├── .github └── workflows │ └── python-publish.yml ├── .gitignore ├── .pre-commit-config.yaml ├── LICENSE ├── MANIFEST.in ├── README.md ├── requirements.txt ├── s3tokenizer ├── __init__.py ├── assets │ ├── BAC009S0764W0121.wav │ ├── BAC009S0764W0122.wav │ └── mel_filters.npz ├── cli.py ├── model.py ├── model_v2.py └── utils.py ├── setup.py └── test └── test_onnx.py /.github/workflows/python-publish.yml: -------------------------------------------------------------------------------- 1 | name: Release 2 | 3 | on: 4 | push: 5 | branches: 6 | - main 7 | jobs: 8 | deploy: 9 | runs-on: ubuntu-latest 10 | steps: 11 | - uses: actions/checkout@v3 12 | - uses: actions-ecosystem/action-regex-match@v2 13 | id: regex-match 14 | with: 15 | text: ${{ github.event.head_commit.message }} 16 | regex: '^Release ([^ ]+)' 17 | - name: Set up Python 18 | uses: actions/setup-python@v4 19 | with: 20 | python-version: '3.8' 21 | - name: Install dependencies 22 | run: | 23 | python -m pip install --upgrade pip 24 | pip install setuptools wheel twine 25 | - name: Release 26 | if: ${{ steps.regex-match.outputs.match != '' }} 27 | uses: softprops/action-gh-release@v1 28 | with: 29 | tag_name: v${{ steps.regex-match.outputs.group1 }} 30 | - name: Build and publish 31 | if: ${{ steps.regex-match.outputs.match != '' }} 32 | env: 33 | TWINE_USERNAME: __token__ 34 | TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }} 35 | run: | 36 | python setup.py sdist 37 | twine upload dist/* 38 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # poetry 98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 102 | #poetry.lock 103 | 104 | # pdm 105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 106 | #pdm.lock 107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 108 | # in version control. 109 | # https://pdm.fming.dev/latest/usage/project/#working-with-version-control 110 | .pdm.toml 111 | .pdm-python 112 | .pdm-build/ 113 | 114 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 115 | __pypackages__/ 116 | 117 | # Celery stuff 118 | celerybeat-schedule 119 | celerybeat.pid 120 | 121 | # SageMath parsed files 122 | *.sage.py 123 | 124 | # Environments 125 | .env 126 | .venv 127 | env/ 128 | venv/ 129 | ENV/ 130 | env.bak/ 131 | venv.bak/ 132 | 133 | # Spyder project settings 134 | .spyderproject 135 | .spyproject 136 | 137 | # Rope project settings 138 | .ropeproject 139 | 140 | # mkdocs documentation 141 | /site 142 | 143 | # mypy 144 | .mypy_cache/ 145 | .dmypy.json 146 | dmypy.json 147 | 148 | # Pyre type checker 149 | .pyre/ 150 | 151 | # pytype static type analyzer 152 | .pytype/ 153 | 154 | # Cython debug symbols 155 | cython_debug/ 156 | 157 | # PyCharm 158 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 159 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 160 | # and can be added to the global gitignore or merged into this file. For a more nuclear 161 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 162 | #.idea/ 163 | -------------------------------------------------------------------------------- /.pre-commit-config.yaml: -------------------------------------------------------------------------------- 1 | repos: 2 | - repo: https://github.com/pre-commit/pre-commit-hooks 3 | rev: v4.5.0 4 | hooks: 5 | - id: trailing-whitespace 6 | exclude: 's3tokenizer/assets/.*' 7 | - repo: https://github.com/pre-commit/mirrors-yapf 8 | rev: 'v0.32.0' 9 | hooks: 10 | - id: yapf 11 | - repo: https://github.com/pycqa/flake8 12 | rev: '3.8.2' 13 | hooks: 14 | - id: flake8 15 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /MANIFEST.in: -------------------------------------------------------------------------------- 1 | include requirements.txt 2 | include README.md 3 | include LICENSE 4 | include s3tokenizer/assets/* -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Reverse Engineering of S3Tokenizer 2 | 3 |
4 | Description 5 |

Supervised Semantic Speech Tokenizer (S3Tokenizer)

6 |
7 | 8 | S3Tokenizer was initially introduced in CosyVoice [[Paper]](https://arxiv.org/abs/2407.04051v2) [[Repo]](https://github.com/FunAudioLLM/CosyVoice), it is a Supervised Semantic Speech Tokenizer based on the pre-trained SenseVoice-Large model, which enhances the semantic relationship of extracted tokens to textual and paralinguistic information, is robust to data noise, and reduces the reliance on clean data collection, thereby enabling the use of a broader range of data for model training. 9 | 10 | However, as indicated in this [[issue]](https://github.com/FunAudioLLM/CosyVoice/issues/70), the authors have no intention to open-source the PyTorch implementation of the S3Tokenizer, and only plan to release an ONNX file. Additionally, users aiming to fine-tune CosyVoice must extract speech codes offline, with the batch size restricted to 1, a process that is notably time-consuming (refer to [[cosyvoice/tools/extract_speech_token.py]](https://github.com/FunAudioLLM/CosyVoice/blob/main/tools/extract_speech_token.py)). 11 | 12 | This repository undertakes a reverse engineering of the S3Tokenizer, offering: 13 | 1. A pure PyTorch implementation of S3Tokenizer (see [[model.py]](https://github.com/xingchensong/S3Tokenizer/blob/main/s3tokenizer/model.py)), compatible with initializing weights from the released ONNX file (see [[utils.py::onnx2torch()]](https://github.com/xingchensong/S3Tokenizer/blob/main/s3tokenizer/utils.py)). 14 | 2. High-throughput (distributed) batch inference, achieving a ~790x speedup compared to the original inference pipeline in [[cosyvoice/tools/extract_speech_token.py]](https://github.com/FunAudioLLM/CosyVoice/blob/main/tools/extract_speech_token.py). 15 | 3. The capability to perform online speech code extraction during SpeechLLM training. 16 | 17 | ## Supported Models 🔥 18 | - [x] [S3Tokenizer V1 50hz](https://modelscope.cn/models/iic/CosyVoice-300M) 19 | - [x] [S3Tokenizer V1 25hz](https://modelscope.cn/models/iic/CosyVoice-300M-25Hz) 20 | - [x] [S3Tokenizer V2 25hz](https://modelscope.cn/models/iic/CosyVoice2-0.5B) 21 | 22 | 23 | # Setup 24 | 25 | ```sh 26 | pip install s3tokenizer 27 | ``` 28 | 29 | # Usage-1: Offline batch inference 30 | 31 | ```py 32 | import s3tokenizer 33 | 34 | tokenizer = s3tokenizer.load_model("speech_tokenizer_v1").cuda() # or "speech_tokenizer_v1_25hz speech_tokenizer_v2_25hz" 35 | 36 | mels = [] 37 | wav_paths = ["s3tokenizer/assets/BAC009S0764W0121.wav", "s3tokenizer/assets/BAC009S0764W0122.wav"] 38 | for wav_path in wav_paths: 39 | audio = s3tokenizer.load_audio(wav_path) 40 | mels.append(s3tokenizer.log_mel_spectrogram(audio)) 41 | mels, mels_lens = s3tokenizer.padding(mels) 42 | codes, codes_lens = tokenizer.quantize(mels.cuda(), mels_lens.cuda()) 43 | 44 | for i in range(len(wav_paths)): 45 | print(codes[i, :codes_lens[i].item()]) 46 | ``` 47 | 48 | # Usage-2: Distributed offline batch inference via command-line tools 49 | 50 | ## 2.1 CPU batch inference 51 | 52 | ```sh 53 | s3tokenizer --wav_scp xxx.scp \ 54 | --device "cpu" \ 55 | --output_dir "./" \ 56 | --batch_size 32 \ 57 | --model "speech_tokenizer_v1" # or "speech_tokenizer_v1_25hz speech_tokenizer_v2_25hz" 58 | ``` 59 | 60 | 61 | 62 | https://github.com/user-attachments/assets/d37d10fd-0e13-46a3-86b0-4cbec309086f 63 | 64 | 65 | 