 91 | categorized. For purposes of this Public License, the rights 92 | specified in Section 2(b)(1)-(2) are not Copyright and Similar 93 | Rights. 94 | d. Effective Technological Measures means those measures that, in the 95 | absence of proper authority, may not be circumvented under laws 96 | fulfilling obligations under Article 11 of the WIPO Copyright 97 | Treaty adopted on December 20, 1996, and/or similar international 98 | agreements. 99 | 100 | e. Exceptions and Limitations means fair use, fair dealing, and/or 101 | any other exception or limitation to Copyright and Similar Rights 102 | that applies to Your use of the Licensed Material. 103 | 104 | f. Licensed Material means the artistic or literary work, database, 105 | or other material to which the Licensor applied this Public 106 | License. 107 | 108 | g. Licensed Rights means the rights granted to You subject to the 109 | terms and conditions of this Public License, which are limited to 110 | all Copyright and Similar Rights that apply to Your use of the 111 | Licensed Material and that the Licensor has authority to license. 112 | 113 | h. Licensor means the individual(s) or entity(ies) granting rights 114 | under this Public License. 115 | 116 | i. NonCommercial means not primarily intended for or directed towards 117 | commercial advantage or monetary compensation. For purposes of 118 | this Public License, the exchange of the Licensed Material for 119 | other material subject to Copyright and Similar Rights by digital 120 | file-sharing or similar means is NonCommercial provided there is 121 | no payment of monetary compensation in connection with the 122 | exchange. 123 | 124 | j. Share means to provide material to the public by any means or 125 | process that requires permission under the Licensed Rights, such 126 | as reproduction, public display, public performance, distribution, 127 | dissemination, communication, or importation, and to make material 128 | available to the public including in ways that members of the 129 | public may access the material from a place and at a time 130 | individually chosen by them. 131 | 132 | k. Sui Generis Database Rights means rights other than copyright 133 | resulting from Directive 96/9/EC of the European Parliament and of 134 | the Council of 11 March 1996 on the legal protection of databases, 135 | as amended and/or succeeded, as well as other essentially 136 | equivalent rights anywhere in the world. 137 | 138 | l. You means the individual or entity exercising the Licensed Rights 139 | under this Public License. Your has a corresponding meaning. 140 | 141 | 142 | Section 2 -- Scope. 143 | 144 | a. License grant. 145 | 146 | 1. Subject to the terms and conditions of this Public License, 147 | the Licensor hereby grants You a worldwide, royalty-free, 148 | non-sublicensable, non-exclusive, irrevocable license to 149 | exercise the Licensed Rights in the Licensed Material to: 150 | 151 | a. reproduce and Share the Licensed Material, in whole or 152 | in part, for NonCommercial purposes only; and 153 | 154 | b. produce, reproduce, and Share Adapted Material for 155 | NonCommercial purposes only. 156 | 157 | 2. Exceptions and Limitations. For the avoidance of doubt, where 158 | Exceptions and Limitations apply to Your use, this Public 159 | License does not apply, and You do not need to comply with 160 | its terms and conditions. 161 | 162 | 3. Term. The term of this Public License is specified in Section 163 | 6(a). 164 | 165 | 4. Media and formats; technical modifications allowed. The 166 | Licensor authorizes You to exercise the Licensed Rights in 167 | all media and formats whether now known or hereafter created, 168 | and to make technical modifications necessary to do so. The 169 | Licensor waives and/or agrees not to assert any right or 170 | authority to forbid You from making technical modifications 171 | necessary to exercise the Licensed Rights, including 172 | technical modifications necessary to circumvent Effective 173 | Technological Measures. For purposes of this Public License, 174 | simply making modifications authorized by this Section 2(a) 175 | (4) never produces Adapted Material. 176 | 177 | 5. Downstream recipients. 178 | 179 | a. Offer from the Licensor -- Licensed Material. Every 180 | recipient of the Licensed Material automatically 181 | receives an offer from the Licensor to exercise the 182 | Licensed Rights under the terms and conditions of this 183 | Public License. 184 | 185 | b. No downstream restrictions. You may not offer or impose 186 | any additional or different terms or conditions on, or 187 | apply any Effective Technological Measures to, the 188 | Licensed Material if doing so restricts exercise of the 189 | Licensed Rights by any recipient of the Licensed 190 | Material. 191 | 192 | 6. No endorsement. Nothing in this Public License constitutes or 193 | may be construed as permission to assert or imply that You 194 | are, or that Your use of the Licensed Material is, connected 195 | with, or sponsored, endorsed, or granted official status by, 196 | the Licensor or others designated to receive attribution as 197 | provided in Section 3(a)(1)(A)(i). 