Transcriber

Local browser-based transcription app for Apple Silicon Macs. It serves a FastAPI backend and vanilla HTML/CSS/JS frontend, then transcribes audio locally with MLX Whisper, NVIDIA Parakeet on MLX, Cohere Transcribe on ONNX Runtime/Core ML, or the optional Microsoft VibeVoice ASR backend for diarized output.

Requirements

• macOS on Apple Silicon
• Python 3.12+
• uv
• For vibevoice-7b: enough memory for the 7B ASR model; browser .webm recordings also need the ffmpeg command available on PATH for audio decoding

Setup

uv sync

The VibeVoice ASR source needed by this app is vendored in vibevoice/ under the upstream MIT license, so no second VibeVoice checkout is required.

Run

uv run python -m app.main

Open http://127.0.0.1:8000.

Choose a language and model, then either record from the browser microphone or choose an audio file. The UI shows upload and transcription progress, result metadata, and lets you copy or save the transcript as .txt.

Models

Configured model options live in config.yaml:

• tiny, base, small, medium: progressively larger multilingual Whisper models optimized for Apple Silicon through MLX
• large (Large v3 Turbo), large-v2, and large-v3: higher-accuracy multilingual MLX Whisper options
• distil-large-v3-en: a faster English-only Distil-Whisper model; selecting it limits the language picker to English
• parakeet-v3: NVIDIA Parakeet TDT 0.6B V3, converted to MLX for fast multilingual transcription on Apple Silicon
• parakeet-v2-en: the English-only NVIDIA Parakeet TDT 0.6B V2 MLX checkpoint
• cohere-transcribe-int8: Cohere Transcribe 03-2026 using the approximately 2.7 GB INT8 ONNX export and ONNX Runtime's ARM CPU provider
• vibevoice-7b: VibeVoice ASR with speaker segments, loaded through the local vendored VibeVoice package

Models download on first use and remain cached on disk. Consecutive runs with the same backend reuse the loaded model in memory. Switching between backend families automatically unloads the previous backend to protect Apple unified memory, so the first run after switching back must reload that model. Parakeet uses MLX/Metal; Cohere uses ONNX Runtime's optimized ARM CPU provider by default because the current INT8 export is incompatible with Core ML at runtime. VibeVoice also downloads its Hugging Face model on first real use and can require substantial memory; keep transcription.vibevoice_max_new_tokens conservative unless you know the target machine can handle more.

API

POST /api/transcriptions

Multipart form fields:

• audio: audio file or browser recording
• language: configured language code, currently de or en
• model_size: configured model id, for example small, large, or vibevoice-7b

Successful response:

{
  "transcript": "Text...",
  "language": "de",
  "model_size": "small",
  "duration_seconds": 12.3,
  "segments": []
}

Error response:

{
  "error": {
    "code": "unsupported_language",
    "message": "Choose German or English."
  }
}

Invalid requests return 400, oversized uploads return 413, and transcription failures return 500.

Configuration

Edit config.yaml for runtime settings:

• transcription.cache_dir: local model cache directory
• transcription.max_upload_size_mb: upload size limit
• transcription.upload_chunk_size_mb: temp-file streaming chunk size
• transcription.supported_languages: language picker and backend validation
• transcription.supported_model_sizes: model picker, repo ids, backend routing, and optional per-model language_codes
• transcription.parakeet_*: long-audio chunk and overlap durations for Parakeet
• transcription.cohere_chunk_duration_seconds: Cohere audio chunk duration, capped at the official processor's 35-second limit; long audio is split near quiet boundaries
• transcription.cohere_max_new_tokens: Cohere decoder limit, capped at the official 256-token generation limit per chunk
• transcription.cohere_execution_provider: auto, coreml, or cpu; auto and cpu use the compatible CPU provider, while coreml is an explicit experimental override
• transcription.vibevoice_*: VibeVoice device, dtype, attention mode, and generation settings
• logging.level: structured JSON log level
• ui.*: browser labels, status messages, and progress timing heuristics

server.host and server.port control the bind address used by uv run python -m app.main.

Development

uv run ruff format .
uv run ruff check .
uv run pytest -v

Project layout:

• app/main.py: FastAPI app, config loading, frontend template rendering
• app/api/: transcription API
• app/services/: backend routing, mlx-whisper, and VibeVoice integration
• app/static/: browser UI
• tests/: API, config, and service tests
• vibevoice/: vendored VibeVoice ASR model and processor code from Microsoft VibeVoice, MIT licensed

License

MIT License. Vendored VibeVoice code under vibevoice/ is also MIT licensed; see vibevoice/LICENSE.