README.md

metadata

title: SenseVoice Audio Transcription
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.0.2
app_file: app.py
pinned: false

Multilingual Audio Transcription with SenseVoice

This application transcribes audio using SenseVoice Small model with multilingual support, providing accurate transcription for Chinese, English, Japanese, Korean, and Cantonese.

Features

• Multilingual Support: Chinese (zh), English (en), Japanese (ja), Korean (ko), Cantonese (yue)
• Multiple Audio Sources:
  • Uploaded audio files
  • Direct URLs to audio files (no YouTube support due to cookie requirements)
• Model Options:
  • Local SenseVoice model
  • Hugging Face model: FunAudioLLM/SenseVoiceSmall
• Advanced Features:
  • Audio trimming with start/end time
  • Proxy support for downloads
  • Verbose logging output
  • Automatic inverse text normalization (ITN)

Model Setup

For Hugging Face Spaces Deployment

The app is configured to work with:

1. Local Model: "SenseVoiceSmall" - Model files in the same directory
2. HF Model: "FunAudioLLM/SenseVoiceSmall" - Auto-downloaded from Hugging Face

For Local Development

• Update MODEL_PATH_LIST in app.py to use your custom models
• Supports local paths and Hugging Face repository names

How to Use

1. Upload Audio: Click "Upload or Record Audio" to select your audio file
2. Select Model: Choose from available models in the dropdown
3. Configure Options:
   • Set start/end time for audio trimming
   • Enable verbose output for debugging
4. Transcribe: Click "Transcribe" to start the process

Git LFS Setup for Large Models

Since this project uses large model files, Git LFS is recommended:

# Initialize Git LFS
git lfs install

# Track large model files
git lfs track "*.bin"
git lfs track "*.safetensors"

# Add and commit
git add .gitattributes
git add .
git commit -m "Add SenseVoice model with LFS tracking"

Deployment Notes

Hugging Face Spaces

• Use git push huggingface main to deploy
• Models are automatically cached during runtime
• First load may be slower due to model download

Model Repository Structure

your-repo/
├── app.py
├── README.md
├── requirements.txt
└── SenseVoiceSmall/  # Model directory
    ├── config.json
    ├── model.bin
    └── other model files...

Output

The application provides:

• Transcription Text: Full processed transcription with ITN
• Metrics: Processing time and file information
• Download: Text file with transcription results

Supported Languages

• 🇨🇳 Chinese (Mandarin)
• 🇺🇸 English
• 🇯🇵 Japanese
• 🇰🇷 Korean
• 🇭🇰 Cantonese

Feedback and Contributions

Welcome feedback and contributions to improve this multilingual transcription tool.