198 | 199 | b. Other rights. 200 | 201 | 1. Moral rights, such as the right of integrity, are not 202 | licensed under this Public License, nor are publicity, 203 | privacy, and/or other similar personality rights; however, to 204 | the extent possible, the Licensor waives and/or agrees not to 205 | assert any such rights held by the Licensor to the limited 206 | extent necessary to allow You to exercise the Licensed 207 | Rights, but not otherwise. 208 | 209 | 2. Patent and trademark rights are not licensed under this 210 | Public License. 211 | 212 | 3. To the extent possible, the Licensor waives any right to 213 | collect royalties from You for the exercise of the Licensed 214 | Rights, whether directly or through a collecting society 215 | under any voluntary or waivable statutory or compulsory 216 | licensing scheme. In all other cases the Licensor expressly 217 | reserves any right to collect such royalties, including when 218 | the Licensed Material is used other than for NonCommercial 219 | purposes. 220 | 221 | 222 | Section 3 -- License Conditions. 223 | 224 | Your exercise of the Licensed Rights is expressly made subject to the 225 | following conditions. 226 | 227 | a. Attribution. 228 | 229 | 1. If You Share the Licensed Material (including in modified 230 | form), You must: 231 | 232 | a. retain the following if it is supplied by the Licensor 233 | with the Licensed Material: 234 | 235 | i. identification of the creator(s) of the Licensed 236 | Material and any others designated to receive 237 | attribution, in any reasonable manner requested by 238 | the Licensor (including by pseudonym if 239 | designated); 240 | 241 | ii. a copyright notice; 242 | 243 | iii. a notice that refers to this Public License; 244 | 245 | iv. a notice that refers to the disclaimer of 246 | warranties; 247 | 248 | v. a URI or hyperlink to the Licensed Material to the 249 | extent reasonably practicable; 250 | 251 | b. indicate if You modified the Licensed Material and 252 | retain an indication of any previous modifications; and 253 | 254 | c. indicate the Licensed Material is licensed under this 255 | Public License, and include the text of, or the URI or 256 | hyperlink to, this Public License. 257 | 258 | 2. You may satisfy the conditions in Section 3(a)(1) in any 259 | reasonable manner based on the medium, means, and context in 260 | which You Share the Licensed Material. For example, it may be 261 | reasonable to satisfy the conditions by providing a URI or 262 | hyperlink to a resource that includes the required 263 | information. 264 | 265 | 3. If requested by the Licensor, You must remove any of the 266 | information required by Section 3(a)(1)(A) to the extent 267 | reasonably practicable. 268 | 269 | 4. If You Share Adapted Material You produce, the Adapter's 270 | License You apply must not prevent recipients of the Adapted 271 | Material from complying with this Public License. 272 | 273 | 274 | Section 4 -- Sui Generis Database Rights. 275 | 276 | Where the Licensed Rights include Sui Generis Database Rights that 277 | apply to Your use of the Licensed Material: 278 | 279 | a. for the avoidance of doubt, Section 2(a)(1) grants You the right 280 | to extract, reuse, reproduce, and Share all or a substantial 281 | portion of the contents of the database for NonCommercial purposes 282 | only; 283 | 284 | b. if You include all or a substantial portion of the database 285 | contents in a database in which You have Sui Generis Database 286 | Rights, then the database in which You have Sui Generis Database 287 | Rights (but not its individual contents) is Adapted Material; and 288 | 289 | c. You must comply with the conditions in Section 3(a) if You Share 290 | all or a substantial portion of the contents of the database. 291 | 292 | For the avoidance of doubt, this Section 4 supplements and does not 293 | replace Your obligations under this Public License where the Licensed 294 | Rights include other Copyright and Similar Rights. 295 | 296 | 297 | Section 5 -- Disclaimer of Warranties and Limitation of Liability. 298 | 299 | a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE 300 | EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS 301 | AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF 302 | ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, 303 | IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, 304 | WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR 305 | PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, 306 | ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT 307 | KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT 308 | ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. 309 | 310 | b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE 311 | TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, 312 | NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, 313 | INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, 314 | COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR 315 | USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN 316 | ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR 317 | DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR 318 | IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. 