├── .dockerignore
├── .gitignore
├── .gitmodules
├── Dockerfile
├── LICENSE
├── README.md
├── bin
└── .gitkeep
├── changelog.md
├── config
├── .gitignore
└── rvc.json
├── deepspeed-0.8.3+6eca037c-cp39-cp39-win_amd64.whl
├── download_ffmpeg.bat
├── models
├── .gitignore
├── .template.dlas.yaml
├── rvc_models
│ └── .gitignore
├── tokenizers
│ ├── ipa.json
│ └── spanish_bpe_tokenizer.json
└── tortoise
│ └── train_diffusion_vocoder_22k_level.yml
├── modules
├── .gitignore
└── rvc
│ ├── .gitignore
│ ├── Dockerfile
│ ├── LICENSE
│ ├── MIT协议暨相关引用库协议
│ ├── README.md
│ ├── Retrieval_based_Voice_Conversion_WebUI.ipynb
│ ├── Retrieval_based_Voice_Conversion_WebUI_v2.ipynb
│ ├── changelog.md
│ ├── configs
│ ├── config.json
│ ├── config.py
│ ├── v1
│ │ ├── 32k.json
│ │ ├── 40k.json
│ │ └── 48k.json
│ └── v2
│ │ ├── 32k.json
│ │ └── 48k.json
│ ├── docker-compose.yml
│ ├── docs
│ ├── cn
│ │ ├── Changelog_CN.md
│ │ └── faq.md
│ ├── en
│ │ ├── Changelog_EN.md
│ │ ├── README.en.md
│ │ ├── faiss_tips_en.md
│ │ ├── faq_en.md
│ │ └── training_tips_en.md
│ ├── fr
│ │ ├── Changelog_FR.md
│ │ ├── README.fr.md
│ │ ├── faiss_tips_fr.md
│ │ ├── faq_fr.md
│ │ └── training_tips_fr.md
│ ├── jp
│ │ ├── README.ja.md
│ │ ├── faiss_tips_ja.md
│ │ └── training_tips_ja.md
│ ├── kr
│ │ ├── Changelog_KO.md
│ │ ├── README.ko.han.md
│ │ ├── README.ko.md
│ │ ├── faiss_tips_ko.md
│ │ └── training_tips_ko.md
│ ├── tr
│ │ ├── Changelog_TR.md
│ │ ├── README.tr.md
│ │ ├── faiss_tips_tr.md
│ │ ├── faq_tr.md
│ │ └── training_tips_tr.md
│ └── 小白简易教程.doc
│ ├── environment_dml.yaml
│ ├── go-realtime-gui-dml.bat
│ ├── go-realtime-gui.bat
│ ├── go-web-dml.bat
│ ├── go-web.bat
│ ├── gui_v1.py
│ ├── i18n
│ ├── i18n.py
│ ├── locale
│ │ ├── en_US.json
│ │ ├── es_ES.json
│ │ ├── fr_FR.json
│ │ ├── it_IT.json
│ │ ├── ja_JP.json
│ │ ├── ru_RU.json
│ │ ├── tr_TR.json
│ │ ├── zh_CN.json
│ │ ├── zh_HK.json
│ │ ├── zh_SG.json
│ │ └── zh_TW.json
│ ├── locale_diff.py
│ └── scan_i18n.py
│ ├── infer-web.py
│ ├── infer
│ └── modules
│ │ ├── ipex
│ │ ├── __init__.py
│ │ ├── attention.py
│ │ ├── gradscaler.py
│ │ └── hijacks.py
│ │ ├── onnx
│ │ └── export.py
│ │ ├── train
│ │ ├── extract
│ │ │ ├── extract_f0_print.py
│ │ │ ├── extract_f0_rmvpe.py
│ │ │ └── extract_f0_rmvpe_dml.py
│ │ ├── extract_feature_print.py
│ │ ├── preprocess.py
│ │ └── train.py
│ │ ├── uvr5
│ │ ├── mdxnet.py
│ │ ├── modules.py
│ │ └── preprocess.py
│ │ └── vc
│ │ ├── __init__.py
│ │ ├── modules.py
│ │ ├── pipeline.py
│ │ └── utils.py
│ ├── modules.py
│ ├── poetry.lock
│ ├── pyproject.toml
│ ├── requirements-amd.txt
│ ├── requirements-dml.txt
│ ├── requirements-ipex.txt
│ ├── requirements-win-for-realtime_vc_gui-dml.txt
│ ├── requirements-win-for-realtime_vc_gui.txt
│ ├── requirements.txt
│ ├── run.sh
│ ├── tools
│ ├── app.py
│ ├── calc_rvc_model_similarity.py
│ ├── dlmodels.bat
│ ├── dlmodels.sh
│ ├── export_onnx.py
│ ├── infer
│ │ ├── infer-pm-index256.py
│ │ ├── train-index-v2.py
│ │ ├── train-index.py
│ │ └── trans_weights.py
│ ├── infer_batch_rvc.py
│ ├── infer_cli.py
│ ├── onnx_inference_demo.py
│ ├── rvc_for_realtime.py
│ └── torchgate
│ │ ├── __init__.py
│ │ ├── torchgate.py
│ │ └── utils.py
│ └── venv.sh
├── notebook_colab.ipynb
├── notebook_paperspace.ipynb
├── reload_flag.txt
├── requirements.txt
├── setup-cuda-bnb.bat
├── setup-cuda-cpu.bat
├── setup-cuda-cpu.sh
├── setup-cuda.bat
├── setup-cuda.sh
├── setup-directml.bat
├── setup-docker.sh
├── setup-rocm-bnb.sh
├── setup-rocm.sh
├── setup-whisperx.bat
├── src
├── api
│ └── websocket_server.py
├── cli.py
├── list_devices.py
├── main.py
├── train.py
├── train_jard.py
├── utils.py
└── webui.py
├── start-docker.sh
├── start.bat
├── start.sh
├── train-docker.sh
├── train.bat
├── train.sh
├── training
└── .gitignore
├── update-force.bat
├── update-force.sh
├── update.bat
├── update.sh
├── update_package.bat
└── voices
├── .gitignore
└── random
├── cond_latents_d1f79232.pth
└── cond_latents_f540259e.pth
/.dockerignore:
--------------------------------------------------------------------------------
1 | /models
2 | /training
3 | /voices
4 | /bin
5 | Dockerfile
6 |
--------------------------------------------------------------------------------
/.gitignore:
--------------------------------------------------------------------------------
1 | # ignores user files
2 | /ffmpeg/
3 | /personal_testing/
4 | *.7z
5 | /venv/
6 | /venv_cpu/
7 | /training/*
8 | /config/*
9 | output/
10 | *.wav
11 | fairseq-0.12.2-cp39-cp39-win_amd64.whl
12 | rvc/
13 | *.zip
14 | pyfastmp3decoder/
15 | .vscode/
16 | onnx_test_stuff/
17 | *.whl
18 | *.exe
19 |
20 | # Byte-compiled / optimized / DLL files
21 | __pycache__/
22 | *.py[cod]
23 | *$py.class
24 |
25 | # C extensions
26 | *.so
27 |
28 | # Distribution / packaging
29 | .Python
30 | build/
31 | develop-eggs/
32 | dist/
33 | downloads/
34 | eggs/
35 | .eggs/
36 | lib/
37 | lib64/
38 | parts/
39 | sdist/
40 | var/
41 | wheels/
42 | pip-wheel-metadata/
43 | share/python-wheels/
44 | *.egg-info/
45 | .installed.cfg
46 | *.egg
47 | MANIFEST
48 |
49 | # PyInstaller
50 | # Usually these files are written by a python script from a template
51 | # before PyInstaller builds the exe, so as to inject date/other infos into it.
52 | *.manifest
53 | *.spec
54 |
55 | # Installer logs
56 | pip-log.txt
57 | pip-delete-this-directory.txt
58 |
59 | # Unit test / coverage reports
60 | htmlcov/
61 | .tox/
62 | .nox/
63 | .coverage
64 | .coverage.*
65 | .cache
66 | nosetests.xml
67 | coverage.xml
68 | *.cover
69 | *.py,cover
70 | .hypothesis/
71 | .pytest_cache/
72 |
73 | # Translations
74 | *.mo
75 | *.pot
76 |
77 | # Django stuff:
78 | *.log
79 | local_settings.py
80 | db.sqlite3
81 | db.sqlite3-journal
82 |
83 | # Flask stuff:
84 | instance/
85 | .webassets-cache
86 |
87 | # Scrapy stuff:
88 | .scrapy
89 |
90 | # Sphinx documentation
91 | docs/_build/
92 |
93 | # PyBuilder
94 | target/
95 |
96 | # Jupyter Notebook
97 | .ipynb_checkpoints
98 |
99 | # IPython
100 | profile_default/
101 | ipython_config.py
102 |
103 | # pyenv
104 | .python-version
105 |
106 | # pipenv
107 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
108 | # However, in case of collaboration, if having platform-specific dependencies or dependencies
109 | # having no cross-platform support, pipenv may install dependencies that don't work, or not
110 | # install all needed dependencies.