66 | ## 2.2 (Multi) GPU batch inference (a.k.a Distributed inference) 67 | 68 | ```sh 69 | torchrun --nproc_per_node=8 --nnodes=1 \ 70 | --rdzv_id=2024 --rdzv_backend="c10d" --rdzv_endpoint="localhost:0" \ 71 | `which s3tokenizer` --wav_scp xxx.scp \ 72 | --device "cuda" \ 73 | --output_dir "./" \ 74 | --batch_size 32 \ 75 | --model "speech_tokenizer_v1" # or "speech_tokenizer_v1_25hz speech_tokenizer_v2_25hz" 76 | ``` 77 | 78 | 79 | 80 | https://github.com/user-attachments/assets/79a3fb11-7199-4ee2-8a35-9682a3b4d94a 81 | 82 | 83 | 84 | ## 2.3 Performance Benchmark 85 | 86 | | Method | Time cost on Aishell Test Set | Relative speed up | Miss Rate | 87 | |:------:|:----------:|:--------------:|:-----:| 88 | | [[cosyvoice/tools/extract_speech_token.py]](https://github.com/FunAudioLLM/CosyVoice/blob/main/tools/extract_speech_token.py), cpu | 9 hours | ~ | ~ | 89 | | cpu, batchsize 32 | 1.5h | ~6x | 0.00% | 90 | | 4 gpus (3090), batchsize 32 per gpu | 41s | ~790x | 0.00% | 91 | 92 | The miss rate represents the proportion of tokens that are inconsistent between the batch inference predictions and the ONNX (batch=1) inference predictions. 93 | 94 | # Usage-3: Online speech code extraction 95 | 96 | 97 | 98 | 99 | 100 | 101 | 102 | 139 | 140 |
Before (extract code offline)After (extract code online)
103 | 104 | 105 | ```py 106 | 107 | class SpeechLLM(nn.Module): 108 | ... 109 | def __init__(self, ...): 110 | ... 111 | 112 | def forward(self, speech_codes: Tensor, text_ids: Tensor, ...): 113 | ... 114 | ``` 115 | 116 | 117 | 118 | 119 | 120 | ```py 121 | import s3tokenizer 122 | 123 | class SpeechLLM(nn.Module): 124 | ... 125 | def __init__(self, ...): 126 | ... 127 | self.speech_tokenizer = s3tokenizer.load_model("speech_tokenizer_v1") # or "speech_tokenizer_v1_25hz" 128 | self.speech_tokenizer.freeze() 129 | 130 | def forward(self, speech: Tensor, speech_lens: Tensor, text_ids: Tensor, ...): 131 | ... 132 | speech_codes, speech_codes_lens = self.speech_tokenizer.quantize(speech, speech_lens) 133 | speech_codes = speech_codes.clone() # for backward compatbility 134 | speech_codes_lens = speeech_codes_lens.clone() # for backward compatbility 135 | ``` 136 | 137 | 138 |
141 | 142 | 143 | # TODO 144 | 145 | - [x] Usage-1: Offline batch inference 146 | - [x] Usage-2: Distributed offline batch inference via command-line tools 147 | - [x] Usage-3: Online speech code extraction 148 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | pre-commit 2 | numpy 3 | torch 4 | onnx 5 | tqdm 6 | torchaudio 7 | einops 8 | -------------------------------------------------------------------------------- /s3tokenizer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2023 OpenAI. (authors: Whisper Team) 2 | # 2024 Tsinghua Univ. (authors: Xingchen Song) 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Modified from 16 | https://github.com/openai/whisper/blob/main/whisper/__init__.py 17 | """ 18 | 19 | import hashlib 20 | import os 21 | import urllib 22 | import warnings 23 | from typing import List, Union 24 | 25 | from tqdm import tqdm 26 | 27 | from s3tokenizer.model_v2 import S3TokenizerV2 28 | 29 | from .model import S3Tokenizer 30 | from .utils import (load_audio, log_mel_spectrogram, make_non_pad_mask, 31 | mask_to_bias, onnx2torch, padding) 32 | 33 | __all__ = [ 34 | 'load_audio', 'log_mel_spectrogram', 'make_non_pad_mask', 'mask_to_bias', 35 | 'onnx2torch', 'padding' 36 | ] 37 | _MODELS = { 38 | "speech_tokenizer_v1": 39 | "https://www.modelscope.cn/models/iic/cosyvoice-300m/" 40 | "resolve/master/speech_tokenizer_v1.onnx", 41 | "speech_tokenizer_v1_25hz": 42 | "https://www.modelscope.cn/models/iic/CosyVoice-300M-25Hz/" 43 | "resolve/master/speech_tokenizer_v1.onnx", 44 | "speech_tokenizer_v2_25hz": 45 | "https://www.modelscope.cn/models/iic/CosyVoice2-0.5B/" 46 | "resolve/master/speech_tokenizer_v2.onnx", 47 | } 48 | 49 | _SHA256S = { 50 | "speech_tokenizer_v1": 51 | "23b5a723ed9143aebfd9ffda14ac4c21231f31c35ef837b6a13bb9e5488abb1e", 52 | "speech_tokenizer_v1_25hz": 53 | "56285ddd4a83e883ee0cb9f8d69c1089b53a94b1f78ff7e4a0224a27eb4cb486", 54 | "speech_tokenizer_v2_25hz": 55 | "d43342aa12163a80bf07bffb94c9de2e120a8df2f9917cd2f642e7f4219c6f71", 56 | } 57 | 58 | 59 | def _download(name: str, root: str) -> Union[bytes, str]: 60 | os.makedirs(root, exist_ok=True) 61 | 62 | expected_sha256 = _SHA256S[name] 63 | url = _MODELS[name] 64 | download_target = os.path.join(root, f"{name}.onnx") 65 | 66 | if os.path.exists(download_target) and not os.path.isfile(download_target): 67 | raise RuntimeError( 68 | f"{download_target} exists and is not a regular file") 69 | 70 | if os.path.isfile(download_target): 71 | with open(download_target, "rb") as f: 72 | model_bytes = f.read() 73 | if hashlib.sha256(model_bytes).hexdigest() == expected_sha256: 74 | return download_target 75 | else: 76 | warnings.warn( 77 | f"{download_target} exists, but the SHA256 checksum does not" 78 | " match; re-downloading the file") 79 | 80 | with urllib.request.urlopen(url) as source, open(download_target, 81 | "wb") as output: 82 | with tqdm( 83 | total=int(source.info().get("Content-Length")), 84 | ncols=80, 85 | unit="iB", 86 | unit_scale=True, 87 | unit_divisor=1024, 88 | desc="Downloading onnx checkpoint", 89 | ) as loop: 90 | while True: 91 | buffer = source.read(8192) 92 | if not buffer: 93 | break 94 | 95 | output.write(buffer) 96 | loop.update(len(buffer)) 97 | 98 | model_bytes = open(download_target, "rb").read() 99 | if hashlib.sha256(model_bytes).hexdigest() != expected_sha256: 100 | raise RuntimeError( 101 | "Model has been downloaded but the SHA256 checksum does not not" 102 | " match. Please retry loading the model.") 103 | 104 | return download_target 105 | 106 | 107 | def available_models() -> List[str]: 108 | """Returns the names of available models""" 109 | return list(_MODELS.keys()) 110 | 111 | 112 | def load_model( 113 | name: str, 114 | download_root: str = None, 115 | ) -> S3Tokenizer: 116 | """ 117 | Load a S3Tokenizer ASR model 118 | 119 | Parameters 120 | ---------- 121 | name : str 122 | one of the official model names listed by 123 | `s3tokenizer.available_models()`, or path to a model checkpoint 124 | containing the model dimensions and the model state_dict. 125 | download_root: str 126 | path to download the model files; by default, 127 | it uses "~/.cache/s3tokenizer" 128 | 129 | Returns 130 | ------- 131 | model : S3Tokenizer 132 | The S3Tokenizer model instance 133 | """ 134 | 135 | if download_root is None: 136 | default = os.path.join(os.path.expanduser("~"), ".cache") 137 | download_root = os.path.join(os.getenv("XDG_CACHE_HOME", default), 138 | "s3tokenizer") 139 | 140 | if name in _MODELS: 141 | checkpoint_file = _download(name, download_root) 142 | elif os.path.isfile(name): 143 | checkpoint_file = name 144 | else: 145 | raise RuntimeError( 146 | f"Model {name} not found; available models = {available_models()}") 147 | if 'v2' in name: 148 | model = S3TokenizerV2(name) 149 | else: 150 | model = S3Tokenizer(name) 151 | model.init_from_onnx(checkpoint_file) 152 | 153 | return model 154 | -------------------------------------------------------------------------------- /s3tokenizer/assets/BAC009S0764W0121.wav: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xingchensong/S3Tokenizer/dc95bac8bce0dee347c40acca90b0005d8eba711/s3tokenizer/assets/BAC009S0764W0121.wav -------------------------------------------------------------------------------- /s3tokenizer/assets/BAC009S0764W0122.wav: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xingchensong/S3Tokenizer/dc95bac8bce0dee347c40acca90b0005d8eba711/s3tokenizer/assets/BAC009S0764W0122.wav -------------------------------------------------------------------------------- /s3tokenizer/assets/mel_filters.