319 | 320 | c. The disclaimer of warranties and limitation of liability provided 321 | above shall be interpreted in a manner that, to the extent 322 | possible, most closely approximates an absolute disclaimer and 323 | waiver of all liability. 324 | 325 | 326 | Section 6 -- Term and Termination. 327 | 328 | a. This Public License applies for the term of the Copyright and 329 | Similar Rights licensed here. However, if You fail to comply with 330 | this Public License, then Your rights under this Public License 331 | terminate automatically. 332 | 333 | b. Where Your right to use the Licensed Material has terminated under 334 | Section 6(a), it reinstates: 335 | 336 | 1. automatically as of the date the violation is cured, provided 337 | it is cured within 30 days of Your discovery of the 338 | violation; or 339 | 340 | 2. upon express reinstatement by the Licensor. 341 | 342 | For the avoidance of doubt, this Section 6(b) does not affect any 343 | right the Licensor may have to seek remedies for Your violations 344 | of this Public License. 345 | 346 | c. For the avoidance of doubt, the Licensor may also offer the 347 | Licensed Material under separate terms or conditions or stop 348 | distributing the Licensed Material at any time; however, doing so 349 | will not terminate this Public License. 350 | 351 | d. Sections 1, 5, 6, 7, and 8 survive termination of this Public 352 | License. 353 | 354 | 355 | Section 7 -- Other Terms and Conditions. 356 | 357 | a. The Licensor shall not be bound by any additional or different 358 | terms or conditions communicated by You unless expressly agreed. 359 | 360 | b. Any arrangements, understandings, or agreements regarding the 361 | Licensed Material not stated herein are separate from and 362 | independent of the terms and conditions of this Public License. 363 | 364 | 365 | Section 8 -- Interpretation. 366 | 367 | a. For the avoidance of doubt, this Public License does not, and 368 | shall not be interpreted to, reduce, limit, restrict, or impose 369 | conditions on any use of the Licensed Material that could lawfully 370 | be made without permission under this Public License. 371 | 372 | b. To the extent possible, if any provision of this Public License is 373 | deemed unenforceable, it shall be automatically reformed to the 374 | minimum extent necessary to make it enforceable. If the provision 375 | cannot be reformed, it shall be severed from this Public License 376 | without affecting the enforceability of the remaining terms and 377 | conditions. 378 | 379 | c. No term or condition of this Public License will be waived and no 380 | failure to comply consented to unless expressly agreed to by the 381 | Licensor. 382 | 383 | d. Nothing in this Public License constitutes or may be interpreted 384 | as a limitation upon, or waiver of, any privileges and immunities 385 | that apply to the Licensor or You, including from the legal 386 | processes of any jurisdiction or authority. 387 | 388 | ======================================================================= 389 | 390 | Creative Commons is not a party to its public 391 | licenses. Notwithstanding, Creative Commons may elect to apply one of 392 | its public licenses to material it publishes and in those instances 393 | will be considered the “Licensor.” The text of the Creative Commons 394 | public licenses is dedicated to the public domain under the CC0 Public 395 | Domain Dedication. Except for the limited purpose of indicating that 396 | material is shared under a Creative Commons public license or as 397 | otherwise permitted by the Creative Commons policies published at 398 | creativecommons.org/policies, Creative Commons does not authorize the 399 | use of the trademark "Creative Commons" or any other trademark or logo 400 | of Creative Commons without its prior written consent including, 401 | without limitation, in connection with any unauthorized modifications 402 | to any of its public licenses or any other arrangements, 403 | understandings, or agreements concerning use of licensed material. For 404 | the avoidance of doubt, this paragraph does not form part of the 405 | public licenses. 406 | 407 | Creative Commons may be contacted at creativecommons.org. 408 | -------------------------------------------------------------------------------- /Makefile: -------------------------------------------------------------------------------- 1 | webui_api: 2 | streamlit run webui_api.py --browser.gatherUsageStats False 3 | 4 | server: 5 | uvicorn api:app --host 0.0.0.0 --port 8080 6 | 7 | client: 8 | python client.py 9 | 10 | image_tag=24.06.06 11 | image_name=jackiexiao/chat_tts_api_ui:${image_tag} 12 | build_docker: 13 | docker build . -f Dockerfile -t ${image_name} 14 | docker tag ${image_name} jackiexiao/chat_tts_api_ui:latest 15 | 16 | 17 | docker_push: 18 | docker push ${image_name} 19 | docker puash jackiexiao/chat_tts_api_ui:latest 20 | 21 | tencent_prefix=ccr.ccs.tencentyun.com/text-to-speech/chat_tts_api_ui 22 | docker_push_tencent: 23 | docker tag ${image_name} ${tencent_prefix}:${image_tag} 24 | docker push ${tencent_prefix}:${image_tag} 25 | docker tag ${image_name} ${tencent_prefix}:latest 26 | docker push ${tencent_prefix}:latest 27 | 28 | run_docker: 29 | docker run --name chat_tts \ 30 | --gpus all --ipc=host \ 31 | --ulimit memlock=-1 --ulimit stack=67108864 \ 32 | -p 8080:8080 -p 8501:8501 \ 33 | ${tencent_prefix}:latest 34 | 35 | save_docker: # docker load -i chat-tts.tar.gz 36 | docker save ${image_name} | gzip > chat-tts.tar.gz 37 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # ChatTTS-api-ui-docker 一键启动! 2 | 3 | - 一行命令启动一个带有 Web 界面的 ChatTTS API 服务器 4 | - 前提条件：支持 CUDA 的 GPU，Docker, 4GB GPU 内存 5 | - 使用腾讯云容器镜像服务, 不用担心无法下载 Docker Image 6 | 7 | ## Usage 使用方法 8 | 9 | 启动服务 (ps: 镜像压缩后 5.16 GB) 10 | ```bash 11 | docker run --name chat_tts \ 12 | --gpus all --ipc=host \ 13 | --ulimit memlock=-1 --ulimit stack=67108864 \ 14 | -p 8080:8080 -p 8501:8501 \ 15 | ccr.ccs.tencentyun.com/text-to-speech/chat_tts_api_ui:latest 16 | ``` 17 | 18 | 删除服务 19 | ``` 20 | docker rm -f chat_tts 21 | ``` 22 | 23 | - 等待服务启动完成后，访问 http://localhost:8501 即可查看 Web 界面。 24 | - Web UI 25 | - URL 网页: http://localhost:8501 26 | - build with streamlit 27 | - Api 28 | - API 地址: http://localhost:8080 29 | - API 文档: http://localhost:8080/docs 30 | - 客户端示例 client: `python client.py` 31 | - 采样率为 24000, 默认返回 mp3 32 | - build with fastapi 33 | 34 | ## 备注 35 | - ChatTTS 的简略流程 36 | - Text -> GPT -> 离散 Token -> VAE Decoder（实际上用的 Hidden latents 不是离散 Token） -> Mel -> Vocos -> Audio 37 | - 优点 38 | - 自回归模型：更高的自然度，可以达到类似 openai 的水平 39 | - 更大的模型更大的数据：通常使用 1-10万小时训练，模型在 0.3B-1B 之间 40 | - 少样本克隆：模型不需要通过微调就可以克隆新的音色（当然也可以微调获得更好的效果）但训练代码还没开源 41 | - 更高的可玩性/可控性：可以通过文本或者音频 prompt 控制发音风格，合成多样性的语音 42 | - 基于 token 而不基于拼音输入的方式可以非常容易扩展到多语种，不需要每个语种开发文本转拼音的工具包 43 | - 缺点 44 | - 训练成本较高（开源版本4万小时, 275M参数量左右，据说有10万小时训练的更大参数量的版本） 45 | - 推理成本较高，推理速度较慢（4090单卡 RTF 0.3），需要后续优化 46 | - 目前合成不够稳定，容易出现多字，漏字，或者出现杂音。但后续迭代可解决 47 | - 目前 Badcase 较多，比如数字，英文字母，标点符号，多音字，部分英文单词常常念错，停顿上错误较多 48 | - 基于 token 的输入：缺点是如果训练数据量不够大，或者缺乏高质量数据，很多生僻字和特殊的人名地名将无法正确合成，多音字也容易出错。 49 | - 目前开源的版本基于4万小时训练的小参数量模型，音质和稳定性上比较一般 50 | - 无法生成超过 30 秒的音频，需要手动分句合成 51 | - 虽然是随机音色，但年龄都在20-45岁之间。没有更年轻或者更老的音色 52 | 53 | 54 | ## 类似项目 55 | > 也提供了 UI 和 Api 56 | - https://github.com/jianchang512/ChatTTS-ui 57 | 58 | # ChatTTS 59 | [**English**](./README.md) | [**中文简体**](./README_CN.md) 60 | 61 | ChatTTS is a text-to-speech model designed specifically for dialogue scenario such as LLM assistant. It supports both English and Chinese languages. Our model is trained with 100,000+ hours composed of chinese and english. The open-source version on **[HuggingFace](https://huggingface.co/2Noise/ChatTTS)** is a 40,000 hours pre trained model without SFT. 62 | 63 | For formal inquiries about model and roadmap, please contact us at **open-source@2noise.com**. You could join our QQ group: ~~808364215 (Full)~~ 230696694 (Group 2) for discussion. Adding github issues is always welcomed. 64 | 65 | --- 66 | ## Highlights 67 | 1. **Conversational TTS**: ChatTTS is optimized for dialogue-based tasks, enabling natural and expressive speech synthesis. It supports multiple speakers, facilitating interactive conversations. 68 | 2. **Fine-grained Control**: The model could predict and control fine-grained prosodic features, including laughter, pauses, and interjections. 69 | 3. **Better Prosody**: ChatTTS surpasses most of open-source TTS models in terms of prosody. We provide pretrained models to support further research and development. 70 | 71 | For the detailed description of the model, you can refer to **[video on Bilibili](https://www.bilibili.com/video/BV1zn4y1o7iV)** 72 | 73 | --- 74 | 75 | ## Disclaimer 76 | 77 | This repo is for academic purposes only. It is intended for educational and research use, and should not be used for any commercial or legal purposes. The authors do not guarantee the accuracy, completeness, or reliability of the information. The information and data used in this repo, are for academic and research purposes only. The data obtained from publicly available sources, and the authors do not claim any ownership or copyright over the data. 78 | 79 | ChatTTS is a powerful text-to-speech system. However, it is very important to utilize this technology responsibly and ethically. To limit the use of ChatTTS, we added a small amount of high-frequency noise during the training of the 40,000-hour model, and compressed the audio quality as much as possible using MP3 format, to prevent malicious actors from potentially using it for criminal purposes. At the same time, we have internally trained a detection model and plan to open-source it in the future. 80 | 81 | 82 | --- 83 | ## Usage 84 | 85 |

Basic usage

86 | 87 | ```python 88 | import ChatTTS 89 | from IPython.display import Audio 90 | 91 | chat = ChatTTS.Chat() 92 | chat.load_models(compile=False) # Set to True for better performance 93 | 94 | texts = ["PUT YOUR TEXT HERE",] 95 | 96 | wavs = chat.infer(texts, ) 97 | 98 | torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000) 99 | ``` 100 | 101 |

