Fablio/Samples-SBV2

Style-Bert-VITS2 Kariusagi Demo

This is a demo for Style-Bert-VITS2 using the kariusagi voice model.

Features

• —Japanese text-to-speech synthesis
• —Emotion and style control
• —Real-time audio generation

Model

This demo uses the kariusagi model trained with Style-Bert-VITS2.

Hugging Face Spaces Deployment

When deploying to Hugging Face Spaces, set the following environment variable:

DISABLE_PYOPENJTALK_WORKER=True

This is required for proper operation in the Hugging Face Spaces environment.

Model Download

The application requires two models that are automatically downloaded during Docker build:

1. 1.BERT Model: deberta-v2-large-japanese-char-wwm
2. 2.Downloaded from: Fablio/samples-style-bert-vits2-models
3. 3.Files: pytorch_model.bin, config.json, tokenizer_config.json, special_tokens_map.json, vocab.txt

1. 1.Kariusagi Voice Model:
2. 2.Downloaded from: Fablio/samples-style-bert-vits2-models
3. 3.Files: config.json, kariusagi_v0.0_e135_s9000.safetensors, style_vectors.npy

Manual Download for Development

If you need to download the models manually for local development:

from huggingface_hub import hf_hub_download
import os

# Create directories
os.makedirs('bert/deberta-v2-large-japanese-char-wwm', exist_ok=True)
os.makedirs('model_assets/kariusagi', exist_ok=True)

# Download BERT model files
bert_files = ['pytorch_model.bin', 'config.json', 'tokenizer_config.json', 'special_tokens_map.json', 'vocab.txt']
for file in bert_files:
    hf_hub_download(
        repo_id='Fablio/samples-style-bert-vits2-models',
        filename=f'deberta-v2-large-japanese-char-wwm/{file}',
        local_dir='bert',
        local_dir_use_symlinks=False
    )

# Download Kariusagi model files
kariusagi_files = ['config.json', 'kariusagi_v0.0_e135_s9000.safetensors', 'style_vectors.npy']
for file in kariusagi_files:
    hf_hub_download(
        repo_id='Fablio/samples-style-bert-vits2-models',
        filename=f'kariusagi/{file}',
        local_dir='model_assets',
        local_dir_use_symlinks=False
    )

After downloading, run the initialization script:

python initialize.py --skip_default_models