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xingchensong/S3Tokenizer/dc95bac8bce0dee347c40acca90b0005d8eba711/s3tokenizer/assets/mel_filters.npz -------------------------------------------------------------------------------- /s3tokenizer/cli.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2024 Tsinghua Univ. (authors: Xingchen Song) 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ Example Usage 15 | cpu: 16 | 17 | s3tokenizer --wav_scp xxx.scp \ 18 | --device "cpu" \ 19 | --output_dir "./" \ 20 | --batch_size 32 21 | 22 | gpu: 23 | 24 | torchrun --nproc_per_node=8 --nnodes=1 \ 25 | --rdzv_id=2024 --rdzv_backend="c10d" --rdzv_endpoint="localhost:0" \ 26 | `which s3tokenizer` --wav_scp xxx.scp \ 27 | --device "cuda" \ 28 | --output_dir "./" \ 29 | --batch_size 32 30 | 31 | """ 32 | 33 | import argparse 34 | import json 35 | import os 36 | 37 | import torch 38 | import torch.distributed as dist 39 | from torch.utils.data import DataLoader, Dataset, DistributedSampler 40 | from tqdm import tqdm 41 | 42 | import s3tokenizer 43 | 44 | 45 | class AudioDataset(Dataset): 46 | 47 | def __init__(self, wav_scp): 48 | self.data = [] 49 | self.keys = [] 50 | 51 | with open(wav_scp, 'r', encoding='utf-8') as f: 52 | for line in f: 53 | key, file_path = line.strip().split() 54 | self.data.append(file_path) 55 | self.keys.append(key) 56 | 57 | def __len__(self): 58 | return len(self.data) 59 | 60 | def __getitem__(self, idx): 61 | file_path = self.data[idx] 62 | key = self.keys[idx] 63 | audio = s3tokenizer.load_audio(file_path) 64 | if audio.shape[0] / 16000 > 30: 65 | print( 66 | f'do not support extract speech token for audio longer than 30s, file_path: {file_path}' # noqa 67 | ) 68 | mel = torch.zeros(128, 0) 69 | else: 70 | mel = s3tokenizer.log_mel_spectrogram(audio) 71 | return key, mel 72 | 73 | 74 | def collate_fn(batch): 75 | keys = [item[0] for item in batch] 76 | mels = [item[1] for item in batch] 77 | mels, mels_lens = s3tokenizer.padding(mels) 78 | return keys, mels, mels_lens 79 | 80 | 81 | def init_distributed(): 82 | world_size = int(os.environ.get('WORLD_SIZE', 1)) 83 | local_rank = int(os.environ.get('LOCAL_RANK', 0)) 84 | rank = int(os.environ.get('RANK', 0)) 85 | print('Inference on multiple gpus, this gpu {}'.format(local_rank) + 86 | ', rank {}, world_size {}'.format(rank, world_size)) 87 | torch.cuda.set_device(local_rank) 88 | dist.init_process_group("nccl") 89 | return world_size, local_rank, rank 90 | 91 | 92 | def get_args(): 93 | parser = argparse.ArgumentParser(description='extract speech code') 94 | parser.add_argument('--model', 95 | required=True, 96 | type=str, 97 | choices=[ 98 | "speech_tokenizer_v1", "speech_tokenizer_v1_25hz", 99 | "speech_tokenizer_v2_25hz" 100 | ], 101 | help='model version') 102 | parser.add_argument('--wav_scp', 103 | required=True, 104 | type=str, 105 | help='each line contains `wav_name wav_path`') 106 | parser.add_argument('--device', 107 | required=True, 108 | type=str, 109 | choices=["cuda", "cpu"], 110 | help='device for inference') 111 | parser.add_argument('--output_dir', 112 | required=True, 113 | type=str, 114 | help='dir to save result') 115 | parser.add_argument('--batch_size', 116 | required=True, 117 | type=int, 118 | help='batch size (per-device) for inference') 119 | parser.add_argument('--num_workers', 120 | type=int, 121 | default=4, 122 | help='workers for dataloader') 123 | parser.add_argument('--prefetch', 124 | type=int, 125 | default=5, 126 | help='prefetch for dataloader') 127 | args = parser.parse_args() 128 | return args 129 | 130 | 131 | def main(): 132 | args = get_args() 133 | os.makedirs(args.output_dir, exist_ok=True) 134 | 135 | if args.device == "cuda": 136 | assert (torch.cuda.is_available()) 137 | world_size, local_rank, rank = init_distributed() 138 | else: 139 | world_size, local_rank, rank = 1, 0, 0 140 | 141 | device = torch.device(args.device) 142 | model = s3tokenizer.load_model(args.model).to(device) 143 | dataset = AudioDataset(args.wav_scp) 144 | 145 | if args.device == "cuda": 146 | model = torch.nn.parallel.DistributedDataParallel( 147 | model, device_ids=[local_rank]) 148 | sampler = DistributedSampler(dataset, 149 | num_replicas=world_size, 150 | rank=rank) 151 | else: 152 | sampler = None 153 | 154 | dataloader = DataLoader(dataset, 155 | batch_size=args.batch_size, 156 | sampler=sampler, 157 | shuffle=False, 158 | num_workers=args.num_workers, 159 | prefetch_factor=args.prefetch, 160 | collate_fn=collate_fn) 161 | 162 | total_steps = len(dataset) 163 | 164 | if rank == 0: 165 | progress_bar = tqdm(total=total_steps, desc="Processing", unit="wavs") 166 | 167 | writer = open(f"{args.output_dir}/part_{rank + 1}_of_{world_size}", "w") 168 | for keys, mels, mels_lens in dataloader: 169 | codes, codes_lens = model(mels.to(device), mels_lens.to(device)) 170 | for i, k in enumerate(keys): 171 | code = codes[i, :codes_lens[i].item()].tolist() 172 | writer.write( 173 | json.dumps({ 174 | "key": k, 175 | "code": code 176 | }, ensure_ascii=False) + "\n") 177 | if rank == 0: 178 | progress_bar.update(world_size * len(keys)) 179 | 180 | if rank == 0: 181 | progress_bar.close() 182 | writer.close() 183 | if args.device == "cuda": 184 | dist.barrier() 185 | dist.destroy_process_group() 186 | 187 | 188 | if __name__ == "__main__": 189 | main() 190 | -------------------------------------------------------------------------------- /s3tokenizer/model.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2023 OpenAI. (authors: Whisper Team) 2 | # 2024 Tsinghua Univ. (authors: Xingchen Song) 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Modified from https://github.com/openai/whisper/blob/main/whisper/model.py 16 | Add EuclideanCodebook & VectorQuantization 17 | """ 18 | 19 | from dataclasses import dataclass 20 | from typing import Iterable, Optional, Tuple 21 | 22 | import numpy as np 23 | import torch 24 | import torch.nn.functional as F 25 | from einops import rearrange 26 | from torch import Tensor, nn 27 | 28 | from .utils import make_non_pad_mask, mask_to_bias, onnx2torch 29 | 30 | 31 | @dataclass 32 | class ModelConfig: 33 | n_mels: int = 128 34 | n_audio_ctx: int = 1500 35 | n_audio_state: int = 1280 36 | n_audio_head: int = 20 37 | n_audio_layer: int = 6 38 | n_codebook_size: int = 4096 39 | 40 | use_sdpa: bool = False 41 | 42 | 43 | class LayerNorm(nn.LayerNorm): 44 | 45 | def forward(self, x: Tensor) -> Tensor: 46 | return super().forward(x.float()).type(x.dtype) 47 | 48 | 49 | class Linear(nn.Linear): 50 | 51 | def forward(self, x: Tensor) -> Tensor: 52 | return F.linear( 53 | x, 54 | self.weight.to(x.dtype), 55 | None if self.bias is None else self.bias.to(x.dtype), 56 | ) 57 | 58 | 59 | class Conv1d(nn.Conv1d): 60 | 61 | def _conv_forward(self, x: Tensor, weight: Tensor, 62 | bias: Optional[Tensor]) -> Tensor: 63 | return super()._conv_forward( 64 | x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)) 65 | 66 | 67 | def sinusoids(length, channels, max_timescale=10000): 68 | """Returns sinusoids for positional embedding""" 69 | assert channels % 2 == 0 70 | log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1) 71 | inv_timescales = torch.exp(-log_timescale_increment * 72 | torch.arange(channels // 2)) 73 | scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[ 74 | np.newaxis, :] 75 | return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1) 76 | 77 | 78 | class MultiHeadAttention(nn.Module): 79 | 80 | def __init__(self, n_state: int, n_head: int, use_sdpa: bool = False): 81 | super().