vcon-mac-wtf

MLX Whisper transcription server for Apple Silicon with WTF/vCon support.

Runs OpenAI's Whisper speech-to-text model locally on Apple Silicon (M1/M2/M3/M4) using the MLX framework for GPU-accelerated inference. Outputs transcriptions in World Transcription Format (WTF) and can enrich vCon documents with transcription analysis.

Features

• GPU-accelerated Whisper inference on Apple Silicon via MLX
• OpenAI-compatible API (POST /v1/audio/transcriptions) — drop-in replacement
• vCon-native API (POST /transcribe) — accepts a vCon, returns enriched vCon with WTF transcription
• Word-level timestamps
• Multiple model sizes (tiny through large-v3)
• Integrates with wtf-server as the mlx-whisper provider

Prerequisites

• Apple Silicon Mac (M1/M2/M3/M4)
• Python 3.12+
• uv (recommended) or pip
• ffmpeg: brew install ffmpeg

Quickstart

# Install (choose one)
pip install vcon-mac-wtf          # PyPI
brew install your-username/vcon/vcon-mac-wtf  # Homebrew tap

# Or from source
uv sync --all-extras

# Start the server (downloads model on first run)
vcon-mac-wtf
# or: make run
# or: uv run uvicorn vcon_mac_wtf.main:app --host 0.0.0.0 --port 8000

API Endpoints

Health

curl http://localhost:8000/health
curl http://localhost:8000/health/ready

Transcribe Audio (OpenAI-compatible)

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -F "file=@recording.wav" \
  -F "model=mlx-community/whisper-turbo" \
  -F "response_format=verbose_json"

Response formats: json, text, verbose_json, wtf

Transcribe vCon

curl -X POST http://localhost:8000/transcribe \
  -H "Content-Type: application/json" \
  -d @my_vcon.json

Returns the vCon with WTF transcription analysis appended.

List Models

curl http://localhost:8000/v1/models

Configuration

Copy .env.example to .env to customize. Add HF_TOKEN=hf_xxx to .env before first run to speed up model downloads and avoid rate limits.

Available Models

Integration with wtf-server

This server works as a provider for the existing TypeScript wtf-server. Set in the wtf-server .env:

ASR_PROVIDER=mlx-whisper
MLX_WHISPER_URL=http://localhost:8000
MLX_WHISPER_MODEL=mlx-community/whisper-turbo

Publishing

PyPI:

pip install build twine
python -m build && twine upload dist/*

Homebrew: See homebrew/README.md for tap setup. After publishing to PyPI, update the formula url and sha256, then add to your tap.

Development

make install    # Install all deps including dev
make dev        # Run with auto-reload
make test       # Run unit tests
make test-all   # Run all tests including integration
make lint       # Lint with ruff
make format     # Format with ruff

Testing

# Unit tests (no MLX required, uses mocks)
make test

# Integration tests (requires Apple Silicon + MLX)
make test-all

# Coverage
make test-cov