README.md

metadata

title: VTT with Diarization
emoji: 🎙️
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit

Voice-to-Text with Speaker Diarization

Powered by faster-whisper and pyannote.audio running locally on this Space.

Features

• 🎯 High-quality transcription using faster-whisper (2-4x faster than OpenAI Whisper)
• 👥 Speaker diarization with pyannote.audio 3.1
• 🌍 Multi-language support (Arabic, English, French, German, Spanish, Russian, Chinese, etc.)
• ⚙️ Configurable parameters (beam size, best_of, model size)
• 🔧 Optimized for Arabic customer service calls with specialized prompts

Usage

1. Upload an audio file (mp3, wav, m4a, flac, etc.)
2. Select language (or leave blank for auto-detect)
3. Enable speaker diarization if needed (requires HF_TOKEN)
4. Adjust quality parameters if desired
5. Click "Transcribe"

Configuration

Set these in Space Settings → Variables:

• WHISPER_MODEL_SIZE: Model size (tiny, base, small, medium, large-v3) - default: small
• WHISPER_DEVICE: Device (cpu or cuda) - default: cpu
• WHISPER_COMPUTE_TYPE: Compute type (int8, int16, float32) - default: int8
• DEFAULT_LANGUAGE: Default language code - default: ar (Arabic)
• WHISPER_BEAM_SIZE: Beam search size (1-10) - default: 5
• WHISPER_BEST_OF: Best of candidates (1-10) - default: 5

Secrets (required for diarization):

• HF_TOKEN: Your Hugging Face token with access to pyannote/speaker-diarization-3.1

Model Information

Whisper Models

Model	Size	RAM	Quality	Speed
tiny	75MB	1GB	⭐⭐	Very Fast
base	150MB	1GB	⭐⭐⭐	Fast
small	500MB	2GB	⭐⭐⭐⭐	Moderate
medium	1.5GB	5GB	⭐⭐⭐⭐⭐	Slow
large-v3	3GB	10GB	⭐⭐⭐⭐⭐⭐	Very Slow

First Run

• First transcription will download the selected Whisper model automatically
• Diarization downloads ~700MB on first use (cached afterward)
• Models are stored in the Space's persistent storage

Technical Details

• Uses the same model loading approach as the Django backend
• faster-whisper automatically downloads models from Hugging Face
• Diarization pipeline is downloaded locally to avoid repeated API calls
• All processing happens on this Space (no external inference APIs)

Credits

• faster-whisper by Guillaume Klein
• pyannote.audio by Hervé Bredin
• Original Django backend by IZI Techs

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference