litagin's picture
init
2916d61
|
raw
history blame
3.81 kB

CLI

0. Install and global paths settings

git clone https://github.com/litagin02/Style-Bert-VITS2.git
cd Style-Bert-VITS2
python -m venv venv
venv\Scripts\activate
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Then download the necessary models and the default TTS model, and set the global paths.

python initialize.py [--skip_jvnv] [--dataset_root <path>] [--assets_root <path>]

Optional:

  • --skip_jvnv: Skip downloading the default JVNV voice models (use this if you only have to train your own models).
  • --dataset_root: Default: Data. Root directory of the training dataset. The training dataset of {model_name} should be placed in {dataset_root}/{model_name}.
  • --assets_root: Default: model_assets. Root directory of the model assets (for inference). In training, the model assets will be saved to {assets_root}/{model_name}, and in inference, we load all the models from {assets_root}.

1. Dataset preparation

1.1. Slice wavs

python slice.py --model_name <model_name> [-i <input_dir>] [-m <min_sec>] [-M <max_sec>]

Required:

  • model_name: Name of the speaker (to be used as the name of the trained model).

Optional:

  • input_dir: Path to the directory containing the audio files to slice (default: inputs)
  • min_sec: Minimum duration of the sliced audio files in seconds (default: 2).
  • max_sec: Maximum duration of the sliced audio files in seconds (default: 12).

1.2. Transcribe wavs

python transcribe.py --model_name <model_name>

Required:

  • model_name: Name of the speaker (to be used as the name of the trained model).

Optional

  • --initial_prompt: Initial prompt to use for the transcription (default value is specific to Japanese).
  • --device: cuda or cpu (default: cuda).
  • --language: jp, en, or en (default: jp).
  • --model: Whisper model, default: large-v3
  • --compute_type: default: bfloat16

2. Preprocess

python preprocess_all.py -m <model_name> [--use_jp_extra] [-b <batch_size>] [-e <epochs>] [-s <save_every_steps>] [--num_processes <num_processes>] [--normalize] [--trim] [--val_per_lang <val_per_lang>] [--log_interval <log_interval>] [--freeze_EN_bert] [--freeze_JP_bert] [--freeze_ZH_bert] [--freeze_style] [--freeze_decoder]

Required:

  • model_name: Name of the speaker (to be used as the name of the trained model).

Optional:

  • --batch_size, -b: Batch size (default: 2).
  • --epochs, -e: Number of epochs (default: 100).
  • --save_every_steps, -s: Save every steps (default: 1000).
  • --num_processes: Number of processes (default: half of the number of CPU cores).
  • --normalize: Loudness normalize audio.
  • --trim: Trim silence.
  • --freeze_EN_bert: Freeze English BERT.
  • --freeze_JP_bert: Freeze Japanese BERT.
  • --freeze_ZH_bert: Freeze Chinese BERT.
  • --freeze_style: Freeze style vector.
  • --freeze_decoder: Freeze decoder.
  • --use_jp_extra: Use JP-Extra model.
  • --val_per_lang: Validation data per language (default: 0).
  • --log_interval: Log interval (default: 200).

3. Train

Training settings are automatically loaded from the above process.

If NOT using JP-Extra model:

python train_ms.py [--repo_id <username>/<repo_name>]

If using JP-Extra model:

python train_ms_jp_extra.py [--repo_id <username>/<repo_name>] [--skip_default_style]

Optional:

  • --repo_id: Hugging Face repository ID to upload the trained model to. You should have logged in using huggingface-cli login before running this command.
  • --skip_default_style: Skip making the default style vector. Use this if you want to resume training (since the default style vector is already made).