111 | #Pipfile.lock
112 |
113 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow
114 | __pypackages__/
115 |
116 | # Celery stuff
117 | celerybeat-schedule
118 | celerybeat.pid
119 |
120 | # SageMath parsed files
121 | *.sage.py
122 |
123 | # Environments
124 | .env
125 | .venv
126 | env/
127 | venv/
128 | ENV/
129 | env.bak/
130 | venv.bak/
131 |
132 | # Spyder project settings
133 | .spyderproject
134 | .spyproject
135 |
136 | # Rope project settings
137 | .ropeproject
138 |
139 | # mkdocs documentation
140 | /site
141 |
142 | # mypy
143 | .mypy_cache/
144 | .dmypy.json
145 | dmypy.json
146 |
147 | # Pyre type checker
148 | .pyre/
149 |
150 | .idea/*
151 |
152 | .custom/*
153 | results/*
154 | debug_states/*
155 | bin/*
156 |
157 |
--------------------------------------------------------------------------------
/.gitmodules:
--------------------------------------------------------------------------------
1 | [submodule "modules/tortoise-tts"]
2 | path = modules/tortoise-tts
3 | url = https://github.com/JarodMica/tortoise-tts.git
4 | [submodule "modules/dlas"]
5 | path = modules/dlas
6 | url = https://github.com/JarodMica/DL-Art-School.git
7 | [submodule "modules/tortoise_dataset_tools"]
8 | path = modules/tortoise_dataset_tools
9 | url = https://github.com/JarodMica/tortoise_dataset_tools.git
10 |
--------------------------------------------------------------------------------
/Dockerfile:
--------------------------------------------------------------------------------
1 | FROM nvidia/cuda:12.2.0-devel-ubuntu22.04
2 |
3 | ARG DEBIAN_FRONTEND=noninteractive
4 | ARG TZ=UTC
5 | ARG MINICONDA_VERSION=23.1.0-1
6 | ARG PYTHON_VERSION=3.11
7 | ARG UID=1000
8 | ARG GID=1000
9 |
10 | # TZ
11 | RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone
12 |
13 | # Prereqs
14 | RUN apt-get update
15 | RUN apt-get install -y \
16 | curl \
17 | wget \
18 | git \
19 | ffmpeg \
20 | p7zip-full \
21 | gcc \
22 | g++ \
23 | vim
24 |
25 | # User
26 | RUN groupadd --gid $GID user
27 | RUN useradd --no-log-init --create-home --shell /bin/bash --uid $UID --gid $GID user
28 | USER user
29 | ENV HOME=/home/user
30 | WORKDIR $HOME
31 | RUN mkdir $HOME/.cache $HOME/.config && chmod -R 777 $HOME
32 |
33 | # Python
34 | RUN wget https://repo.anaconda.com/miniconda/Miniconda3-py39_$MINICONDA_VERSION-Linux-x86_64.sh
35 | RUN chmod +x Miniconda3-py39_$MINICONDA_VERSION-Linux-x86_64.sh
36 | RUN ./Miniconda3-py39_$MINICONDA_VERSION-Linux-x86_64.sh -b -p /home/user/miniconda
37 | ENV PATH="$HOME/miniconda/bin:$PATH"
38 | RUN conda init
39 | RUN conda install python=$PYTHON_VERSION
40 | RUN python3 -m pip install --upgrade pip
41 | RUN pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
42 |
43 | # Base path
44 | RUN mkdir $HOME/ai-voice-cloning
45 | WORKDIR $HOME/ai-voice-cloning
46 |
47 | # Built in modules
48 | COPY --chown=user:user modules modules
49 | RUN python3 -m pip install -r ./modules/tortoise-tts/requirements.txt
50 | RUN python3 -m pip install -e ./modules/tortoise-tts/
51 | RUN python3 -m pip install -r ./modules/dlas/requirements.txt
52 | RUN python3 -m pip install -e ./modules/dlas/
53 |
54 | # RVC
55 | RUN \
56 | curl -L -o /tmp/rvc.zip https://huggingface.co/Jmica/rvc/resolve/main/rvc_lightweight.zip?download=true &&\
57 | 7z x /tmp/rvc.zip &&\
58 | rm -f /tmp/rvc.zip
59 | USER root
60 | RUN \
61 | chown user:user rvc -R &&\
62 | chmod -R u+rwX,go+rX,go-w rvc
63 | USER user
64 | RUN python3 -m pip install -r ./rvc/requirements.txt
65 |
66 | # Fairseq
67 | # Using patched version for Python 3.11 due to https://github.com/facebookresearch/fairseq/issues/5012
68 | RUN python3 -m pip install git+https://github.com/liyaodev/fairseq
69 |
70 | # RVC Pipeline
71 | RUN python3 -m pip install git+https://github.com/JarodMica/rvc-tts-pipeline.git@lightweight#egg=rvc_tts_pipe
72 |
73 | # Deepspeed
74 | RUN python3 -m pip install deepspeed
75 |
76 | # PyFastMP3Decoder
77 | RUN python3 -m pip install cython
78 | RUN git clone https://github.com/neonbjb/pyfastmp3decoder.git
79 | RUN \
80 | cd pyfastmp3decoder &&\
81 | git submodule update --init --recursive &&\
82 | python setup.py install &&\
83 | cd ..
84 |
85 | # WhisperX
86 | RUN python3 -m pip install git+https://github.com/m-bain/whisperx.git
87 |
88 | # Main requirements
89 | ADD requirements.txt requirements.txt
90 | RUN python3 -m pip install -r ./requirements.txt
91 |
92 | # The app
93 | ADD --chown=user:user . $HOME/ai-voice-cloning
94 |
95 | ENV IN_DOCKER=true
96 |
97 | CMD ["./start.sh"]
98 |
--------------------------------------------------------------------------------
/bin/.gitkeep:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/JarodMica/ai-voice-cloning/61e46bcbc2676e21623724e66269f48d015b2ccb/bin/.gitkeep
--------------------------------------------------------------------------------
/config/.gitignore:
--------------------------------------------------------------------------------
1 | *
2 | !.gitignore
--------------------------------------------------------------------------------
/config/rvc.json:
--------------------------------------------------------------------------------
1 | {
2 | "rvc_model": "azasu.pth",
3 | "f0_up_key": 12,
4 | "file_index": "",
5 | "index_rate": 0,
6 | "filter_radius": 3,
7 | "resample_sr": 48000,
8 | "rms_mix_rate": 0.25,
9 | "protect": 0.33
10 | }
--------------------------------------------------------------------------------
/deepspeed-0.8.3+6eca037c-cp39-cp39-win_amd64.whl:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/JarodMica/ai-voice-cloning/61e46bcbc2676e21623724e66269f48d015b2ccb/deepspeed-0.8.3+6eca037c-cp39-cp39-win_amd64.whl
--------------------------------------------------------------------------------
/download_ffmpeg.bat:
--------------------------------------------------------------------------------
1 | @echo off
2 |
3 | set ffmpeg_url=https://www.gyan.dev/ffmpeg/builds/ffmpeg-git-essentials.7z?download=true
4 | set ffmpeg_folder=ffmpeg
5 | set ffmpeg_zip=ffmpeg.7z
6 |
7 | curl -o 7z.exe "https://www.7-zip.org/a/7zr.exe"
8 |
9 | if not exist "%ffmpeg_folder%" (
10 | if not exist "%ffmpeg_zip%" (
11 | echo Downloading %ffmpeg_zip%...
12 | curl -L -o "%ffmpeg_zip%" "%ffmpeg_url%"
13 | if errorlevel 1 (
14 | echo Download failed. Please check your internet connection or the URL and try again.
15 | exit /b 1
16 | )
17 | ) else (
18 | echo File %ffmpeg_zip% already exists, skipping download.
19 | )
20 |
21 | echo Extracting %ffmpeg_zip%...
22 | 7z.exe x %ffmpeg_zip% -o%ffmpeg_folder%
23 | echo FFmpeg has finished downloading and extracting.
24 | ) else (
25 | echo FFmpeg folder %ffmpeg_folder% already exists, skipping download and extraction.
26 | )
27 |
28 | :: Move ffmpeg.exe and ffprobe.exe to the ffmpeg folder root
29 | for /D %%i in ("%ffmpeg_folder%\*") do (
30 | if exist "%%i\bin\ffmpeg.exe" move "%%i\bin\ffmpeg.exe" "ffmpeg.exe"
31 | if exist "%%i\bin\ffprobe.exe" move "%%i\bin\ffprobe.exe" "ffprobe.exe"
32 | )
33 |
34 | echo FFmpeg moved out of downloaded folder
--------------------------------------------------------------------------------
/models/.gitignore:
--------------------------------------------------------------------------------
1 | *
2 | !tokenizers/
3 | !.gitignore
4 | !rvc_models/
5 | !.template.dlas.yaml
6 | !tortoise/
--------------------------------------------------------------------------------
/models/.template.dlas.yaml:
--------------------------------------------------------------------------------
1 | name: '${voice}'
2 | model: extensibletrainer
3 | scale: 1
4 | gpu_ids: [0] # Manually edit this if the GPU you want to train on is not your primary, as this will set the env var that exposes CUDA devices
5 | start_step: 0
6 | checkpointing_enabled: true
7 | fp16: ${half_p}
8 | bitsandbytes: ${bitsandbytes}
9 | gpus: ${gpus}
10 |
11 | datasets:
12 | train:
13 | name: training
14 | n_workers: ${workers}
15 | batch_size: ${batch_size}
16 | mode: paired_voice_audio
17 | path: ${dataset_path}
18 | fetcher_mode: ['lj']
19 | phase: train
20 | max_wav_length: 255995 # ~11.6 seconds
21 | max_text_length: 200
22 | sample_rate: 22050
23 | load_conditioning: True
24 | num_conditioning_candidates: 2
25 | conditioning_length: 44000
26 | use_bpe_tokenizer: True
27 | tokenizer_vocab: ${tokenizer_json} # ./models/tortoise/bpe_lowercase_asr_256.json
28 | load_aligned_codes: False
29 | val:
30 | name: validation
31 | n_workers: ${workers}
32 | batch_size: ${validation_batch_size}
33 | mode: paired_voice_audio
34 | path: ${validation_path}
35 | fetcher_mode: ['lj']
36 | phase: val
37 | max_wav_length: 255995
38 | max_text_length: 200
39 | sample_rate: 22050
40 | load_conditioning: True
41 | num_conditioning_candidates: 2
42 | conditioning_length: 44000
43 | use_bpe_tokenizer: True
44 | tokenizer_vocab: ${tokenizer_json} # ./models/tortoise/bpe_lowercase_asr_256.json
45 | load_aligned_codes: False
46 |
47 | steps:
48 | gpt_train:
49 | training: gpt
50 | loss_log_buffer: 500
51 |
52 | # Generally follows the recipe from the DALLE paper.
53 | optimizer: ${optimizer} # this should be adamw_zero if you're using distributed training
54 | optimizer_params:
55 | lr: !!float ${learning_rate} # originally: 1e-4
56 | weight_decay: !!float 1e-2
57 | beta1: 0.9
58 | beta2: 0.96
59 | clip_grad_eps: 4
60 |
61 | injectors:
62 | paired_to_mel:
63 | type: torch_mel_spectrogram
64 | mel_norm_file: ./modules/tortoise-tts/tortoise/data/mel_norms.pth # ./models/tortoise/clips_mel_norms.pth
65 | in: wav
66 | out: paired_mel
67 | paired_cond_to_mel:
68 | type: for_each
69 | subtype: torch_mel_spectrogram
70 | mel_norm_file: ./modules/tortoise-tts/tortoise/data/mel_norms.pth # ./models/tortoise/clips_mel_norms.pth
71 | in: conditioning
72 | out: paired_conditioning_mel
73 | to_codes:
74 | type: discrete_token
75 | in: paired_mel
76 | out: paired_mel_codes
77 | dvae_config: "./models/tortoise/train_diffusion_vocoder_22k_level.yml"
78 | paired_fwd_text:
79 | type: generator
80 | generator: gpt
81 | in: [paired_conditioning_mel, padded_text, text_lengths, paired_mel_codes, wav_lengths]
82 | out: [loss_text_ce, loss_mel_ce, logits]
83 | losses:
84 | text_ce:
85 | type: direct
86 | weight: ${text_lr_weight}
87 | key: loss_text_ce
88 | mel_ce:
89 | type: direct
90 | weight: ${mel_lr_weight}
91 | key: loss_mel_ce
92 |
93 | networks:
94 | gpt:
95 | type: generator
96 | which_model_G: unified_voice2
97 | kwargs:
98 | layers: 30 # originally: 8
99 | model_dim: 1024 # originally: 512
100 | heads: 16 # originally: 8
101 | max_text_tokens: 402 # originally: 120
102 | max_mel_tokens: 604 # originally: 250
103 | max_conditioning_inputs: 2 # originally: 1
104 | mel_length_compression: 1024
105 | number_text_tokens: 256 # supposed to be 255 for newer unified_voice files
106 | number_mel_codes: 8194
107 | start_mel_token: 8192
108 | stop_mel_token: 8193
109 | start_text_token: 255
110 | train_solo_embeddings: False # missing in uv3/4
111 | use_mel_codes_as_input: True # ditto
112 | checkpointing: True
113 | tortoise_compat: True
114 | # freeze_everything_but_position_embeddings: True
115 |
116 | path:
117 | strict_load: true
118 | ${source_model}
119 | ${resume_state}
120 |
121 | train:
122 | niter: ${iterations}
123 | warmup_iter: -1
124 | mega_batch_factor: ${gradient_accumulation_size}
125 | val_freq: ${validation_rate}
126 |
127 | ema_enabled: false # I really don't think EMA matters
128 |
129 | ${learning_rate_scheme}
130 |
131 | eval:
132 | pure: ${validation_enabled}
133 | output_state: gen
134 |
135 | logger:
136 | save_checkpoint_freq: ${save_rate}
137 | visuals: [gen, mel]
138 | visual_debug_rate: ${save_rate}
139 | is_mel_spectrogram: true
--------------------------------------------------------------------------------
/models/rvc_models/.gitignore:
--------------------------------------------------------------------------------
1 | *
2 | !.gitignore
3 | !.gitkeep
--------------------------------------------------------------------------------
/models/tokenizers/ipa.json:
--------------------------------------------------------------------------------
1 | {
2 | "version": "1.0",
3 | "truncation": null,
4 | "padding": null,
5 | "normalizer": null,
6 | "pre_tokenizer": null,
7 | "post_processor": null,
8 | "decoder": null,
9 | "added_tokens": [
10 | {
11 | "id": 0,
12 | "special": true,
13 | "content": "[STOP]",
14 | "single_word": false,
15 | "lstrip": false,
16 | "rstrip": false,
17 | "normalized": false
18 | },
19 | {
20 | "id": 1,
21 | "special": true,
22 | "content": "[UNK]",
23 | "single_word": false,
24 | "lstrip": false,
25 | "rstrip": false,
26 | "normalized": false
27 | },
28 | {
29 | "id": 2,
30 | "special": true,
31 | "content": "[SPACE]",
32 | "single_word": false,
33 | "lstrip": false,
34 | "rstrip": false,
35 | "normalized": false
36 | }
37 | ],
38 | "model": {
39 | "type": "BPE",
40 | "dropout": null,
41 | "unk_token": "[UNK]",
42 | "continuing_subword_prefix": null,
43 | "end_of_word_suffix": null,
44 | "fuse_unk": false,
45 | "vocab": {
46 | "[STOP]": 0,
47 | "[UNK]": 1,
48 | "[SPACE]": 2,
49 | "!": 3,
50 | "'": 4,
51 | "(": 5,
52 | ")": 6,
53 | ",": 7,
54 | "-": 8,
55 | ".": 9,
56 | "/": 10,
57 | ":": 11,
58 | ";": 12,
59 | "?": 13,
60 | "a": 14,
61 | "b": 15,
62 | "c": 16,
63 | "d": 17,
64 | "e": 18,
65 | "f": 19,
66 | "g": 20,
67 | "h": 21,
68 | "i": 22,
69 | "j": 23,
70 | "k": 24,
71 | "l": 25,
72 | "m": 26,
73 | "n": 27,
74 | "o": 28,
75 | "p": 29,
76 | "q": 30,
77 | "r": 31,
78 | "s": 32,
79 | "t": 33,
80 | "u": 34,
81 | "v": 35,
82 | "w": 36,
83 | "x": 37,
84 | "y": 38,
85 | "z": 39,
86 | "d͡": 40,
87 | "t͡": 41,
88 | "|": 42,
89 | "æ": 43,
90 | "ð": 44,
91 | "ŋ": 45,
92 | "ɑ": 46,
93 | "ɔ": 47,
94 | "ə": 48,
95 | "ɚ": 49,
96 | "ɛ": 50,
97 | "ɡ": 51,
98 | "ɪ": 52,
99 | "ɹ": 53,
100 | "ʃ": 54,
101 | "ʊ": 55,
102 | "ʌ": 56,
103 | "ʒ": 57,
104 | "θ": 58,
105 | "ɐ": 59,
106 | "ɜ": 60,
107 | "ᵻ": 61,
108 | "ɾ": 62,
109 | "n\u0329": 63,
110 | "ː": 64,
111 | "ˈ": 65,
112 | "ˌ": 66,
113 | "ʔ": 67,
114 | "d͡ʒ": 68,
115 | "aɪ": 69,
116 | "aʊ": 70,
117 | "eɪ": 71,
118 | "oʊ": 72,
119 | "t͡ʃ": 73,
120 | "ɔɪ": 74,
121 | "ɔː": 75,
122 | "uː": 76,
123 | "iː": 77,
124 | "ɑː": 78,
125 | "oː": 79,
126 | "ɜː": 80
127 | },
128 | "merges": [
129 | "a ɪ",
130 | "e ɪ",
131 | "ɔ ɪ",
132 | "a ʊ",
133 | "o ʊ",
134 | "d͡ ʒ",
135 | "t͡ ʃ",
136 | "i ː",
137 | "o ː",
138 | "u ː",
139 | "ɑ ː",
140 | "ɔ ː",
141 | "ɜ ː"
142 | ]
143 | }
144 | }
--------------------------------------------------------------------------------
/models/tortoise/train_diffusion_vocoder_22k_level.yml:
--------------------------------------------------------------------------------
1 | path:
2 | pretrain_model_dvae: './models/tortoise/dvae.pth'
3 | strict_load: true
4 | #resume_state: ../experiments/train_imgnet_vqvae_stage1/training_state/0.state # <-- Set this to resume from a previous training state.
5 | networks:
6 | dvae:
7 | type: generator
8 | which_model_G: lucidrains_dvae
9 | kwargs:
10 | channels: 80
11 | codebook_dim: 512
12 | hidden_dim: 512
13 | kernel_size: 3
14 | num_layers: 2
15 | num_resnet_blocks: 3
16 | num_tokens: 8192
17 | positional_dims: 1
18 | use_transposed_convs: false
19 |
--------------------------------------------------------------------------------
/modules/.gitignore:
--------------------------------------------------------------------------------
1 | bitsandbytes-windows/
--------------------------------------------------------------------------------
/modules/rvc/.gitignore:
--------------------------------------------------------------------------------
1 | .DS_Store
2 | __pycache__
3 | /TEMP
4 | *.pyd
5 | .venv
6 | /opt
7 | tools/aria2c/
8 | tools/flag.txt
9 | venv/
10 |
11 | # Imported from huggingface.co/lj1995/VoiceConversionWebUI
12 | /pretrained
13 | /pretrained_v2
14 | /uvr5_weights
15 | hubert_base.pt
16 | rmvpe.onnx
17 | rmvpe.pt
18 | assets/
19 |
20 | # Generated by RVC
21 | /logs
22 | /weights
23 |
24 | # To set a Python version for the project
25 | .tool-versions
26 |
--------------------------------------------------------------------------------
/modules/rvc/Dockerfile:
--------------------------------------------------------------------------------
1 | # syntax=docker/dockerfile:1
2 |
3 | FROM nvidia/cuda:11.6.2-cudnn8-runtime-ubuntu20.04
4 |
5 | EXPOSE 7865
6 |
7 | WORKDIR /app
8 |
9 | COPY . .