Advanced usage

102 | 103 | ```python 104 | ################################### 105 | # Sample a speaker from Gaussian. 106 | 107 | rand_spk = chat.sample_random_speaker() 108 | 109 | params_infer_code = { 110 | 'spk_emb': rand_spk, # add sampled speaker 111 | 'temperature': .3, # using custom temperature 112 | 'top_P': 0.7, # top P decode 113 | 'top_K': 20, # top K decode 114 | } 115 | 116 | ################################### 117 | # For sentence level manual control. 118 | 119 | # use oral_(0-9), laugh_(0-2), break_(0-7) 120 | # to generate special token in text to synthesize. 121 | params_refine_text = { 122 | 'prompt': '[oral_2][laugh_0][break_6]' 123 | } 124 | 125 | wav = chat.infer(texts, params_refine_text=params_refine_text, params_infer_code=params_infer_code) 126 | 127 | ################################### 128 | # For word level manual control. 129 | text = 'What is [uv_break]your favorite english food?[laugh][lbreak]' 130 | wav = chat.infer(text, skip_refine_text=True, params_refine_text=params_refine_text, params_infer_code=params_infer_code) 131 | torchaudio.save("output2.wav", torch.from_numpy(wavs[0]), 24000) 132 | ``` 133 | 134 |
135 |

Example: self introduction

 136 | 137 | ```python 138 | inputs_en = """ 139 | chat T T S is a text to speech model designed for dialogue applications. 140 | [uv_break]it supports mixed language input [uv_break]and offers multi speaker 141 | capabilities with precise control over prosodic elements [laugh]like like 142 | [uv_break]laughter[laugh], [uv_break]pauses, [uv_break]and intonation. 143 | [uv_break]it delivers natural and expressive speech,[uv_break]so please 144 | [uv_break] use the project responsibly at your own risk.[uv_break] 145 | """.replace('\n', '') # English is still experimental. 146 | 147 | params_refine_text = { 148 | 'prompt': '[oral_2][laugh_0][break_4]' 149 | } 150 | # audio_array_cn = chat.infer(inputs_cn, params_refine_text=params_refine_text) 151 | audio_array_en = chat.infer(inputs_en, params_refine_text=params_refine_text) 152 | torchaudio.save("output3.wav", torch.from_numpy(audio_array_en[0]), 24000) 153 | ``` 154 | [male speaker](https://github.com/2noise/ChatTTS/assets/130631963/e0f51251-db7f-4d39-a0e9-3e095bb65de1) 155 | 156 | [female speaker](https://github.com/2noise/ChatTTS/assets/130631963/f5dcdd01-1091-47c5-8241-c4f6aaaa8bbd) 157 |
158 | 159 | --- 160 | ## Roadmap 161 | - [x] Open-source the 40k hour base model and spk_stats file 162 | - [ ] Open-source VQ encoder and Lora training code 163 | - [ ] Streaming audio generation without refining the text* 164 | - [ ] Open-source the 40k hour version with multi-emotion control 165 | - [ ] ChatTTS.cpp maybe? (PR or new repo are welcomed.) 166 | 167 | ---- 168 | ## FAQ 169 | 170 | ##### How much VRAM do I need? How about infer speed? 171 | For a 30-second audio clip, at least 4GB of GPU memory is required. For the 4090 GPU, it can generate audio corresponding to approximately 7 semantic tokens per second. The Real-Time Factor (RTF) is around 0.3. 172 | 173 | ##### model stability is not good enough, with issues such as multi speakers or poor audio quality. 174 | 175 | This is a problem that typically occurs with autoregressive models(for bark and valle). It's generally difficult to avoid. One can try multiple samples to find a suitable result. 176 | 177 | ##### Besides laughter, can we control anything else? Can we control other emotions? 178 | 179 | In the current released model, the only token-level control units are [laugh], [uv_break], and [lbreak]. In future versions, we may open-source models with additional emotional control capabilities. 180 | 181 | --- 182 | ## Acknowledgements 183 | - [bark](https://github.com/suno-ai/bark), [XTTSv2](https://github.com/coqui-ai/TTS) and [valle](https://arxiv.org/abs/2301.02111) demostrate a remarkable TTS result by a autoregressive-style system. 184 | - [fish-speech](https://github.com/fishaudio/fish-speech) reveals capability of GVQ as audio tokenizer for LLM modeling. 185 | - [vocos](https://github.com/gemelo-ai/vocos) which is used as a pretrained vocoder. 186 | 187 | --- 188 | ## Special Appreciation 189 | - [wlu-audio lab](https://audio.westlake.edu.cn/) for early algorithm experiments. 190 | -------------------------------------------------------------------------------- /README_CN.md: -------------------------------------------------------------------------------- 1 | # ChatTTS 2 | [**English**](./README.md) | [**中文简体**](./README_CN.md) 3 | 4 | ChatTTS是专门为对话场景设计的文本转语音模型，例如LLM助手对话任务。它支持英文和中文两种语言。最大的模型使用了10万小时以上的中英文数据进行训练。在HuggingFace中开源的版本为4万小时训练且未SFT的版本. 5 | 6 | 如需就模型进行正式商业咨询，请发送邮件至 **open-source@2noise.com**。对于中文用户，您可以加入我们的QQ群：~~808364215 (已满)~~ 230696694 (二群) 进行讨论。同时欢迎在GitHub上提出问题。如果遇到无法使用 **[HuggingFace](https://huggingface.co/2Noise/ChatTTS)** 的情况,可以在 [modelscope](https://www.modelscope.cn/models/pzc163/chatTTS) 上进行下载. 7 | 8 | --- 9 | ## 亮点 10 | 1. **对话式 TTS**: ChatTTS针对对话式任务进行了优化，实现了自然流畅的语音合成，同时支持多说话人。 11 | 2. **细粒度控制**: 该模型能够预测和控制细粒度的韵律特征，包括笑声、停顿和插入词等。 12 | 3. **更好的韵律**: ChatTTS在韵律方面超越了大部分开源TTS模型。同时提供预训练模型，支持进一步的研究。 13 | 14 | 对于模型的具体介绍, 可以参考B站的 **[宣传视频](https://www.bilibili.com/video/BV1zn4y1o7iV)** 15 | 16 | --- 17 | 18 | ## 免责声明 19 | 本文件中的信息仅供学术交流使用。其目的是用于教育和研究，不得用于任何商业或法律目的。作者不保证信息的准确性、完整性或可靠性。本文件中使用的信息和数据，仅用于学术研究目的。这些数据来自公开可用的来源，作者不对数据的所有权或版权提出任何主张。 20 | 21 | ChatTTS是一个强大的文本转语音系统。然而，负责任地和符合伦理地利用这项技术是非常重要的。为了限制ChatTTS的使用，我们在4w小时模型的训练过程中添加了少量额外的高频噪音，并用mp3格式尽可能压低了音质，以防不法分子用于潜在的犯罪可能。同时我们在内部训练了检测模型，并计划在未来开放。 22 | 23 | --- 24 | ## 用法 25 | 26 |