__init__() 82 | self.n_head = n_head 83 | self.query = Linear(n_state, n_state) 84 | self.key = Linear(n_state, n_state, bias=False) 85 | self.value = Linear(n_state, n_state) 86 | self.out = Linear(n_state, n_state) 87 | 88 | self.use_sdpa = use_sdpa 89 | 90 | def forward( 91 | self, 92 | x: Tensor, 93 | mask: Optional[Tensor] = None, 94 | ): 95 | q = self.query(x) 96 | k = self.key(x) 97 | v = self.value(x) 98 | 99 | wv, qk = self.qkv_attention(q, k, v, mask) 100 | return self.out(wv), qk 101 | 102 | def qkv_attention(self, 103 | q: Tensor, 104 | k: Tensor, 105 | v: Tensor, 106 | mask: Optional[Tensor] = None): 107 | _, _, D = q.shape 108 | scale = (D // self.n_head)**-0.25 109 | q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale 110 | k = k.view(*k.shape[:2], self.n_head, -1) 111 | v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) 112 | 113 | if not self.use_sdpa: 114 | k = k.permute(0, 2, 3, 1) * scale 115 | qk = q @ k # (B, n_head, T, T) 116 | if mask is not None: 117 | qk = qk + mask 118 | qk = qk.float() 119 | w = torch.nn.functional.softmax(qk, dim=-1).to(q.dtype) 120 | return (w @ v).permute(0, 2, 1, 121 | 3).flatten(start_dim=2), qk.detach() 122 | else: 123 | k = k.permute(0, 2, 1, 3) * scale 124 | assert mask is not None 125 | output = torch.nn.functional.scaled_dot_product_attention( 126 | q, 127 | k, 128 | v, 129 | attn_mask=mask, 130 | dropout_p=0., 131 | scale=1., 132 | ) 133 | output = (output.transpose(1, 134 | 2).contiguous().view(q.size(0), -1, D) 135 | ) # (batch, time1, d_model) 136 | return output, None 137 | 138 | 139 | class ResidualAttentionBlock(nn.Module): 140 | 141 | def __init__(self, n_state: int, n_head: int, use_sdpa: bool): 142 | super().__init__() 143 | 144 | self.attn = MultiHeadAttention(n_state, n_head, use_sdpa=use_sdpa) 145 | self.attn_ln = LayerNorm(n_state) 146 | 147 | n_mlp = n_state * 4 148 | self.mlp = nn.Sequential(Linear(n_state, n_mlp), nn.GELU(), 149 | Linear(n_mlp, n_state)) 150 | self.mlp_ln = LayerNorm(n_state) 151 | 152 | def forward( 153 | self, 154 | x: Tensor, 155 | mask: Optional[Tensor] = None, 156 | ): 157 | x = x + self.attn(self.attn_ln(x), mask=mask)[0] 158 | x = x + self.mlp(self.mlp_ln(x)) 159 | return x 160 | 161 | 162 | class AudioEncoder(nn.Module): 163 | 164 | def __init__( 165 | self, 166 | n_mels: int, 167 | n_ctx: int, 168 | n_state: int, 169 | n_head: int, 170 | n_layer: int, 171 | stride: int, 172 | use_sdpa: bool, 173 | ): 174 | super().__init__() 175 | self.stride = stride 176 | self.conv1 = Conv1d(n_mels, 177 | n_state, 178 | kernel_size=3, 179 | stride=stride, 180 | padding=1) 181 | self.conv2 = Conv1d(n_state, 182 | n_state, 183 | kernel_size=3, 184 | stride=2, 185 | padding=1) 186 | self.register_buffer("positional_embedding", sinusoids(n_ctx, n_state)) 187 | 188 | self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList([ 189 | ResidualAttentionBlock(n_state, n_head, use_sdpa=use_sdpa) 190 | for _ in range(n_layer) 191 | ]) 192 | 193 | def forward(self, x: Tensor, x_len: Tensor) -> Tuple[Tensor, Tensor]: 194 | """ 195 | x : torch.Tensor, shape = (batch_size, n_mels, T) 196 | the mel spectrogram of the audio 197 | x_len: torch.Tensor, shape = (batch_size,) 198 | length of each audio in x 199 | """ 200 | mask = make_non_pad_mask(x_len).unsqueeze(1) 201 | x = F.gelu(self.conv1(x * mask)) 202 | x_len = (x_len + 2 - 1 * (3 - 1) - 1) // self.stride + 1 203 | mask = make_non_pad_mask(x_len).unsqueeze(1) 204 | x = F.gelu(self.conv2(x * mask)) 205 | x_len = (x_len + 2 - 1 * (3 - 1) - 1) // 2 + 1 206 | mask = make_non_pad_mask(x_len).unsqueeze(1) 207 | x = x.permute(0, 2, 1) # (B, T // 2, n_state) 208 | 209 | mask = mask_to_bias(mask, x.dtype) 210 | 211 | x = (x + self.positional_embedding[:x.shape[1], :]).to(x.dtype) 212 | 213 | for block in self.blocks: 214 | x = block(x, mask.unsqueeze(1)) 215 | 216 | return x, x_len 217 | 218 | 219 | class EuclideanCodebook(nn.Module): 220 | """Codebook with Euclidean distance (inference-only). 221 | Args: 222 | dim (int): Dimension. 223 | codebook_size (int): Codebook size. 224 | """ 225 | 226 | def __init__(self, dim: int, codebook_size: int): 227 | super().__init__() 228 | embed = torch.zeros(codebook_size, dim) 229 | self.codebook_size = codebook_size 230 | self.register_buffer("embed", embed) 231 | 232 | @torch.inference_mode() 233 | def preprocess(self, x: Tensor) -> Tensor: 234 | x = rearrange(x, "... d -> (...) d") 235 | return x 236 | 237 | @torch.inference_mode() 238 | def quantize(self, x: Tensor) -> Tensor: 239 | embed = self.embed.t() 240 | dist = -(x.pow(2).sum(1, keepdim=True) - 2 * x @ embed + 241 | embed.pow(2).sum(0, keepdim=True)) 242 | embed_ind = dist.max(dim=-1).indices 243 | return embed_ind 244 | 245 | @torch.inference_mode() 246 | def postprocess_emb(self, embed_ind, shape): 247 | return embed_ind.view(*shape[:-1]) 248 | 249 | @torch.inference_mode() 250 | def dequantize(self, embed_ind: Tensor) -> Tensor: 251 | quantize = F.embedding(embed_ind, self.embed) 252 | return quantize 253 | 254 | @torch.inference_mode() 255 | def encode(self, x: Tensor) -> Tensor: 256 | shape = x.shape 257 | # pre-process 258 | x = self.preprocess(x) 259 | # quantize 260 | embed_ind = self.quantize(x) 261 | # post-process 262 | embed_ind = self.postprocess_emb(embed_ind, shape) 263 | return embed_ind 264 | 265 | @torch.inference_mode() 266 | def decode(self, embed_ind: Tensor) -> Tensor: 267 | quantize = self.dequantize(embed_ind) 268 | return quantize 269 | 270 | 271 | class VectorQuantization(nn.Module): 272 | """Vector quantization implementation (inference-only). 273 | Args: 274 | dim (int): Dimension 275 | codebook_size (int): Codebook size 276 | """ 277 | 278 | def __init__(self, dim: int, codebook_size: int): 279 | super().__init__() 280 | self._codebook = EuclideanCodebook(dim=dim, 281 | codebook_size=codebook_size) 282 | self.codebook_size = codebook_size 283 | 284 | @property 285 | def codebook(self): 286 | return self._codebook.embed 287 | 288 | @torch.inference_mode() 289 | def encode(self, x: Tensor) -> Tensor: 290 | x = F.normalize(x, p=2, dim=-1) 291 | embed_in = self._codebook.encode(x) 292 | return embed_in 293 | 294 | @torch.inference_mode() 295 | def decode(self, embed_ind: Tensor) -> Tensor: 296 | quantize = self._codebook.decode(embed_ind) 297 | quantize = rearrange(quantize, "b n d -> b d n") 298 | return quantize 299 | 300 | 301 | class S3Tokenizer(nn.Module): 302 | """S3 tokenizer implementation (inference-only). 