10 |
11 | # Install dependenceis to add PPAs
12 | RUN apt-get update && \
13 | apt-get install -y -qq ffmpeg aria2 && apt clean && \
14 | apt-get install -y software-properties-common && \
15 | apt-get clean && \
16 | rm -rf /var/lib/apt/lists/*
17 |
18 | # Add the deadsnakes PPA to get Python 3.9
19 | RUN add-apt-repository ppa:deadsnakes/ppa
20 |
21 | # Install Python 3.9 and pip
22 | RUN apt-get update && \
23 | apt-get install -y build-essential python-dev python3-dev python3.9-distutils python3.9-dev python3.9 curl && \
24 | apt-get clean && \
25 | update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.9 1 && \
26 | curl https://bootstrap.pypa.io/get-pip.py | python3.9
27 |
28 | # Set Python 3.9 as the default
29 | RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.9 1
30 |
31 | RUN python3 -m pip install --no-cache-dir -r requirements.txt
32 |
33 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/D40k.pth -d assets/pretrained_v2/ -o D40k.pth
34 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/G40k.pth -d assets/pretrained_v2/ -o G40k.pth
35 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0D40k.pth -d assets/pretrained_v2/ -o f0D40k.pth
36 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0G40k.pth -d assets/pretrained_v2/ -o f0G40k.pth
37 |
38 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP2-人声vocals+非人声instrumentals.pth -d assets/uvr5_weights/ -o HP2-人声vocals+非人声instrumentals.pth
39 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP5-主旋律人声vocals+其他instrumentals.pth -d assets/uvr5_weights/ -o HP5-主旋律人声vocals+其他instrumentals.pth
40 |
41 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt -d assets/hubert -o hubert_base.pt
42 |
43 | RUN aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/rmvpe.pt -d assets/rmvpe -o rmvpe.pt
44 |
45 | VOLUME [ "/app/weights", "/app/opt" ]
46 |
47 | CMD ["python3", "infer-web.py"]
48 |
--------------------------------------------------------------------------------
/modules/rvc/LICENSE:
--------------------------------------------------------------------------------
1 | MIT License
2 |
3 | Copyright (c) 2023 liujing04
4 | Copyright (c) 2023 源文雨
5 | Copyright (c) 2023 Ftps
6 |
7 | Permission is hereby granted, free of charge, to any person obtaining a copy
8 | of this software and associated documentation files (the "Software"), to deal
9 | in the Software without restriction, including without limitation the rights
10 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 | copies of the Software, and to permit persons to whom the Software is
12 | furnished to do so, subject to the following conditions:
13 |
14 | The above copyright notice and this permission notice shall be included in all
15 | copies or substantial portions of the Software.
16 |
17 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
20 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 | SOFTWARE.
24 |
--------------------------------------------------------------------------------
/modules/rvc/MIT协议暨相关引用库协议:
--------------------------------------------------------------------------------
1 | 本软件及其相关代码以MIT协议开源，作者不对软件具备任何控制力，使用软件者、传播软件导出的声音者自负全责。
2 | 如不认可该条款，则不能使用或引用软件包内任何代码和文件。
3 |
4 | 特此授予任何获得本软件和相关文档文件（以下简称“软件”）副本的人免费使用、复制、修改、合并、出版、分发、再授权和/或销售本软件的权利，以及授予本软件所提供的人使用本软件的权利，但须符合以下条件：
5 | 上述版权声明和本许可声明应包含在软件的所有副本或实质部分中。
6 | 软件是“按原样”提供的，没有任何明示或暗示的保证，包括但不限于适销性、适用于特定目的和不侵权的保证。在任何情况下，作者或版权持有人均不承担因软件或软件的使用或其他交易而产生、产生或与之相关的任何索赔、损害赔偿或其他责任，无论是在合同诉讼、侵权诉讼还是其他诉讼中。
7 |
8 |
9 | The LICENCEs for related libraries are as follows.
10 | 相关引用库协议如下：
11 |
12 | ContentVec
13 | https://github.com/auspicious3000/contentvec/blob/main/LICENSE
14 | MIT License
15 |
16 | VITS
17 | https://github.com/jaywalnut310/vits/blob/main/LICENSE
18 | MIT License
19 |
20 | HIFIGAN
21 | https://github.com/jik876/hifi-gan/blob/master/LICENSE
22 | MIT License
23 |
24 | gradio
25 | https://github.com/gradio-app/gradio/blob/main/LICENSE
26 | Apache License 2.0
27 |
28 | ffmpeg
29 | https://github.com/FFmpeg/FFmpeg/blob/master/COPYING.LGPLv3
30 | https://github.com/BtbN/FFmpeg-Builds/releases/download/autobuild-2021-02-28-12-32/ffmpeg-n4.3.2-160-gfbb9368226-win64-lgpl-4.3.zip
31 | LPGLv3 License
32 | MIT License
33 |
34 | ultimatevocalremovergui
35 | https://github.com/Anjok07/ultimatevocalremovergui/blob/master/LICENSE
36 | https://github.com/yang123qwe/vocal_separation_by_uvr5
37 | MIT License
38 |
39 | audio-slicer
40 | https://github.com/openvpi/audio-slicer/blob/main/LICENSE
41 | MIT License
42 |
43 | PySimpleGUI
44 | https://github.com/PySimpleGUI/PySimpleGUI/blob/master/license.txt
45 | LPGLv3 License
46 |
--------------------------------------------------------------------------------
/modules/rvc/README.md:
--------------------------------------------------------------------------------
1 | # Notes
2 | This is a fork of RVC and may not contain the latest features. It is mainly used in my projects so that I have an "rvc package" that is useable for functions that I need.
3 | - I removed fairseq, so this needs to be manually installed or use the prebuilt wheels file
4 |
5 |
6 |
7 |
Retrieval-based-Voice-Conversion-WebUI
8 | 一个基于VITS的简单易用的语音转换（变声器）框架
9 |
10 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
12 |
13 |
14 |
15 | [](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI.ipynb)
16 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
17 | [](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
18 |
19 | [](https://discord.gg/HcsmBBGyVk)
20 |
21 | [**更新日志**](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/docs/Changelog_CN.md) | [**常见问题解答**](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%B8%B8%E8%A7%81%E9%97%AE%E9%A2%98%E8%A7%A3%E7%AD%94) | [**AutoDL·5毛钱训练AI歌手**](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B) | [**对照实验记录**](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%AF%B9%E7%85%A7%E5%AE%9E%E9%AA%8C%C2%B7%E5%AE%9E%E9%AA%8C%E8%AE%B0%E5%BD%95)) | [**在线演示**](https://huggingface.co/spaces/Ricecake123/RVC-demo)
22 |
23 |
148 |
149 |
--------------------------------------------------------------------------------
/modules/rvc/changelog.md:
--------------------------------------------------------------------------------
1 | # Changelog
2 |
3 | ## 10/6/2023
4 | I removed fairseq from requirements.txt as if you don't have microsoft build tools for c++ 2019, it will fail to install packages. This can be remedied by either installing the build tools, or using a prebuilt wheels (which is the approach I'm going to use for my applications that I distribute)
5 | - Added weights_only=True to all torch.load calls to help mitigate possible security issues with pickle.
6 | - I have not tested to see if the package will still work for training with this new parameter, but I will have to test this at a later time when I get to that fuctionality.