基本用法

27 | 28 | ```python 29 | import ChatTTS 30 | from IPython.display import Audio 31 | 32 | chat = ChatTTS.Chat() 33 | chat.load_models(compile=False) # 设置为True以获得更快速度 34 | 35 | texts = ["在这里输入你的文本",] 36 | 37 | wavs = chat.infer(texts, use_decoder=True) 38 | 39 | torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000) 40 | ``` 41 | 42 |

进阶用法

43 | 44 | ```python 45 | ################################### 46 | # Sample a speaker from Gaussian. 47 | 48 | rand_spk = chat.sample_random_speaker() 49 | 50 | params_infer_code = { 51 | 'spk_emb': rand_spk, # add sampled speaker 52 | 'temperature': .3, # using custom temperature 53 | 'top_P': 0.7, # top P decode 54 | 'top_K': 20, # top K decode 55 | } 56 | 57 | ################################### 58 | # For sentence level manual control. 59 | 60 | # use oral_(0-9), laugh_(0-2), break_(0-7) 61 | # to generate special token in text to synthesize. 62 | params_refine_text = { 63 | 'prompt': '[oral_2][laugh_0][break_6]' 64 | } 65 | 66 | wav = chat.infer(texts, params_refine_text=params_refine_text, params_infer_code=params_infer_code) 67 | 68 | ################################### 69 | # For word level manual control. 70 | # use_decoder=False to infer faster with a bit worse quality 71 | text = 'What is [uv_break]your favorite english food?[laugh][lbreak]' 72 | wav = chat.infer(text, skip_refine_text=True, params_infer_code=params_infer_code, use_decoder=False) 73 | 74 | torchaudio.save("output2.wav", torch.from_numpy(wavs[0]), 24000) 75 | ``` 76 | 77 |
78 |

自我介绍样例

 79 | 80 | ```python 81 | inputs_cn = """ 82 | chat T T S 是一款强大的对话式文本转语音模型。它有中英混读和多说话人的能力。 83 | chat T T S 不仅能够生成自然流畅的语音，还能控制[laugh]笑声啊[laugh]， 84 | 停顿啊[uv_break]语气词啊等副语言现象[uv_break]。这个韵律超越了许多开源模型[uv_break]。 85 | 请注意，chat T T S 的使用应遵守法律和伦理准则，避免滥用的安全风险。[uv_break]' 86 | """.replace('\n', '') 87 | 88 | params_refine_text = { 89 | 'prompt': '[oral_2][laugh_0][break_4]' 90 | } 91 | audio_array_cn = chat.infer(inputs_cn, params_refine_text=params_refine_text) 92 | # audio_array_en = chat.infer(inputs_en, params_refine_text=params_refine_text) 93 | 94 | torchaudio.save("output3.wav", torch.from_numpy(audio_array_cn[0]), 24000) 95 | ``` 96 | [男说话人](https://github.com/2noise/ChatTTS/assets/130631963/bbfa3b83-2b67-4bb6-9315-64c992b63788) 97 | 98 | [女说话人](https://github.com/2noise/ChatTTS/assets/130631963/e061f230-0e05-45e6-8e4e-0189f2d260c4) 99 |
100 | 101 | 102 | --- 103 | ## 计划路线 104 | - [x] 开源4w小时基础模型和spk_stats文件 105 | - [ ] 开源VQ encoder和Lora 训练代码 106 | - [ ] 在非refine text情况下, 流式生成音频* 107 | - [ ] 开源多情感可控的4w小时版本 108 | - [ ] ChatTTS.cpp maybe? (欢迎社区PR或独立的新repo) 109 | 110 | --- 111 | ## 常见问题 112 | 113 | ##### 连不上HuggingFace 114 | 请使用[modelscope](https://www.modelscope.cn/models/pzc163/chatTTS)的版本. 并设置cache的位置: 115 | ```python 116 | chat.load_models(source='local', local_path='你的下载位置') 117 | ``` 118 | 119 | ##### 我要多少显存? Infer的速度是怎么样的? 120 | 对于30s的音频, 至少需要4G的显存. 对于4090, 1s生成约7个字所对应的音频. RTF约0.3. 121 | 122 | ##### 模型稳定性似乎不够好, 会出现其他说话人或音质很差的现象. 123 | 这是自回归模型通常都会出现的问题. 说话人可能会在中间变化, 可能会采样到音质非常差的结果, 这通常难以避免. 可以多采样几次来找到合适的结果. 124 | 125 | ##### 除了笑声还能控制什么吗? 还能控制其他情感吗? 126 | 在现在放出的模型版本中, 只有[laugh]和[uv_break], [lbreak]作为字级别的控制单元. 在未来的版本中我们可能会开源其他情感控制的版本. 127 | 128 | --- 129 | ## 致谢 130 | - [bark](https://github.com/suno-ai/bark),[XTTSv2](https://github.com/coqui-ai/TTS)和[valle](https://arxiv.org/abs/2301.02111)展示了自回归任务用于TTS任务的可能性. 131 | - [fish-speech](https://github.com/fishaudio/fish-speech)一个优秀的自回归TTS模型, 揭示了GVQ用于LLM任务的可能性. 132 | - [vocos](https://github.com/gemelo-ai/vocos)作为模型中的vocoder. 133 | 134 | --- 135 | ## 特别致谢 136 | - [wlu-audio lab](https://audio.westlake.edu.cn/)为我们提供了早期算法试验的支持. 137 | -------------------------------------------------------------------------------- /api.py: -------------------------------------------------------------------------------- 1 | import os 2 | import re 3 | import sys 4 | 5 | import numpy as np 6 | import torch 7 | from fastapi import FastAPI, Response 8 | from loguru import logger 9 | from pydantic import BaseModel, Field 10 | from pydub import AudioSegment 11 | 12 | import ChatTTS 13 | 14 | LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO") 15 | LOG_FILE = os.getenv("LOG_FILE", "service.log") 16 | valid_pattern = re.compile(r"[^\u4e00-\u9fffA-Za-z，。、,\. \[\]\_]") 17 | 18 | 19 | def logger_setting(logfile: str, log_level: str = "INFO"): 20 | logger.remove() 21 | logger.add(sys.stderr, level=log_level) 22 | logger.add(logfile, level=log_level, rotation="10MB", retention=20) 23 | 24 | 25 | logger_setting(LOG_FILE, LOG_LEVEL) 26 | 27 | 28 | def deterministic(seed=0): 29 | torch.manual_seed(seed) 30 | np.random.seed(seed) 31 | torch.cuda.manual_seed(seed) 32 | torch.backends.cudnn.deterministic = True 33 | torch.backends.cudnn.benchmark = False 34 | 35 | 36 | app = FastAPI(title="ChatTTS Api") 37 | 38 | logger.info("Loading models...") 