303 | Args: 304 | config (ModelConfig): Config 305 | """ 306 | 307 | def __init__(self, name: str, config: ModelConfig = ModelConfig()): 308 | super().__init__() 309 | self.config = config 310 | self.encoder = AudioEncoder( 311 | self.config.n_mels, 312 | self.config.n_audio_ctx, 313 | self.config.n_audio_state, 314 | self.config.n_audio_head, 315 | self.config.n_audio_layer, 316 | 2 if name == "speech_tokenizer_v1_25hz" else 1, 317 | self.config.use_sdpa, 318 | ) 319 | self.quantizer = VectorQuantization(self.config.n_audio_state, 320 | self.config.n_codebook_size) 321 | 322 | def forward(self, mel: Tensor, mel_len: Tensor) -> Tuple[Tensor, Tensor]: 323 | return self.quantize(mel, mel_len) 324 | 325 | @torch.inference_mode() 326 | def quantize(self, mel: Tensor, mel_len: Tensor) -> Tuple[Tensor, Tensor]: 327 | hidden, code_len = self.encoder(mel, mel_len) 328 | code = self.quantizer.encode(hidden) 329 | return code, code_len 330 | 331 | @property 332 | def device(self): 333 | return next(self.parameters()).device 334 | 335 | def init_from_onnx(self, onnx_path: str): 336 | ckpt = onnx2torch(onnx_path, None, False) 337 | self.load_state_dict(ckpt, strict=True) 338 | 339 | def init_from_pt(self, ckpt_path: str): 340 | ckpt = torch.load(ckpt_path, map_location="cpu", mmap=True) 341 | self.load_state_dict(ckpt, strict=True) 342 | 343 | def freeze(self): 344 | for _, param in self.named_parameters(): 345 | param.requires_grad = False 346 | -------------------------------------------------------------------------------- /s3tokenizer/model_v2.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) (Mddct: Dinghao Zhou) 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from dataclasses import dataclass 16 | from typing import Optional, Tuple 17 | 18 | import torch 19 | from einops import rearrange 20 | 21 | from s3tokenizer.model import Conv1d, LayerNorm, Linear, MultiHeadAttention 22 | from s3tokenizer.utils import make_non_pad_mask, mask_to_bias, onnx2torch 23 | 24 | 25 | @dataclass 26 | class ModelConfig: 27 | n_mels: int = 128 28 | n_audio_ctx: int = 1500 29 | n_audio_state: int = 1280 30 | n_audio_head: int = 20 31 | n_audio_layer: int = 6 32 | n_codebook_size: int = 3**8 33 | 34 | use_sdpa: bool = False 35 | 36 | 37 | def precompute_freqs_cis(dim: int, 38 | end: int, 39 | theta: float = 10000.0, 40 | scaling=None): 41 | freqs = 1.0 / (theta**(torch.arange(0, dim, 2)[:(dim // 2)].float() / dim)) 42 | t = torch.arange(end, device=freqs.device) # type: ignore 43 | if scaling is not None: 44 | t = t * scaling 45 | freqs = torch.outer(t, freqs).float() # type: ignore 46 | freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 47 | 48 | return torch.cat((freqs_cis, freqs_cis), dim=-1) 49 | 50 | 51 | def apply_rotary_emb( 52 | xq: torch.Tensor, 53 | xk: torch.Tensor, 54 | freqs_cis: torch.Tensor, 55 | ) -> Tuple[torch.Tensor, torch.Tensor]: 56 | real = torch.view_as_real(freqs_cis) 57 | cos, sin = real[:, :, 0], real[:, :, 1] 58 | cos = cos.unsqueeze(0).unsqueeze(2) 59 | sin = sin.unsqueeze(0).unsqueeze(2) 60 | 61 | D = xq.shape[-1] 62 | half_l, half_r = xq[:, :, :, :D // 2], xq[:, :, :, D // 2:] 63 | xq_r = torch.cat((-half_r, half_l), dim=-1) 64 | 65 | D = xk.shape[-1] 66 | 67 | half_l, half_r = xk[:, :, :, :D // 2], xk[:, :, :, D // 2:] 68 | xk_r = torch.cat((-half_r, half_l), dim=-1) 69 | 70 | return xq * cos + xq_r * sin, xk * cos + xk_r * sin 71 | 72 | 73 | def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor): 74 | ndim = x.ndim 75 | assert 0 <= 1 < ndim 76 | assert freqs_cis.shape == (x.shape[1], x.shape[-1]) 77 | shape = [ 78 | d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape) 79 | ] 80 | return freqs_cis.view(*shape) 81 | 82 | 83 | class FSQCodebook(torch.nn.Module): 84 | 85 | def __init__(self, dim: int, level: int = 3): 86 | super().__init__() 87 | self.project_down = torch.nn.Linear(dim, 8) 88 | self.level = level 89 | self.embed = None 90 | 91 | @torch.inference_mode() 92 | def preprocess(self, x: torch.Tensor) -> torch.Tensor: 93 | x = rearrange(x, "... d -> (...) d") 94 | return x 95 | 96 | @torch.inference_mode() 97 | def encode(self, x: torch.Tensor) -> torch.Tensor: 98 | x_shape = x.shape 99 | # pre-process 100 | x = self.preprocess(x) 101 | # quantize 102 | h = self.project_down(x).float() 103 | h = h.tanh() 104 | h = h * 0.9990000128746033 105 | h = h.round() + 1 106 | # h = ((self.level - 1) * h).round() # range [-k, k] 107 | powers = torch.pow( 108 | self.level, 109 | torch.arange(2**self.level, device=x.device, dtype=h.dtype)) 110 | mu = torch.sum(h * powers.unsqueeze(0), dim=-1) 111 | ind = mu.reshape(x_shape[0], x_shape[1]).int() 112 | return ind 113 | 114 | @torch.inference_mode() 115 | def decode(self, embed_ind: torch.Tensor) -> torch.Tensor: 116 | raise NotImplementedError( 117 | 'There is no official up project component provided') 118 | 119 | 120 | class FSQVectorQuantization(torch.nn.Module): 121 | """Vector quantization implementation (inference-only). 122 | Args: 123 | dim (int): Dimension 124 | codebook_size (int): Codebook size 125 | """ 126 | 127 | def __init__( 128 | self, 129 | dim: int, 130 | codebook_size: int, 131 | ): 132 | super().__init__() 133 | assert 3**8 == codebook_size 134 | self._codebook = FSQCodebook(dim=dim, level=3) 135 | self.codebook_size = codebook_size 136 | 137 | @property 138 | def codebook(self): 139 | return self._codebook.embed 140 | 141 | @torch.inference_mode() 142 | def encode(self, x: torch.Tensor) -> torch.Tensor: 143 | return self._codebook.encode(x) 144 | 145 | @torch.inference_mode() 146 | def decode(self, embed_ind: torch.Tensor) -> torch.Tensor: 147 | quantize = self._codebook.decode(embed_ind) 148 | quantize = rearrange(quantize, "b n d -> b d n") 149 | return quantize 150 | 151 | 152 | class FSMNMultiHeadAttention(MultiHeadAttention): 153 | 154 | def __init__( 155 | self, 156 | n_state: int, 157 | n_head: int, 158 | kernel_size: int = 31, 159 | use_sdpa: bool = False, 160 | ): 161 | super().__init__(n_state, n_head) 162 | 163 | self.fsmn_block = torch.nn.Conv1d(n_state, 164 | n_state, 165 | kernel_size, 166 | stride=1, 167 | padding=0, 168 | groups=n_state, 169 | bias=False) 170 | self.left_padding = (kernel_size - 1) // 2 171 | self.right_padding = kernel_size - 1 - self.left_padding 172 | self.pad_fn = torch.nn.ConstantPad1d( 173 | (self.left_padding, self.right_padding), 0.0) 174 | 175 | self.use_sdpa = use_sdpa 176 | 177 | def forward_fsmn(self, 178 | inputs: torch.Tensor, 179 | mask: Optional[torch.Tensor] = None): 180 | b, t, _, _ = inputs.size() 181 | inputs = inputs.view(b, t, -1) 182 | if mask is not None and mask.size(2) > 0: # time2 > 0 183 | inputs = inputs * mask 184 | x = inputs.transpose(1, 2) 185 | x = self.pad_fn(x) 186 | x = self.fsmn_block(x) 187 | x = x.transpose(1, 2) 188 | x += inputs 189 | return x * mask 190 | 191 | def qkv_attention(self, 192 | q: torch.Tensor, 193 | k: torch.Tensor, 194 | v: torch.Tensor, 195 | mask: Optional[torch.Tensor] = None, 196 | mask_pad: Optional[torch.Tensor] = None, 197 | freqs_cis: Optional[torch.Tensor] = None): 198 | _, _, D = q.shape 199 | scale = (D // self.n_head)**-0.25 200 | q = q.view(*q.shape[:2], self.n_head, -1) 