--------------------------------------------------------------------------------
/modules/rvc/configs/config.json:
--------------------------------------------------------------------------------
1 | {
2 | "pth_path": "assets/weights/kikiV1.pth",
3 | "index_path": "logs/kikiV1.index",
4 | "sg_input_device": "VoiceMeeter Output (VB-Audio Vo (MME)",
5 | "sg_output_device": "VoiceMeeter Aux Input (VB-Audio (MME)",
6 | "threhold": -45.0,
7 | "pitch": 12.0,
8 | "index_rate": 0.0,
9 | "rms_mix_rate": 0.0,
10 | "block_time": 0.25,
11 | "crossfade_length": 0.04,
12 | "extra_time": 2.0,
13 | "n_cpu": 6.0,
14 | "f0method": "rmvpe"
15 | }
16 |
--------------------------------------------------------------------------------
/modules/rvc/configs/v1/32k.json:
--------------------------------------------------------------------------------
1 | {
2 | "train": {
3 | "log_interval": 200,
4 | "seed": 1234,
5 | "epochs": 20000,
6 | "learning_rate": 1e-4,
7 | "betas": [0.8, 0.99],
8 | "eps": 1e-9,
9 | "batch_size": 4,
10 | "fp16_run": true,
11 | "lr_decay": 0.999875,
12 | "segment_size": 12800,
13 | "init_lr_ratio": 1,
14 | "warmup_epochs": 0,
15 | "c_mel": 45,
16 | "c_kl": 1.0
17 | },
18 | "data": {
19 | "max_wav_value": 32768.0,
20 | "sampling_rate": 32000,
21 | "filter_length": 1024,
22 | "hop_length": 320,
23 | "win_length": 1024,
24 | "n_mel_channels": 80,
25 | "mel_fmin": 0.0,
26 | "mel_fmax": null
27 | },
28 | "model": {
29 | "inter_channels": 192,
30 | "hidden_channels": 192,
31 | "filter_channels": 768,
32 | "n_heads": 2,
33 | "n_layers": 6,
34 | "kernel_size": 3,
35 | "p_dropout": 0,
36 | "resblock": "1",
37 | "resblock_kernel_sizes": [3,7,11],
38 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
39 | "upsample_rates": [10,4,2,2,2],
40 | "upsample_initial_channel": 512,
41 | "upsample_kernel_sizes": [16,16,4,4,4],
42 | "use_spectral_norm": false,
43 | "gin_channels": 256,
44 | "spk_embed_dim": 109
45 | }
46 | }
47 |
--------------------------------------------------------------------------------
/modules/rvc/configs/v1/40k.json:
--------------------------------------------------------------------------------
1 | {
2 | "train": {
3 | "log_interval": 200,
4 | "seed": 1234,
5 | "epochs": 20000,
6 | "learning_rate": 1e-4,
7 | "betas": [0.8, 0.99],
8 | "eps": 1e-9,
9 | "batch_size": 4,
10 | "fp16_run": true,
11 | "lr_decay": 0.999875,
12 | "segment_size": 12800,
13 | "init_lr_ratio": 1,
14 | "warmup_epochs": 0,
15 | "c_mel": 45,
16 | "c_kl": 1.0
17 | },
18 | "data": {
19 | "max_wav_value": 32768.0,
20 | "sampling_rate": 40000,
21 | "filter_length": 2048,
22 | "hop_length": 400,
23 | "win_length": 2048,
24 | "n_mel_channels": 125,
25 | "mel_fmin": 0.0,
26 | "mel_fmax": null
27 | },
28 | "model": {
29 | "inter_channels": 192,
30 | "hidden_channels": 192,
31 | "filter_channels": 768,
32 | "n_heads": 2,
33 | "n_layers": 6,
34 | "kernel_size": 3,
35 | "p_dropout": 0,
36 | "resblock": "1",
37 | "resblock_kernel_sizes": [3,7,11],
38 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
39 | "upsample_rates": [10,10,2,2],
40 | "upsample_initial_channel": 512,
41 | "upsample_kernel_sizes": [16,16,4,4],
42 | "use_spectral_norm": false,
43 | "gin_channels": 256,
44 | "spk_embed_dim": 109
45 | }
46 | }
47 |
--------------------------------------------------------------------------------
/modules/rvc/configs/v1/48k.json:
--------------------------------------------------------------------------------
1 | {
2 | "train": {
3 | "log_interval": 200,
4 | "seed": 1234,
5 | "epochs": 20000,
6 | "learning_rate": 1e-4,
7 | "betas": [0.8, 0.99],
8 | "eps": 1e-9,
9 | "batch_size": 4,
10 | "fp16_run": true,
11 | "lr_decay": 0.999875,
12 | "segment_size": 11520,
13 | "init_lr_ratio": 1,
14 | "warmup_epochs": 0,
15 | "c_mel": 45,
16 | "c_kl": 1.0
17 | },
18 | "data": {
19 | "max_wav_value": 32768.0,
20 | "sampling_rate": 48000,
21 | "filter_length": 2048,
22 | "hop_length": 480,
23 | "win_length": 2048,
24 | "n_mel_channels": 128,
25 | "mel_fmin": 0.0,
26 | "mel_fmax": null
27 | },
28 | "model": {
29 | "inter_channels": 192,
30 | "hidden_channels": 192,
31 | "filter_channels": 768,
32 | "n_heads": 2,
33 | "n_layers": 6,
34 | "kernel_size": 3,
35 | "p_dropout": 0,
36 | "resblock": "1",
37 | "resblock_kernel_sizes": [3,7,11],
38 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
39 | "upsample_rates": [10,6,2,2,2],
40 | "upsample_initial_channel": 512,
41 | "upsample_kernel_sizes": [16,16,4,4,4],
42 | "use_spectral_norm": false,
43 | "gin_channels": 256,
44 | "spk_embed_dim": 109
45 | }
46 | }
47 |
--------------------------------------------------------------------------------
/modules/rvc/configs/v2/32k.json:
--------------------------------------------------------------------------------
1 | {
2 | "train": {
3 | "log_interval": 200,
4 | "seed": 1234,
5 | "epochs": 20000,
6 | "learning_rate": 1e-4,
7 | "betas": [0.8, 0.99],
8 | "eps": 1e-9,
9 | "batch_size": 4,
10 | "fp16_run": true,
11 | "lr_decay": 0.999875,
12 | "segment_size": 12800,
13 | "init_lr_ratio": 1,
14 | "warmup_epochs": 0,
15 | "c_mel": 45,
16 | "c_kl": 1.0
17 | },
18 | "data": {
19 | "max_wav_value": 32768.0,
20 | "sampling_rate": 32000,
21 | "filter_length": 1024,
22 | "hop_length": 320,
23 | "win_length": 1024,
24 | "n_mel_channels": 80,
25 | "mel_fmin": 0.0,
26 | "mel_fmax": null
27 | },
28 | "model": {
29 | "inter_channels": 192,
30 | "hidden_channels": 192,
31 | "filter_channels": 768,
32 | "n_heads": 2,
33 | "n_layers": 6,
34 | "kernel_size": 3,
35 | "p_dropout": 0,
36 | "resblock": "1",
37 | "resblock_kernel_sizes": [3,7,11],
38 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
39 | "upsample_rates": [10,8,2,2],
40 | "upsample_initial_channel": 512,
41 | "upsample_kernel_sizes": [20,16,4,4],
42 | "use_spectral_norm": false,
43 | "gin_channels": 256,
44 | "spk_embed_dim": 109
45 | }
46 | }
47 |
--------------------------------------------------------------------------------
/modules/rvc/configs/v2/48k.json:
--------------------------------------------------------------------------------
1 | {
2 | "train": {
3 | "log_interval": 200,
4 | "seed": 1234,
5 | "epochs": 20000,
6 | "learning_rate": 1e-4,
7 | "betas": [0.8, 0.99],
8 | "eps": 1e-9,
9 | "batch_size": 4,
10 | "fp16_run": true,
11 | "lr_decay": 0.999875,
12 | "segment_size": 17280,
13 | "init_lr_ratio": 1,
14 | "warmup_epochs": 0,
15 | "c_mel": 45,
16 | "c_kl": 1.0
17 | },
18 | "data": {
19 | "max_wav_value": 32768.0,
20 | "sampling_rate": 48000,
21 | "filter_length": 2048,
22 | "hop_length": 480,
23 | "win_length": 2048,
24 | "n_mel_channels": 128,
25 | "mel_fmin": 0.0,
26 | "mel_fmax": null
27 | },
28 | "model": {
29 | "inter_channels": 192,
30 | "hidden_channels": 192,
31 | "filter_channels": 768,
32 | "n_heads": 2,
33 | "n_layers": 6,
34 | "kernel_size": 3,
35 | "p_dropout": 0,
36 | "resblock": "1",
37 | "resblock_kernel_sizes": [3,7,11],
38 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
39 | "upsample_rates": [12,10,2,2],
40 | "upsample_initial_channel": 512,
41 | "upsample_kernel_sizes": [24,20,4,4],
42 | "use_spectral_norm": false,
43 | "gin_channels": 256,
44 | "spk_embed_dim": 109
45 | }
46 | }
47 |
--------------------------------------------------------------------------------
/modules/rvc/docker-compose.yml:
--------------------------------------------------------------------------------
1 | version: "3.8"