39 | chat = ChatTTS.Chat() 40 | chat.load_models(source="local", local_path="tts_model", compile=False) 41 | # compile=True is faster, but it will take longer to start 42 | logger.info("Models loaded, warming up...") 43 | chat.infer(["你好"], use_decoder=True) 44 | logger.info("Warmup done, ready to serve requests!") 45 | 46 | 47 | class TTSRequest(BaseModel): 48 | text: str = Field(..., description="Input text to be converted to speech") 49 | seed: int = Field(1111, description="Voice Seed, int") 50 | top_P: float = Field(0.7, description="Top P value for sampling") 51 | top_K: int = Field(20, description="Top K value for sampling") 52 | temperature: float = Field(0.3, description="Temperature value for sampling") 53 | skip_refine_text: bool = Field(False, description="Whether to refine the text") 54 | refine_text_prompt: str = Field( 55 | "[oral_2][laugh_0][break_6]", description="Prompt for refining the text" 56 | ) 57 | 58 | 59 | def wav_to_mp3(wav_data, sample_rate=24000, bitrate="48k"): 60 | audio = AudioSegment( 61 | data=wav_data, sample_width=2, frame_rate=sample_rate, channels=1 62 | ) 63 | return audio.export(format="mp3", bitrate=bitrate) 64 | 65 | 66 | def infer(request): 67 | # remove invalid characters, otherwise the model will raise an error 68 | # text = valid_pattern.sub("", request.text) 69 | text = request.text 70 | logger.info(f"Text after removing invalid characters: {text}") 71 | deterministic(request.seed) 72 | rnd_spk_emb = chat.sample_random_speaker() 73 | 74 | params_infer_code = { 75 | "spk_emb": rnd_spk_emb, 76 | "temperature": request.temperature, 77 | "top_P": request.top_P, 78 | "top_K": request.top_K, 79 | } 80 | 81 | params_refine_text = {} 82 | if request.skip_refine_text: 83 | params_refine_text = {"prompt": request.refine_text_prompt} 84 | 85 | wavs = chat.infer( 86 | [text], 87 | params_refine_text=params_refine_text, 88 | params_infer_code=params_infer_code, 89 | use_decoder=True, 90 | skip_refine_text=request.skip_refine_text, 91 | ) 92 | audio_data = wavs[0][0] 93 | audio_data = audio_data / np.max(np.abs(audio_data)) 94 | audio_data = (audio_data * 32768).astype(np.int16) 95 | return wav_to_mp3(audio_data).read() 96 | 97 | 98 | @app.post("/tts") 99 | async def tts_stream( 100 | request: TTSRequest, 101 | ) -> Response: 102 | logger.info(f"Received request: {request}") 103 | 104 | audio = infer(request) 105 | logger.info("Generated audio") 106 | return Response(audio, media_type="audio/mp3") 107 | -------------------------------------------------------------------------------- /client.py: -------------------------------------------------------------------------------- 1 | import requests 2 | 3 | 4 | def tts( 5 | server_url: str, 6 | text, 7 | seed, 8 | output_file, 9 | timeout: int = 20, 10 | ): 11 | data = { 12 | "text": text, 13 | "seed": seed, 14 | "temperature": 0.3, 15 | "top_P": 0.7, 16 | "top_K": 20, 17 | "skip_refine_text": False, 18 | } 19 | try: 20 | response = requests.post( 21 | server_url + "/tts", 22 | stream=True, 23 | json=data, 24 | timeout=timeout, 25 | ) 26 | response.raise_for_status() 27 | with open(output_file, "wb") as f: 28 | f.write(response.content) 29 | 30 | except requests.exceptions.RequestException as e: 31 | raise ValueError(response.content) from e 32 | 33 | 34 | if __name__ == "__main__": 35 | tts( 36 | "http://127.0.0.1:8080", 37 | "你好世界", 38 | 1111, 39 | "output.mp3", 40 | ) 41 | print("See output.mp3") 42 | -------------------------------------------------------------------------------- /requirements-api-ui.txt: -------------------------------------------------------------------------------- 1 | fastapi 2 | soundfile 3 | loguru 4 | streamlit 5 | gunicorn 6 | pydub -------------------------------------------------------------------------------- /requirements-chat-tts.txt: -------------------------------------------------------------------------------- 1 | omegaconf~=2.3.0 2 | torch 3 | torchaudio 4 | tqdm 5 | einops 6 | vector_quantize_pytorch 7 | transformers~=4.41.1 8 | vocos 9 | WeTextProcessing 10 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | omegaconf~=2.3.0 2 | torch~=2.1.0 3 | tqdm 4 | einops 5 | vector_quantize_pytorch 6 | transformers~=4.41.1 7 | vocos 8 | IPython -------------------------------------------------------------------------------- /webui.py: -------------------------------------------------------------------------------- 1 | import os 2 | import random 3 | import argparse 4 | 5 | import torch 6 | import gradio as gr 7 | import numpy as np 8 | 9 | import ChatTTS 10 | 11 | 12 | def generate_seed(): 13 | new_seed = random.randint(1, 100000000) 14 | return { 15 | "__type__": "update", 16 | "value": new_seed 17 | } 18 | 19 | 20 | def generate_audio(text, temperature, top_P, top_K, audio_seed_input, text_seed_input, refine_text_flag): 21 | 22 | torch.manual_seed(audio_seed_input) 23 | rand_spk = chat.sample_random_speaker() 24 | params_infer_code = { 25 | 'spk_emb': rand_spk, 26 | 