201 | k = k.view(*k.shape[:2], self.n_head, -1) 202 | v = v.view(*v.shape[:2], self.n_head, -1) 203 | 204 | if freqs_cis is not None: 205 | q, k = apply_rotary_emb(q, k, freqs_cis=freqs_cis) 206 | 207 | fsm_memory = self.forward_fsmn(v, mask_pad) 208 | 209 | q = q.permute(0, 2, 1, 3) * scale 210 | v = v.permute(0, 2, 1, 3) 211 | 212 | if not self.use_sdpa: 213 | k = k.permute(0, 2, 3, 1) * scale 214 | qk = q @ k # (B, n_head, T, T) 215 | if mask is not None: 216 | qk = qk + mask 217 | qk = qk.float() 218 | w = torch.nn.functional.softmax(qk, dim=-1).to(q.dtype) 219 | return (w @ v).permute( 220 | 0, 2, 1, 3).flatten(start_dim=2), qk.detach(), fsm_memory 221 | else: 222 | k = k.permute(0, 2, 1, 3) * scale 223 | assert mask is not None 224 | output = torch.nn.functional.scaled_dot_product_attention( 225 | q, 226 | k, 227 | v, 228 | attn_mask=mask, 229 | dropout_p=0., 230 | scale=1., 231 | ) 232 | output = (output.transpose(1, 233 | 2).contiguous().view(q.size(0), -1, D) 234 | ) # (batch, time1, d_model) 235 | return output, None, fsm_memory 236 | 237 | def forward(self, 238 | x: torch.Tensor, 239 | mask: Optional[torch.Tensor] = None, 240 | mask_pad: Optional[torch.Tensor] = None, 241 | freqs_cis: Optional[torch.Tensor] = None): 242 | 243 | q = self.query(x) 244 | k = self.key(x) 245 | v = self.value(x) 246 | 247 | wv, qk, fsm_memory = self.qkv_attention(q, k, v, mask, mask_pad, 248 | freqs_cis) 249 | return self.out(wv) + fsm_memory, qk 250 | 251 | 252 | class ResidualAttentionBlock(torch.nn.Module): 253 | 254 | def __init__( 255 | self, 256 | n_state: int, 257 | n_head: int, 258 | kernel_size: int = 31, 259 | use_sdpa: bool = False, 260 | ): 261 | super().__init__() 262 | 263 | self.attn = FSMNMultiHeadAttention(n_state, 264 | n_head, 265 | kernel_size, 266 | use_sdpa=use_sdpa) 267 | self.attn_ln = LayerNorm(n_state, eps=1e-6) 268 | 269 | n_mlp = n_state * 4 270 | 271 | self.mlp = torch.nn.Sequential(Linear(n_state, n_mlp), torch.nn.GELU(), 272 | Linear(n_mlp, n_state)) 273 | self.mlp_ln = LayerNorm(n_state) 274 | 275 | def forward( 276 | self, 277 | x: torch.Tensor, 278 | mask: Optional[torch.Tensor] = None, 279 | mask_pad: Optional[torch.Tensor] = None, 280 | freqs_cis: Optional[torch.Tensor] = None, 281 | ): 282 | x = x + self.attn( 283 | self.attn_ln(x), mask=mask, mask_pad=mask_pad, 284 | freqs_cis=freqs_cis)[0] 285 | 286 | x = x + self.mlp(self.mlp_ln(x)) 287 | return x 288 | 289 | 290 | class AudioEncoderV2(torch.nn.Module): 291 | 292 | def __init__( 293 | self, 294 | n_mels: int, 295 | n_state: int, 296 | n_head: int, 297 | n_layer: int, 298 | stride: int, 299 | use_sdpa: bool, 300 | ): 301 | super().__init__() 302 | self.stride = stride 303 | 304 | self.conv1 = Conv1d(n_mels, 305 | n_state, 306 | kernel_size=3, 307 | stride=stride, 308 | padding=1) 309 | self.conv2 = Conv1d(n_state, 310 | n_state, 311 | kernel_size=3, 312 | stride=2, 313 | padding=1) 314 | self.freqs_cis = precompute_freqs_cis(64, 1024 * 2) 315 | self.blocks = torch.nn.ModuleList([ 316 | ResidualAttentionBlock(n_state, n_head, use_sdpa=use_sdpa) 317 | for _ in range(n_layer) 318 | ]) 319 | 320 | def forward(self, x: torch.Tensor, 321 | x_len: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 322 | """ 323 | x : torch.Tensor, shape = (batch_size, n_mels, T) 324 | the mel spectrogram of the audio 325 | x_len: torch.Tensor, shape = (batch_size,) 326 | length of each audio in x 327 | """ 328 | mask = make_non_pad_mask(x_len).unsqueeze(1) 329 | x = torch.nn.functional.gelu(self.conv1(x * mask)) 330 | x_len = (x_len + 2 - 1 * (3 - 1) - 1) // self.stride + 1 331 | mask = make_non_pad_mask(x_len).unsqueeze(1) 332 | x = torch.nn.functional.gelu(self.conv2(x * mask)) 333 | x_len = (x_len + 2 - 1 * (3 - 1) - 1) // 2 + 1 334 | mask = make_non_pad_mask(x_len).unsqueeze(1) 335 | x = x.permute(0, 2, 1) # (B, T // 2, n_state) 336 | freqs_cis = self.freqs_cis.to(x.device) 337 | mask_pad = mask.transpose(1, 2) 338 | mask = mask_to_bias(mask, x.dtype) 339 | 340 | tmp = torch.view_as_real(freqs_cis) 341 | cos, sin = tmp[:, :, 0], tmp[:, :, 1] 342 | 343 | cos = torch.cat((cos, cos), dim=-1) 344 | sin = torch.cat((sin, sin), dim=-1) 345 | cos = cos.unsqueeze(0).unsqueeze(2) 346 | sin = sin.unsqueeze(0).unsqueeze(2) 347 | 348 | for block in self.blocks: 349 | x = block(x, mask.unsqueeze(1), mask_pad, freqs_cis[:x.size(1)]) 350 | 351 | return x, x_len 352 | 353 | 354 | class S3TokenizerV2(torch.nn.Module): 355 | """S3 tokenizer v2 implementation (inference-only). 356 | Args: 357 | config (ModelConfig): Config 358 | """ 359 | 360 | def __init__(self, name: str, config: ModelConfig = ModelConfig()): 361 | super().__init__() 362 | if 'v1' not in name: 363 | assert 'v2' in name 364 | # TODO(Mddct): make it configureable 365 | config.n_codebook_size = 3**8 366 | self.config = config 367 | self.encoder = AudioEncoderV2( 368 | self.config.n_mels, 369 | self.config.n_audio_state, 370 | self.config.n_audio_head, 371 | self.config.n_audio_layer, 372 | 2, 373 | self.config.use_sdpa, 374 | ) 375 | self.quantizer = FSQVectorQuantization( 376 | self.config.n_audio_state, 377 | self.config.n_codebook_size, 378 | ) 379 | 380 | def forward(self, mel: torch.Tensor, 381 | mel_len: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 382 | return self.quantize(mel, mel_len) 383 | 384 | @torch.inference_mode() 385 | def quantize(self, mel: torch.Tensor, 386 | mel_len: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 387 | hidden, code_len = self.encoder(mel, mel_len) 388 | code = self.quantizer.encode(hidden) 389 | return code, code_len 390 | 391 | @property 392 | def device(self): 393 | return next(self.parameters()).device 394 | 395 | def init_from_onnx(self, onnx_path: str): 396 | ckpt = onnx2torch(onnx_path, None, False) 397 | self.load_state_dict(ckpt, strict=True) 398 | 399 | def init_from_pt(self, ckpt_path: str): 400 | ckpt = torch.load(ckpt_path, map_location="cpu", mmap=True) 401 | self.load_state_dict(ckpt, strict=True) 402 | 403 | def freeze(self): 404 | for _, param in self.named_parameters(): 405 | param.requires_grad = False 406 | -------------------------------------------------------------------------------- /s3tokenizer/utils.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2023 OpenAI. (authors: Whisper Team) 2 | # 2024 Tsinghua Univ. (authors: Xingchen Song) 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Modified from https://github.com/openai/whisper/blob/main/whisper/audio.py 16 | Add rename_weights() & onnx2torch() & make_non_pad_mask() & mask_to_bias() 17 | """ 18 | 19 | import os 20 | from functools import lru_cache 21 | from typing import List, Optional, Union 22 | 23 | import numpy as np 24 | import onnx 25 | import torch 26 | import torch.nn.functional as F 27 | import torchaudio 28 | from torch.nn.utils.rnn import pad_sequence 29 | 30 | 31 | def _rename_weights(weights_dict: dict): 32 | """ 33 | Rename onnx weights to pytorch format. 34 | 35 | Parameters 36 | ---------- 37 | weight_dict: dict 38 | The dict containing weights in onnx format 39 | 40 | Returns 41 | ------- 42 | A new weight dict containing the weights in pytorch format. 