2 | services:
3 | rvc:
4 | build:
5 | context: .
6 | dockerfile: Dockerfile
7 | container_name: rvc
8 | volumes:
9 | - ./weights:/app/assets/weights
10 | - ./opt:/app/opt
11 | # - ./dataset:/app/dataset # you can use this folder in order to provide your dataset for model training
12 | ports:
13 | - 7865:7865
14 | deploy:
15 | resources:
16 | reservations:
17 | devices:
18 | - driver: nvidia
19 | count: 1
20 | capabilities: [gpu]
--------------------------------------------------------------------------------
/modules/rvc/docs/cn/Changelog_CN.md:
--------------------------------------------------------------------------------
1 | ### 20230813更新
2 | 1-常规bug修复
3 | - 保存频率总轮数最低改为1 总轮数最低改为2
4 | - 修复无pretrain模型训练报错
5 | - 增加伴奏人声分离完毕清理显存
6 | - faiss保存路径绝对路径改为相对路径
7 | - 支持路径包含空格（训练集路径+实验名称均支持，不再会报错）
8 | - filelist取消强制utf8编码
9 | - 解决实时变声中开启索引导致的CPU极大占用问题
10 |
11 | 2-重点更新
12 | - 训练出当前最强开源人声音高提取模型RMVPE，并用于RVC的训练、离线/实时推理，支持pytorch/onnx/DirectML
13 | - 通过pytorch-dml支持A卡和I卡的
14 | （1）实时变声（2）推理（3）人声伴奏分离（4）训练暂未支持，会切换至CPU训练；通过onnx_dml支持rmvpe_gpu的推理
15 |
16 | ### 20230618更新
17 | - v2增加32k和48k两个新预训练模型
18 | - 修复非f0模型推理报错
19 | - 对于超过一小时的训练集的索引建立环节，自动kmeans缩小特征处理以加速索引训练、加入和查询
20 | - 附送一个人声转吉他玩具仓库
21 | - 数据处理剔除异常值切片
22 | - onnx导出选项卡
23 |
24 | 失败的实验：
25 | - ~~特征检索增加时序维度：寄，没啥效果~~
26 | - ~~特征检索增加PCAR降维可选项：寄，数据大用kmeans缩小数据量，数据小降维操作耗时比省下的匹配耗时还多~~
27 | - ~~支持onnx推理（附带仅推理的小压缩包）：寄，生成nsf还是需要pytorch~~
28 | - ~~训练时在音高、gender、eq、噪声等方面对输入进行随机增强：寄，没啥效果~~
29 | - ~~接入小型声码器调研:寄，效果变差~~
30 |
31 | todolist：
32 | - ~~训练集音高识别支持crepe：已经被RMVPE取代，不需要~~
33 | - ~~多进程harvest推理：已经被RMVPE取代，不需要~~
34 | - ~~crepe的精度支持和RVC-config同步：已经被RMVPE取代，不需要。支持这个还要同步torchcrepe的库，麻烦~~
35 | - 对接F0编辑器
36 |
37 |
38 | ### 20230528更新
39 | - 增加v2的jupyter notebook，韩文changelog，增加一些环境依赖
40 | - 增加呼吸、清辅音、齿音保护模式
41 | - 支持crepe-full推理
42 | - UVR5人声伴奏分离加上3个去延迟模型和MDX-Net去混响模型，增加HP3人声提取模型
43 | - 索引名称增加版本和实验名称
44 | - 人声伴奏分离、推理批量导出增加音频导出格式选项
45 | - 废弃32k模型的训练
46 |
47 | ### 20230513更新
48 | - 清除一键包内部老版本runtime内残留的lib.infer_pack和uvr5_pack
49 | - 修复训练集预处理伪多进程的bug
50 | - 增加harvest识别音高可选通过中值滤波削弱哑音现象，可调整中值滤波半径
51 | - 导出音频增加后处理重采样
52 | - 训练n_cpu进程数从"仅调整f0提取"改为"调整数据预处理和f0提取"
53 | - 自动检测logs文件夹下的index路径，提供下拉列表功能
54 | - tab页增加"常见问题解答"（也可参考github-rvc-wiki）
55 | - 相同路径的输入音频推理增加了音高缓存（用途：使用harvest音高提取，整个pipeline会经历漫长且重复的音高提取过程，如果不使用缓存，实验不同音色、索引、音高中值滤波半径参数的用户在第一次测试后的等待结果会非常痛苦）
56 |
57 | ### 20230514更新
58 | - 音量包络对齐输入混合（可以缓解“输入静音输出小幅度噪声”的问题。如果输入音频背景底噪大则不建议开启，默认不开启（值为1可视为不开启））
59 | - 支持按照指定频率保存提取的小模型（假如你想尝试不同epoch下的推理效果，但是不想保存所有大checkpoint并且每次都要ckpt手工处理提取小模型，这项功能会非常实用）
60 | - 通过设置环境变量解决服务端开了系统全局代理导致浏览器连接错误的问题
61 | - 支持v2预训练模型（目前只公开了40k版本进行测试，另外2个采样率还没有训练完全）
62 | - 推理前限制超过1的过大音量
63 | - 微调数据预处理参数
64 |
65 |
66 | ### 20230409更新
67 | - 修正训练参数，提升显卡平均利用率，A100最高从25%提升至90%左右，V100:50%->90%左右，2060S:60%->85%左右，P40:25%->95%左右，训练速度显著提升
68 | - 修正参数：总batch_size改为每张卡的batch_size
69 | - 修正total_epoch：最大限制100解锁至1000；默认10提升至默认20
70 | - 修复ckpt提取识别是否带音高错误导致推理异常的问题
71 | - 修复分布式训练每个rank都保存一次ckpt的问题
72 | - 特征提取进行nan特征过滤
73 | - 修复静音输入输出随机辅音or噪声的问题（老版模型需要重做训练集重训）
74 |
75 | ### 20230416更新
76 | - 新增本地实时变声迷你GUI，双击go-realtime-gui.bat启动
77 | - 训练推理均对<50Hz的频段进行滤波过滤
78 | - 训练推理音高提取pyworld最低音高从默认80下降至50,50-80hz间的男声低音不会哑
79 | - WebUI支持根据系统区域变更语言（现支持en_US，ja_JP，zh_CN，zh_HK，zh_SG，zh_TW，不支持的默认en_US）
80 | - 修正部分显卡识别（例如V100-16G识别失败，P4识别失败）
81 |
82 | ### 20230428更新
83 | - 升级faiss索引设置，速度更快，质量更高
84 | - 取消total_npy依赖，后续分享模型不再需要填写total_npy
85 | - 解锁16系限制。4G显存GPU给到4G的推理设置。
86 | - 修复部分音频格式下UVR5人声伴奏分离的bug
87 | - 实时变声迷你gui增加对非40k与不懈怠音高模型的支持
88 |
89 | ### 后续计划：
90 | 功能：
91 | - 支持多人训练选项卡（至多4人）
92 |
93 | 底模：
94 | - 收集呼吸wav加入训练集修正呼吸变声电音的问题
95 | - 我们正在训练增加了歌声训练集的底模，未来会公开
96 |
97 |
--------------------------------------------------------------------------------
/modules/rvc/docs/cn/faq.md:
--------------------------------------------------------------------------------
1 | ## Q1:ffmpeg error/utf8 error.
2 |
3 | 大概率不是ffmpeg问题，而是音频路径问题；
4 | ffmpeg读取路径带空格、()等特殊符号，可能出现ffmpeg error；训练集音频带中文路径，在写入filelist.txt的时候可能出现utf8 error；
5 |
6 | ## Q2:一键训练结束没有索引
7 |
8 | 显示"Training is done. The program is closed."则模型训练成功，后续紧邻的报错是假的；
9 |
10 | 一键训练结束完成没有added开头的索引文件，可能是因为训练集太大卡住了添加索引的步骤；已通过批处理add索引解决内存add索引对内存需求过大的问题。临时可尝试再次点击"训练索引"按钮。
11 |
12 | ## Q3:训练结束推理没看到训练集的音色
13 | 点刷新音色再看看，如果还没有看看训练有没有报错，控制台和webui的截图，logs/实验名下的log，都可以发给开发者看看。
14 |
15 | ## Q4:如何分享模型
16 | rvc_root/logs/实验名 下面存储的pth不是用来分享模型用来推理的，而是为了存储实验状态供复现，以及继续训练用的。用来分享的模型应该是weights文件夹下大小为60+MB的pth文件；
17 | 后续将把weights/exp_name.pth和logs/exp_name/added_xxx.index合并打包成weights/exp_name.zip省去填写index的步骤，那么zip文件用来分享，不要分享pth文件，除非是想换机器继续训练；
18 | 如果你把logs文件夹下的几百MB的pth文件复制/分享到weights文件夹下强行用于推理，可能会出现f0，tgt_sr等各种key不存在的报错。你需要用ckpt选项卡最下面，手工或自动（本地logs下如果能找到相关信息则会自动）选择是否携带音高、目标音频采样率的选项后进行ckpt小模型提取（输入路径填G开头的那个），提取完在weights文件夹下会出现60+MB的pth文件，刷新音色后可以选择使用。
19 |
20 | ## Q5:Connection Error.
21 | 也许你关闭了控制台（黑色窗口）。
22 |
23 | ## Q6:WebUI弹出Expecting value: line 1 column 1 (char 0).
24 | 请关闭系统局域网代理/全局代理。
25 |
26 | 这个不仅是客户端的代理，也包括服务端的代理（例如你使用autodl设置了http_proxy和https_proxy学术加速，使用时也需要unset关掉）
27 |
28 | ## Q7:不用WebUI如何通过命令训练推理
29 | 训练脚本：
30 | 可先跑通WebUI，消息窗内会显示数据集处理和训练用命令行；
31 |
32 | 推理脚本：
33 | https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/mymodules.rvc.infer.py
34 |
35 | 例子：
36 |
37 | runtime\python.exe mymodules.rvc.infer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
38 |
39 | f0up_key=sys.argv[1]
40 | input_path=sys.argv[2]
41 | index_path=sys.argv[3]
42 | f0method=sys.argv[4]#harvest or pm
43 | opt_path=sys.argv[5]
44 | model_path=sys.argv[6]
45 | index_rate=float(sys.argv[7])
46 | device=sys.argv[8]
47 | is_half=bool(sys.argv[9])
48 |
49 | ## Q8:Cuda error/Cuda out of memory.
50 | 小概率是cuda配置问题、设备不支持；大概率是显存不够（out of memory）；
51 |
52 | 训练的话缩小batch size（如果缩小到1还不够只能更换显卡训练），推理的话酌情缩小config.py结尾的x_pad，x_query，x_center，x_max。4G以下显存（例如1060（3G）和各种2G显卡）可以直接放弃，4G显存显卡还有救。
53 |
54 | ## Q9:total_epoch调多少比较好
55 |
56 | 如果训练集音质差底噪大，20~30足够了，调太高，底模音质无法带高你的低音质训练集
57 | 如果训练集音质高底噪低时长多，可以调高，200是ok的（训练速度很快，既然你有条件准备高音质训练集，显卡想必条件也不错，肯定不在乎多一些训练时间）
58 |
59 | ## Q10:需要多少训练集时长
60 | 推荐10min至50min
61 | 保证音质高底噪低的情况下，如果有个人特色的音色统一，则多多益善
62 | 高水平的训练集（精简+音色有特色），5min至10min也是ok的，仓库作者本人就经常这么玩
63 | 也有人拿1min至2min的数据来训练并且训练成功的，但是成功经验是其他人不可复现的，不太具备参考价值。这要求训练集音色特色非常明显（比如说高频气声较明显的萝莉少女音），且音质高；
64 | 1min以下时长数据目前没见有人尝试（成功）过。不建议进行这种鬼畜行为。
65 |
66 | ## Q11:index rate干嘛用的，怎么调（科普）
67 | 如果底模和推理源的音质高于训练集的音质，他们可以带高推理结果的音质，但代价可能是音色往底模/推理源的音色靠，这种现象叫做"音色泄露"；
68 | index rate用来削减/解决音色泄露问题。调到1，则理论上不存在推理源的音色泄露问题，但音质更倾向于训练集。如果训练集音质比推理源低，则index rate调高可能降低音质。调到0，则不具备利用检索混合来保护训练集音色的效果；
69 | 如果训练集优质时长多，可调高total_epoch，此时模型本身不太会引用推理源和底模的音色，很少存在"音色泄露"问题，此时index_rate不重要，你甚至可以不建立/分享index索引文件。
70 |
71 | ## Q11:推理怎么选gpu
72 | config.py文件里device cuda:后面选择卡号；
73 | 卡号和显卡的映射关系，在训练选项卡的显卡信息栏里能看到。
74 |
75 | ## Q12:如何推理训练中间保存的pth
76 | 通过ckpt选项卡最下面提取小模型。
77 |
78 |
79 | ## Q13:如何中断和继续训练
80 | 现阶段只能关闭WebUI控制台双击go-web.bat重启程序。网页参数也要刷新重新填写；
81 | 继续训练：相同网页参数点训练模型，就会接着上次的checkpoint继续训练。
82 |
83 | ## Q14:训练时出现文件页面/内存error
84 | 进程开太多了，内存炸了。你可能可以通过如下方式解决
85 | 1、"提取音高和处理数据使用的CPU进程数" 酌情拉低；
86 | 2、训练集音频手工切一下，不要太长。
87 |
88 |
89 | ## Q15:如何中途加数据训练
90 | 1、所有数据新建一个实验名；
91 | 2、拷贝上一次的最新的那个G和D文件（或者你想基于哪个中间ckpt训练，也可以拷贝中间的）到新实验名；下
92 | 3、一键训练新实验名，他会继续上一次的最新进度训练。
93 |
94 | ## Q16: error about llvmlite.dll
95 |
96 | OSError: Could not load shared object file: llvmlite.dll
97 |
98 | FileNotFoundError: Could not find module lib\site-packages\llvmlite\binding\llvmlite.dll (or one of its dependencies). Try using the full path with constructor syntax.