'temperature': temperature, 27 | 'top_P': top_P, 28 | 'top_K': top_K, 29 | } 30 | params_refine_text = {'prompt': '[oral_2][laugh_0][break_6]'} 31 | 32 | torch.manual_seed(text_seed_input) 33 | 34 | if refine_text_flag: 35 | text = chat.infer(text, 36 | skip_refine_text=False, 37 | refine_text_only=True, 38 | params_refine_text=params_refine_text, 39 | params_infer_code=params_infer_code 40 | ) 41 | 42 | wav = chat.infer(text, 43 | skip_refine_text=True, 44 | params_refine_text=params_refine_text, 45 | params_infer_code=params_infer_code 46 | ) 47 | 48 | audio_data = np.array(wav[0]).flatten() 49 | sample_rate = 24000 50 | text_data = text[0] if isinstance(text, list) else text 51 | 52 | return [(sample_rate, audio_data), text_data] 53 | 54 | 55 | def main(): 56 | 57 | with gr.Blocks() as demo: 58 | gr.Markdown("# ChatTTS Webui") 59 | gr.Markdown("ChatTTS Model: [2noise/ChatTTS](https://github.com/2noise/ChatTTS)") 60 | 61 | default_text = "四川美食确实以辣闻名，但也有不辣的选择。比如甜水面、赖汤圆、蛋烘糕、叶儿粑等，这些小吃口味温和，甜而不腻，也很受欢迎。" 62 | text_input = gr.Textbox(label="Input Text", lines=4, placeholder="Please Input Text...", value=default_text) 63 | 64 | with gr.Row(): 65 | refine_text_checkbox = gr.Checkbox(label="Refine text", value=True) 66 | temperature_slider = gr.Slider(minimum=0.00001, maximum=1.0, step=0.00001, value=0.3, label="Audio temperature") 67 | top_p_slider = gr.Slider(minimum=0.1, maximum=0.9, step=0.05, value=0.7, label="top_P") 68 | top_k_slider = gr.Slider(minimum=1, maximum=20, step=1, value=20, label="top_K") 69 | 70 | with gr.Row(): 71 | audio_seed_input = gr.Number(value=2, label="Audio Seed") 72 | generate_audio_seed = gr.Button("\U0001F3B2") 73 | text_seed_input = gr.Number(value=42, label="Text Seed") 74 | generate_text_seed = gr.Button("\U0001F3B2") 75 | 76 | generate_button = gr.Button("Generate") 77 | 78 | text_output = gr.Textbox(label="Output Text", interactive=False) 79 | audio_output = gr.Audio(label="Output Audio") 80 | 81 | generate_audio_seed.click(generate_seed, 82 | inputs=[], 83 | outputs=audio_seed_input) 84 | 85 | generate_text_seed.click(generate_seed, 86 | inputs=[], 87 | outputs=text_seed_input) 88 | 89 | generate_button.click(generate_audio, 90 | inputs=[text_input, temperature_slider, top_p_slider, top_k_slider, audio_seed_input, text_seed_input, refine_text_checkbox], 91 | outputs=[audio_output, text_output]) 92 | 93 | parser = argparse.ArgumentParser(description='ChatTTS demo Launch') 94 | parser.add_argument('--server_name', type=str, default='0.0.0.0', help='Server name') 95 | parser.add_argument('--server_port', type=int, default=8080, help='Server port') 96 | parser.add_argument('--local_path', type=str, default=None, help='the local_path if need') 97 | args = parser.parse_args() 98 | 99 | print("loading ChatTTS model...") 100 | global chat 101 | chat = ChatTTS.Chat() 102 | 103 | if args.local_path == None: 104 | chat.load_models() 105 | else: 106 | print('local model path:', args.local_path) 107 | chat.load_models('local', local_path=args.local_path) 108 | 109 | demo.launch(server_name=args.server_name, server_port=args.server_port, inbrowser=True) 110 | 111 | 112 | if __name__ == '__main__': 113 | main() -------------------------------------------------------------------------------- /webui_api.py: -------------------------------------------------------------------------------- 1 | import requests 2 | import streamlit as st 3 | 4 | # 音色选项 5 | voices = { 6 | "音色1": {"seed": 1111}, 7 | "音色2": {"seed": 2222}, 8 | "音色3": {"seed": 3333}, 9 | "音色4": {"seed": 4444}, 10 | "音色5": {"seed": 5555}, 11 | "音色6": {"seed": 6666}, 12 | "音色7": {"seed": 7777}, 13 | "音色8": {"seed": 8888}, 14 | "音色9": {"seed": 9999}, 15 | } 16 | 17 | 18 | def tts( 19 | server_url: str, 20 | data, 21 | output_file, 22 | ): 23 | try: 24 | response = requests.post( 25 | server_url + "/tts", 26 | stream=True, 27 | json=data, 28 | timeout=20, 29 | ) 30 | response.raise_for_status() 31 | with open(output_file, "wb") as f: 32 | f.write(response.content) 33 | 34 | except requests.exceptions.RequestException as e: 35 | raise ValueError(response.content) from e 36 | 37 | 38 | # Streamlit界面设置 39 | st.title("ChatTTS 音频生成器") 40 | 41 | st.header("选择音色") 42 | st.markdown("通过随机种子生成,不保证每次运行结果一致") 43 | 44 | voice_options = list(voices.keys()) 45 | selected_voice = st.selectbox("选择音色", voice_options) 46 | seed = voices[selected_voice]["seed"] 47 | 48 | st.header("输入文本") 49 | text = st.text_area("请输入要合成的文本") 50 | 51 | st.header("参数设置") 52 | temperature = st.slider("Temperature", 0.0, 1.0, 0.3) 53 | top_P = st.slider("Top P", 0.0, 1.0, 0.7) 54 | top_K = st.slider("Top K", 0, 50, 20) 55 | skip_refine_text = st.checkbox("跳过文本调整") 56 | 57 | 58 | # 音频生成 59 | if st.button("生成音频"): 60 | # 调用ChatTTS进行推断 61 | with st.spinner("正在生成音频..."): 62 | tmp_file = "temp.mp3" 63 | data = { 64 | "text": text, 65 | "seed": seed, 66 | "temperature": temperature, 67 | "top_P": top_P, 68 | "top_K": top_K, 69 | "skip_refine_text": skip_refine_text, 70 | } 71 | tts("http://127.0.0.1:8080", data, tmp_file) 72 | 73 | st.audio(tmp_file) 74 | --------------------------------------------------------------------------------