43 | """ 44 | new_weight_dict = {} 45 | for k in weights_dict.keys(): 46 | if "quantizer" in k: # vq or fsq 47 | if k == "/quantizer/rq/model/layers.0/_codebook/Pow_1": 48 | new_weight_dict["quantizer._codebook.embed"] = weights_dict[k] 49 | elif 'project_down' in k: # v2 50 | new_weight_dict[k] = weights_dict[k] 51 | elif "positional_embedding" in k: # positional emb 52 | new_weight_dict[k] = weights_dict[k] 53 | elif "conv" in k: # 1/2 or 1/4 subsample 54 | new_weight_dict[k] = weights_dict[k] 55 | else: # transformer blocks 56 | assert "blocks" in k 57 | new_k = (k[1:].replace('/', '.').replace( 58 | 'MatMul', 'weight').replace('Add_1', 'bias').replace( 59 | 'Mul', 'weight').replace('Add', 'bias').replace( 60 | 'mlp.mlp', 'mlp')).replace('fsmn_block.Conv', 61 | 'fsmn_block.weight') 62 | 63 | new_weight_dict[f"encoder.{new_k}"] = weights_dict[k] 64 | return new_weight_dict 65 | 66 | 67 | def onnx2torch(onnx_path: str, torch_path: str = None, verbose: bool = False): 68 | """ 69 | Open an onnx file and convert to pytorch format. 70 | 71 | Parameters 72 | ---------- 73 | onnx_path: str 74 | The onnx file to open, typically `speech_tokenizer_v1.onnx` 75 | 76 | torch_path: str 77 | The path to save the torch-formated checkpoint. 78 | 79 | verbose: bool 80 | Logging info or not. 81 | 82 | Returns 83 | ------- 84 | A checkpoint dict containing the weights and their names, if torch_path is 85 | None. Otherwise save checkpoint dict to the desired path. 86 | """ 87 | onnx_model = onnx.load(onnx_path) 88 | weights_dict = {} 89 | initializer_map = { 90 | initializer.name: initializer 91 | for initializer in onnx_model.graph.initializer 92 | } 93 | for node in onnx_model.graph.node: 94 | for input_name in node.input: 95 | if input_name in initializer_map: 96 | ln_bias_name, ln_weight_name = None, None # for v2 ln 97 | initializer = initializer_map[input_name] 98 | if input_name in [ 99 | "onnx::Conv_1519", 100 | "encoders.conv1.weight", 101 | "onnx::Conv_2216", 102 | ]: # v1_50hz, v1_25hz, v2_25hz 103 | weight_name = "encoder.conv1.weight" 104 | elif input_name in [ 105 | "onnx::Conv_1520", 106 | "encoders.conv1.bias", 107 | "onnx::Conv_2217", 108 | ]: # v1_50hz, v1_25hz, v2_25hz 109 | weight_name = "encoder.conv1.bias" 110 | elif input_name in [ 111 | "onnx::Conv_1521", 112 | "encoders.conv2.weight", 113 | "onnx::Conv_2218", 114 | ]: 115 | weight_name = "encoder.conv2.weight" 116 | elif input_name in [ 117 | "onnx::Conv_1522", 118 | "encoders.conv2.bias", 119 | "onnx::Conv_2219", 120 | ]: 121 | weight_name = "encoder.conv2.bias" 122 | elif input_name == "encoders.positional_embedding": 123 | weight_name = "encoder.positional_embedding" 124 | elif input_name == 'quantizer.project_in.bias': 125 | weight_name = "quantizer._codebook.project_down.bias" 126 | elif input_name == 'onnx::MatMul_2536': 127 | weight_name = "quantizer._codebook.project_down.weight" 128 | else: 129 | if node.op_type == 'LayerNormalization': # in input_name: 130 | ln_name = node.name.replace('/LayerNormalization', '') 131 | ln_weight_name = ln_name + '.weight' 132 | ln_bias_name = ln_name + '.bias' 133 | else: 134 | weight_name = node.name 135 | if ln_weight_name is not None and ln_bias_name is not None: 136 | ln_inputs = node.input 137 | scale_name = ln_inputs[1] 138 | bias_name = ln_inputs[2] 139 | scale = onnx.numpy_helper.to_array( 140 | initializer_map[scale_name]).copy( 141 | ) if scale_name in initializer_map else None 142 | bias = onnx.numpy_helper.to_array( 143 | initializer_map[bias_name]).copy( 144 | ) if bias_name in initializer_map else None 145 | scale.flags.writeable = True 146 | bias.flags.writeable = True 147 | weight_tensor = torch.from_numpy(scale) 148 | bias_tensor = torch.from_numpy(bias) 149 | 150 | weights_dict[ln_bias_name] = bias_tensor 151 | weights_dict[ln_weight_name] = weight_tensor 152 | else: 153 | weight_array = onnx.numpy_helper.to_array( 154 | initializer).copy() 155 | weight_array.flags.writeable = True 156 | weight_tensor = torch.from_numpy(weight_array) 157 | if len(weight_tensor.shape) > 2 or weight_name in [ 158 | "encoder.positional_embedding" 159 | ]: 160 | weights_dict[weight_name] = weight_tensor 161 | else: 162 | weights_dict[weight_name] = weight_tensor.t() 163 | 164 | new_weights_dict = _rename_weights(weights_dict) 165 | if verbose: 166 | for k, v in new_weights_dict.items(): 167 | print(f"{k} : {v.shape} {v.dtype}") 168 | print(f"PyTorch weights saved to {torch_path}") 169 | del weights_dict, onnx_model 170 | if torch_path: 171 | torch.save(new_weights_dict, torch_path) 172 | else: 173 | return new_weights_dict 174 | 175 | 176 | def load_audio(file: str, sr: int = 16000): 177 | """ 178 | Open an audio file and read as mono waveform, resampling as necessary 179 | 180 | Parameters 181 | ---------- 182 | file: str 183 | The audio file to open 184 | 185 | sr: int 186 | The sample rate to resample the audio if necessary 187 | 188 | Returns 189 | ------- 190 | A torch.Tensor containing the audio waveform, in float32 dtype. 191 | """ 192 | audio, sample_rate = torchaudio.load(file) 193 | if sample_rate != sr: 194 | audio = torchaudio.transforms.Resample(sample_rate, sr)(audio) 195 | audio = audio[0] # get the first channel 196 | return audio 197 | 198 | 199 | @lru_cache(maxsize=None) 200 | def _mel_filters(device, n_mels: int) -> torch.Tensor: 201 | """ 202 | load the mel filterbank matrix for projecting STFT into a Mel spectrogram. 203 | Allows decoupling librosa dependency; saved using: 204 | 205 | np.savez_compressed( 206 | "mel_filters.npz", 207 | mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80), 208 | mel_128=librosa.filters.mel(sr=16000, n_fft=400, n_mels=128), 209 | ) 210 | """ 211 | assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}" 212 | 213 | filters_path = os.path.join(os.path.dirname(__file__), "assets", 214 | "mel_filters.npz") 215 | with np.load(filters_path, allow_pickle=False) as f: 216 | return torch.from_numpy(f[f"mel_{n_mels}"]).to(device) 217 | 218 | 219 | def log_mel_spectrogram( 220 | audio: Union[str, np.ndarray, torch.Tensor], 221 | n_mels: int = 128, 222 | padding: int = 0, 223 | device: Optional[Union[str, torch.device]] = None, 224 | ): 225 | """ 226 | Compute the log-Mel spectrogram of 227 | 228 | Parameters 229 | ---------- 230 | audio: Union[str, np.ndarray, torch.Tensor], shape = (*) 231 | The path to audio or either a NumPy array or Tensor containing the 232 | audio waveform in 16 kHz 233 | 234 | n_mels: int 235 | The number of Mel-frequency filters, only 80 is supported 236 | 237 | padding: int 238 | Number of zero samples to pad to the right 239 | 240 | device: Optional[Union[str, torch.device]] 241 | If given, the audio tensor is moved to this device before STFT 242 | 243 | Returns 244 | ------- 245 | torch.Tensor, shape = (128, n_frames) 246 | A Tensor that contains the Mel spectrogram 247 | """ 248 | if not torch.is_tensor(audio): 249 | if isinstance(audio, str): 250 | audio = load_audio(audio) 251 | audio = torch.from_numpy(audio) 252 | 253 | if device is not None: 254 | audio = audio.to(device) 255 | if padding > 0: 256 | audio = F.pad(audio, (0, padding)) 257 | window = torch.hann_window(400).to(audio.device) 258 | stft = torch.stft(audio, 400, 160, window=window, return_complex=True) 259 | magnitudes = stft[..., :-1].abs()**2 260 | 261 | filters = _mel_filters(audio.device, n_mels) 262 | mel_spec = filters @ magnitudes 263 | 264 | log_spec = torch.clamp(mel_spec, min=1e-10).log10() 265 | log_spec = torch.maximum(log_spec, log_spec.max() - 8.0) 266 | log_spec = (log_spec + 4.0) / 4.0 267 | return log_spec 268 | 269 | 270 | def make_non_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor: 271 | """Make mask tensor containing indices of non-padded part. 272 | 273 | The sequences in a batch may have different lengths. To enable 274 | batch computing, padding is need to make all sequence in same 275 | size. To avoid the padding part pass value to context dependent 276 | block such as attention or convolution , this padding part is 277 | masked. 