99 |
100 | win平台会报这个错，装上https://aka.ms/vs/17/release/vc_redist.x64.exe这个再重启WebUI就好了。
101 |
102 | ## Q17: RuntimeError: The expanded size of the tensor (17280) must match the existing size (0) at non-singleton dimension 1. Target sizes: [1, 17280]. Tensor sizes: [0]
103 |
104 | wavs16k文件夹下，找到文件大小显著比其他都小的一些音频文件，删掉，点击训练模型，就不会报错了，不过由于一键流程中断了你训练完模型还要点训练索引。
105 |
106 | ## Q18: RuntimeError: The size of tensor a (24) must match the size of tensor b (16) at non-singleton dimension 2
107 |
108 | 不要中途变更采样率继续训练。如果一定要变更，应更换实验名从头训练。当然你也可以把上次提取的音高和特征（0/1/2/2b folders）拷贝过去加速训练流程。
109 |
--------------------------------------------------------------------------------
/modules/rvc/docs/en/faiss_tips_en.md:
--------------------------------------------------------------------------------
1 | faiss tuning TIPS
2 | ==================
3 | # about faiss
4 | faiss is a library of neighborhood searches for dense vectors, developed by facebook research, which efficiently implements many approximate neighborhood search methods.
5 | Approximate Neighbor Search finds similar vectors quickly while sacrificing some accuracy.
6 |
7 | ## faiss in RVC
8 | In RVC, for the embedding of features converted by HuBERT, we search for embeddings similar to the embedding generated from the training data and mix them to achieve a conversion that is closer to the original speech. However, since this search takes time if performed naively, high-speed conversion is realized by using approximate neighborhood search.
9 |
10 | # implementation overview
11 | In '/logs/your-experiment/3_feature256' where the model is located, features extracted by HuBERT from each voice data are located.
12 | From here we read the npy files in order sorted by filename and concatenate the vectors to create big_npy. (This vector has shape [N, 256].)
13 | After saving big_npy as /logs/your-experiment/total_fea.npy, train it with faiss.
14 |
15 | In this article, I will explain the meaning of these parameters.
16 |
17 | # Explanation of the method
18 | ## index factory
19 | An index factory is a unique faiss notation that expresses a pipeline that connects multiple approximate neighborhood search methods as a string.
20 | This allows you to try various approximate neighborhood search methods simply by changing the index factory string.
21 | In RVC it is used like this:
22 |
23 | ```python
24 | index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
25 | ```
26 | Among the arguments of index_factory, the first is the number of dimensions of the vector, the second is the index factory string, and the third is the distance to use.
27 |
28 | For more detailed notation
29 | https://github.com/facebookresearch/faiss/wiki/The-index-factory
30 |
31 | ## index for distance
32 | There are two typical indexes used as similarity of embedding as follows.
33 |
34 | - Euclidean distance (METRIC_L2)
35 | - inner product (METRIC_INNER_PRODUCT)
36 |
37 | Euclidean distance takes the squared difference in each dimension, sums the differences in all dimensions, and then takes the square root. This is the same as the distance in 2D and 3D that we use on a daily basis.
38 | The inner product is not used as an index of similarity as it is, and the cosine similarity that takes the inner product after being normalized by the L2 norm is generally used.
39 |
40 | Which is better depends on the case, but cosine similarity is often used in embedding obtained by word2vec and similar image retrieval models learned by ArcFace. If you want to do l2 normalization on vector X with numpy, you can do it with the following code with eps small enough to avoid 0 division.
41 |
42 | ```python
43 | X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
44 | ```
45 |
46 | Also, for the index factory, you can change the distance index used for calculation by choosing the value to pass as the third argument.
47 |
48 | ```python
49 | index = faiss.index_factory(dimention, text, faiss.METRIC_INNER_PRODUCT)
50 | ```
51 |
52 | ## IVF
53 | IVF (Inverted file indexes) is an algorithm similar to the inverted index in full-text search.
54 | During learning, the search target is clustered with kmeans, and Voronoi partitioning is performed using the cluster center. Each data point is assigned a cluster, so we create a dictionary that looks up the data points from the clusters.
55 |
56 | For example, if clusters are assigned as follows
57 | |index|Cluster|
58 | |-----|-------|
59 | |1|A|
60 | |2|B|
61 | |3|A|
62 | |4|C|
63 | |5|B|
64 |
65 | The resulting inverted index looks like this:
66 |
67 | |cluster|index|
68 | |-------|-----|
69 | |A|1, 3|
70 | |B|2, 5|
71 | |C|4|
72 |
73 | When searching, we first search n_probe clusters from the clusters, and then calculate the distances for the data points belonging to each cluster.
74 |
75 | # recommend parameter
76 | There are official guidelines on how to choose an index, so I will explain accordingly.
77 | https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
78 |
79 | For datasets below 1M, 4bit-PQ is the most efficient method available in faiss as of April 2023.
80 | Combining this with IVF, narrowing down the candidates with 4bit-PQ, and finally recalculating the distance with an accurate index can be described by using the following index factory.
81 |
82 | ```python
83 | index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
84 | ```
85 |
86 | ## Recommended parameters for IVF
87 | Consider the case of too many IVFs. For example, if coarse quantization by IVF is performed for the number of data, this is the same as a naive exhaustive search and is inefficient.
88 | For 1M or less, IVF values are recommended between 4*sqrt(N) ~ 16*sqrt(N) for N number of data points.
89 |
90 | Since the calculation time increases in proportion to the number of n_probes, please consult with the accuracy and choose appropriately. Personally, I don't think RVC needs that much accuracy, so n_probe = 1 is fine.
91 |
92 | ## FastScan
93 | FastScan is a method that enables high-speed approximation of distances by Cartesian product quantization by performing them in registers.
94 | Cartesian product quantization performs clustering independently for each d dimension (usually d = 2) during learning, calculates the distance between clusters in advance, and creates a lookup table. At the time of prediction, the distance of each dimension can be calculated in O(1) by looking at the lookup table.
95 | So the number you specify after PQ usually specifies half the dimension of the vector.
96 |
97 | For a more detailed description of FastScan, please refer to the official documentation.
98 | https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
99 |
100 | ## RFlat
101 | RFlat is an instruction to recalculate the rough distance calculated by FastScan with the exact distance specified by the third argument of index factory.
102 | When getting k neighbors, k*k_factor points are recalculated.
103 |
--------------------------------------------------------------------------------
/modules/rvc/docs/en/training_tips_en.md:
--------------------------------------------------------------------------------
1 | Instructions and tips for RVC training
2 | ======================================
3 | This TIPS explains how data training is done.
4 |
5 | # Training flow
6 | I will explain along the steps in the training tab of the GUI.
7 |
8 | ## step1
9 | Set the experiment name here.
10 |
11 | You can also set here whether the model should take pitch into account.
12 | If the model doesn't consider pitch, the model will be lighter, but not suitable for singing.
13 |
14 | Data for each experiment is placed in `/logs/your-experiment-name/`.
15 |
16 | ## step2a
17 | Loads and preprocesses audio.
18 |
19 | ### load audio
20 | If you specify a folder with audio, the audio files in that folder will be read automatically.
21 | For example, if you specify `C:Users\hoge\voices`, `C:Users\hoge\voices\voice.mp3` will be loaded, but `C:Users\hoge\voices\dir\voice.mp3` will Not loaded.
22 |
23 | Since ffmpeg is used internally for reading audio, if the extension is supported by ffmpeg, it will be read automatically.
24 | After converting to int16 with ffmpeg, convert to float32 and normalize between -1 to 1.
25 |
26 | ### denoising
27 | The audio is smoothed by scipy's filtfilt.
28 |
29 | ### Audio Split
30 | First, the input audio is divided by detecting parts of silence that last longer than a certain period (max_sil_kept=5 seconds?). After splitting the audio on silence, split the audio every 4 seconds with an overlap of 0.3 seconds. For audio separated within 4 seconds, after normalizing the volume, convert the wav file to `/logs/your-experiment-name/0_gt_wavs` and then convert it to 16k sampling rate to `/logs/your-experiment-name/1_16k_wavs ` as a wav file.
31 |
32 | ## step2b
33 | ### Extract pitch
34 | Extract pitch information from wav files. Extract the pitch information (=f0) using the method built into parselmouth or pyworld and save it in `/logs/your-experiment-name/2a_f0`. Then logarithmically convert the pitch information to an integer between 1 and 255 and save it in `/logs/your-experiment-name/2b-f0nsf`.
35 |
36 | ### Extract feature_print
37 | Convert the wav file to embedding in advance using HuBERT. Read the wav file saved in `/logs/your-experiment-name/1_16k_wavs`, convert the wav file to 256-dimensional features with HuBERT, and save in npy format in `/logs/your-experiment-name/3_feature256`.
38 |
39 | ## step3
40 | train the model.
41 | ### Glossary for Beginners
42 | In deep learning, the data set is divided and the learning proceeds little by little. In one model update (step), batch_size data are retrieved and predictions and error corrections are performed. Doing this once for a dataset counts as one epoch.
43 |
44 | Therefore, the learning time is the learning time per step x (the number of data in the dataset / batch size) x the number of epochs. In general, the larger the batch size, the more stable the learning becomes (learning time per step ÷ batch size) becomes smaller, but it uses more GPU memory. GPU RAM can be checked with the nvidia-smi command. Learning can be done in a short time by increasing the batch size as much as possible according to the machine of the execution environment.
45 |
46 | ### Specify pretrained model
47 | RVC starts training the model from pretrained weights instead of from 0, so it can be trained with a small dataset.
48 |
49 | By default
50 |
51 | - If you consider pitch, it loads `rvc-location/pretrained/f0G40k.pth` and `rvc-location/pretrained/f0D40k.pth`.
52 | - If you don't consider pitch, it loads `rvc-location/pretrained/f0G40k.pth` and `rvc-location/pretrained/f0D40k.pth`.
53 |
54 | When learning, model parameters are saved in `logs/your-experiment-name/G_{}.pth` and `logs/your-experiment-name/D_{}.pth` for each save_every_epoch, but by specifying this path, you can start learning. You can restart or start training from model weights learned in a different experiment.
55 |
56 | ### learning index
57 | RVC saves the HuBERT feature values used during training, and during inference, searches for feature values that are similar to the feature values used during learning to perform inference. In order to perform this search at high speed, the index is learned in advance.
58 | For index learning, we use the approximate neighborhood search library faiss. Read the feature value of `logs/your-experiment-name/3_feature256` and use it to learn the index, and save it as `logs/your-experiment-name/add_XXX.index`.
59 |
60 | (From the 20230428update version, it is read from the index, and saving / specifying is no longer necessary.)
61 |
62 | ### Button description
63 | - Train model: After executing step2b, press this button to train the model.
64 | - Train feature index: After training the model, perform index learning.
65 | - One-click training: step2b, model training and feature index training all at once.
--------------------------------------------------------------------------------
/modules/rvc/docs/fr/training_tips_fr.md:
--------------------------------------------------------------------------------
1 | Instructions et conseils pour la formation RVC
2 | ======================================
3 | Ces conseils expliquent comment se déroule la formation des données.
4 |
5 | # Flux de formation
6 | Je vais expliquer selon les étapes de l'onglet de formation de l'interface graphique.
7 |
8 | ## étape 1
9 | Définissez ici le nom de l'expérience.
10 |
11 | Vous pouvez également définir ici si le modèle doit prendre en compte le pitch.
12 | Si le modèle ne considère pas le pitch, le modèle sera plus léger, mais pas adapté au chant.
13 |
14 | Les données de chaque expérience sont placées dans `/logs/nom-de-votre-experience/`.
15 |
16 | ## étape 2a
17 | Charge et pré-traite l'audio.
18 |
19 | ### charger l'audio
20 | Si vous spécifiez un dossier avec de l'audio, les fichiers audio de ce dossier seront lus automatiquement.
21 | Par exemple, si vous spécifiez `C:Users\hoge\voices`, `C:Users\hoge\voices\voice.mp3` sera chargé, mais `C:Users\hoge\voices\dir\voice.mp3` ne sera pas chargé.
22 |
23 | Comme ffmpeg est utilisé en interne pour lire l'audio, si l'extension est prise en charge par ffmpeg, elle sera lue automatiquement.
24 | Après la conversion en int16 avec ffmpeg, convertir en float32 et normaliser entre -1 et 1.
25 |
26 | ### débruitage
27 | L'audio est lissé par filtfilt de scipy.
28 |
29 | ### Séparation audio
30 | Tout d'abord, l'audio d'entrée est divisé en détectant des parties de silence qui durent plus d'une certaine période (max_sil_kept = 5 secondes ?). Après avoir séparé l'audio sur le silence, séparez l'audio toutes les 4 secondes avec un chevauchement de 0,3 seconde. Pour l'audio séparé en 4 secondes, après normalisation du volume, convertir le fichier wav en `/logs/nom-de-votre-experience/0_gt_wavs` puis le convertir à un taux d'échantillonnage de 16k dans `/logs/nom-de-votre-experience/1_16k_wavs` sous forme de fichier wav.
31 |
32 | ## étape 2b
33 | ### Extraire le pitch
34 | Extrait les informations de pitch des fichiers wav. Extraire les informations de pitch (=f0) en utilisant la méthode intégrée dans parselmouth ou pyworld et les sauvegarder dans `/logs/nom-de-votre-experience/2a_f0`. Convertissez ensuite logarithmiquement les informations de pitch en un entier entre 1 et 255 et sauvegardez-le dans `/logs/nom-de-votre-experience/2b-f0nsf`.
35 |
36 | ### Extraire l'empreinte de caractéristique
37 | Convertissez le fichier wav en incorporation à l'avance en utilisant HuBERT. Lisez le fichier wav sauvegardé dans `/logs/nom-de-votre-experience/1_16k_wavs`, convertissez le fichier wav en caractéristiques de dimension 256 avec HuBERT, et sauvegardez au format npy dans `/logs/nom-de-votre-experience/3_feature256`.
38 |
39 | ## étape 3
40 | former le modèle.
41 | ### Glossaire pour les débutants
42 | Dans l'apprentissage profond, l'ensemble de données est divisé et l'apprentissage progresse petit à petit. Dans une mise à jour de modèle (étape), les données de batch_size sont récupérées et des prédictions et corrections d'erreur sont effectuées. Faire cela une fois pour un ensemble de données compte comme une époque.
43 |
44 | Par conséquent, le temps d'apprentissage est le temps d'apprentissage par étape x (le nombre de données dans l'ensemble de données / taille du lot) x le nombre d'époques. En général, plus la taille du lot est grande, plus l'apprentissage devient stable (temps d'apprentissage par étape ÷ taille du lot) devient plus petit, mais il utilise plus de mémoire GPU. La RAM GPU peut être vérifiée avec la commande nvidia-smi. L'apprentissage peut être effectué en peu de temps en augmentant la taille du lot autant que possible selon la machine de l'environnement d'exécution.
45 |
46 | ### Spécifier le modèle pré-entraîné
47 | RVC commence à former le modèle à partir de poids pré-entraînés plutôt que de zéro, il peut donc être formé avec un petit ensemble de données.
48 |
49 | Par défaut :
50 |
51 | - Si vous considérez le pitch, il charge `rvc-location/pretrained/f0G40k.pth` et `rvc-location/pretrained/f0D40k.pth`.
52 | - Si vous ne considérez pas le pitch, il charge `rvc-location/pretrained/f0G40k.pth` et `rvc-location/pretrained/f0D40k.pth`.
53 |
54 | Lors de l'apprentissage, les paramètres du modèle sont sauvegardés dans `logs/nom-de-votre-experience/G_{}.pth` et `logs/nom-de-votre-experience/D_{}.pth` pour chaque save_every_epoch, mais en spécifiant ce chemin, vous pouvez démarrer l'apprentissage. Vous pouvez redémarrer ou commencer à former à partir de poids de modèle appris lors d'une expérience différente.
55 |
56 | ### Index d'apprentissage
57 | RVC sauvegarde les valeurs de caractéristique HuBERT utilisées lors de la formation, et pendant l'inférence, recherche les valeurs de caractéristique qui sont similaires aux valeurs de caractéristique utilisées lors de l'apprentissage pour effectuer l'inférence. Afin d'effectuer cette recherche à haute vitesse, l'index est appris à l'avance.
58 | Pour l'apprentissage d'index, nous utilisons la bibliothèque de recherche de voisinage approximatif faiss. Lisez la valeur de caractéristique de `logs/nom-de-votre-experience/3_feature256` et utilisez-la pour apprendre l'index, et sauvegardez-la sous `logs/nom-de-votre-experience/add_XXX.index`.
59 |
60 | (À partir de la version de mise à jour 20230428, elle est lue à partir de l'index, et la sauvegarde / spécification n'est plus nécessaire.)
61 |
62 | ### Description du bouton
63 | - Former le modèle : après avoir exécuté l'étape 2b, appuyez sur ce bouton pour former le modèle.
64 | - Former l'index de caractéristique : après avoir formé le modèle, effectuez un apprentissage d'index.
65 | - Formation en un clic : étape 2b, formation du modèle et formation de l'index de caractéristique tout d'un coup.```
66 |
--------------------------------------------------------------------------------
/modules/rvc/docs/jp/README.ja.md:
--------------------------------------------------------------------------------
1 |
2 |
3 |
Retrieval-based-Voice-Conversion-WebUI
4 | VITSに基づく使いやすい音声変換（voice changer）framework
5 |
6 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
8 |
9 |
10 |
11 | [](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI.ipynb)
12 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
13 | [](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
14 |
15 | [](https://discord.gg/HcsmBBGyVk)
16 |
17 |
2 |
3 |
Retrieval-based-Voice-Conversion-WebUI
4 | VITS基盤의 簡單하고使用하기 쉬운音聲變換틀
5 |
6 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
8 |
9 |
10 |
11 | [](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI.ipynb)
12 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
13 | [](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
14 |
15 | [](https://discord.gg/HcsmBBGyVk)
16 |
17 |
2 |
3 |
Retrieval-based-Voice-Conversion-WebUI
4 | VITS 기반의 간단하고 사용하기 쉬운 음성 변환 프레임워크.
5 |
6 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
8 |
9 |
10 |
11 | [](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI.ipynb)
12 | [](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
13 | [](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
14 |
15 | [](https://discord.gg/HcsmBBGyVk)
16 |
17 |