278 | 279 | 1 for non-padded part and 0 for padded part. 280 | 281 | Parameters 282 | ---------- 283 | lengths (torch.Tensor): Batch of lengths (B,). 284 | 285 | Returns: 286 | ------- 287 | torch.Tensor: Mask tensor containing indices of padded part (B, max_T). 288 | 289 | Examples: 290 | >>> import torch 291 | >>> import s3tokenizer 292 | >>> lengths = torch.tensor([5, 3, 2]) 293 | >>> masks = s3tokenizer.make_non_pad_mask(lengths) 294 | masks = [[1, 1, 1, 1, 1], 295 | [1, 1, 1, 0, 0], 296 | [1, 1, 0, 0, 0]] 297 | """ 298 | batch_size = lengths.size(0) 299 | max_len = max_len if max_len > 0 else lengths.max().item() 300 | seq_range = torch.arange(0, 301 | max_len, 302 | dtype=torch.int64, 303 | device=lengths.device) 304 | seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len) 305 | seq_length_expand = lengths.unsqueeze(-1) 306 | mask = seq_range_expand >= seq_length_expand 307 | return ~mask 308 | 309 | 310 | def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor: 311 | """Convert bool-tensor to float-tensor for flash attention. 312 | 313 | Parameters 314 | ---------- 315 | lengths (torch.Tensor): Batch of lengths (B, ?). 316 | 317 | Returns: 318 | ------- 319 | torch.Tensor: Mask tensor containing indices of padded part (B, ?). 320 | 321 | Examples: 322 | >>> import torch 323 | >>> import s3tokenizer 324 | >>> lengths = torch.tensor([5, 3, 2]) 325 | >>> masks = s3tokenizer.make_non_pad_mask(lengths) 326 | masks = [[1, 1, 1, 1, 1], 327 | [1, 1, 1, 0, 0], 328 | [1, 1, 0, 0, 0]] 329 | >>> new_masks = s3tokenizer.mask_to_bias(masks, torch.float32) 330 | new_masks = 331 | [[-0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00], 332 | [-0.0000e+00, -0.0000e+00, -0.0000e+00, -1.0000e+10, -1.0000e+10], 333 | [-0.0000e+00, -0.0000e+00, -1.0000e+10, -1.0000e+10, -1.0000e+10]] 334 | """ 335 | assert mask.dtype == torch.bool 336 | assert dtype in [torch.float32, torch.bfloat16, torch.float16] 337 | mask = mask.to(dtype) 338 | 339 | # attention mask bias 340 | # NOTE(Mddct): torch.finfo jit issues 341 | # chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min 342 | mask = (1.0 - mask) * -1.0e+10 343 | return mask 344 | 345 | 346 | def padding(data: List[torch.Tensor]): 347 | """ Padding the data into batch data 348 | 349 | Parameters 350 | ---------- 351 | data: List[Tensor], shape of Tensor (128, T) 352 | 353 | Returns: 354 | ------- 355 | feats, feats lengths 356 | """ 357 | sample = data 358 | assert isinstance(sample, list) 359 | feats_lengths = torch.tensor([s.size(1) for s in sample], 360 | dtype=torch.int32) 361 | feats = [s.t() for s in sample] 362 | padded_feats = pad_sequence(feats, batch_first=True, padding_value=0) 363 | 364 | return padded_feats.transpose(1, 2), feats_lengths 365 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | from pathlib import Path 2 | 3 | from setuptools import find_packages, setup 4 | 5 | 6 | def parse_requirements(filename): 7 | """Load requirements from a pip requirements file.""" 8 | with open(filename, 'r') as file: 9 | lines = (line.strip() for line in file) 10 | return [line for line in lines if line and not line.startswith('#')] 11 | 12 | 13 | setup( 14 | name="s3tokenizer", 15 | version="0.1.7", 16 | description=\ 17 | "Reverse Engineering of Supervised Semantic Speech Tokenizer (S3Tokenizer) proposed in CosyVoice", # noqa 18 | long_description=open("README.md", encoding="utf-8").read(), 19 | long_description_content_type="text/markdown", 20 | python_requires=">=3.8", 21 | author="xingchensong", 22 | url="https://github.com/xingchensong/S3Tokenizer", 23 | license="Apache2.0", 24 | packages=find_packages(), 25 | install_requires=parse_requirements( 26 | Path(__file__).with_name("requirements.txt")), 27 | entry_points={ 28 | "console_scripts": ["s3tokenizer=s3tokenizer.cli:main"], 29 | }, 30 | include_package_data=True, 31 | extras_require={"dev": ["pytest", "scipy", "black", "flake8", "isort"]}, 32 | classifiers=[ 33 | "Programming Language :: Python :: 3", 34 | "Operating System :: OS Independent", 35 | "Topic :: Scientific/Engineering", 36 | ], 37 | ) 38 | -------------------------------------------------------------------------------- /test/test_onnx.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | # Copyright [2024-09-27] 4 | 5 | import os 6 | 7 | import numpy as np 8 | import onnxruntime 9 | import s3tokenizer 10 | import torch 11 | 12 | default = os.path.join(os.path.expanduser("~"), ".cache") 13 | download_root = os.path.join(os.getenv("XDG_CACHE_HOME", default), 14 | "s3tokenizer") 15 | name = "speech_tokenizer_v1" 16 | tokenizer = s3tokenizer.load_model(name) 17 | 18 | mels = [] 19 | wav_paths = [ 20 | "s3tokenizer/assets/BAC009S0764W0121.wav", 21 | "s3tokenizer/assets/BAC009S0764W0122.wav" 22 | ] 23 | for wav_path in wav_paths: 24 | audio = s3tokenizer.load_audio(wav_path) 25 | mels.append(s3tokenizer.log_mel_spectrogram(audio)) 26 | print("=========torch=============") 27 | mels, mels_lens = s3tokenizer.padding(mels) 28 | print(f"mels.size: {mels.size()}, mels_lens: {mels_lens}") 29 | codes, codes_lens = tokenizer.quantize(mels, mels_lens) 30 | print(f"codes.size: {codes.size()}, codes_lens: {codes_lens}") 31 | 32 | for i in range(len(wav_paths)): 33 | print(f"wav[{i}]") 34 | print(codes[i, :codes_lens[i].item()]) 35 | 36 | print("=========onnx===============") 37 | option = onnxruntime.SessionOptions() 38 | option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL # noqa 39 | option.intra_op_num_threads = 1 40 | providers = ["CPUExecutionProvider"] 41 | ort_session = onnxruntime.InferenceSession(f"{download_root}/{name}.onnx", 42 | sess_options=option, 43 | providers=providers) 44 | 45 | for i in range(len(wav_paths)): 46 | speech_token = ort_session.run( 47 | None, { 48 | ort_session.get_inputs()[0].name: 49 | mels[i, :, :mels_lens[i].item()].unsqueeze( 50 | 0).detach().cpu().numpy(), 51 | ort_session.get_inputs()[1].name: 52 | np.array([mels_lens[i].item()], dtype=np.int32) 53 | })[0] 54 | if name == 'speech_tokenizer_v2_25hz': 55 | speech_token = np.expand_dims(speech_token, 0) 56 | speech_token = torch.tensor(speech_token[0, 0, :]) 57 | print(f"wav[{i}]") 58 | print(speech_token) 59 | print( 60 | f"all equal: {torch.equal(speech_token, codes[i, :codes_lens[i].item()].cpu())}" # noqa 61 | ) 62 | miss_num = torch.sum( 63 | ~(speech_token == codes[i, :codes_lens[i].item()].cpu())) 64 | total = speech_token.numel() 65 | print(f"miss rate: {miss_num * 100.0 / total}%") 66 | --------------------------------------------------------------------------------