WhisperX

• Source: https://github.com/m-bain/whisperX
• Server: whisperX-FastAPI
• License: MIT
• Alternatives: faster-whisper / Speaches, Parakeet, Vosk, Wav2Vec2

Speech-to-text transcription with speaker diarization (who said what), served by whisperX-FastAPI on an OpenAI-compatible API.

:::note Not the Home Assistant one This is for long recordings — an interview, a meeting, a voice memo — where you want to know who said what. It runs on the GPU and is driven by decree.

Home Assistant's voice pipeline uses wyoming-whisper instead: HA speaks the Wyoming TCP protocol, not HTTP, and short voice commands want low latency far more than they want speaker labels. :::

Setup

Diarization needs a HuggingFace token (WHISPERX_HF_TOKEN, prompted during setup). Accept the gated model terms once while logged in:

• https://huggingface.co/pyannote/speaker-diarization-community-1

Without the token, transcription and alignment still succeed and the task then fails at the diarize step with a HuggingFace 401 — every health route stays green, so check the task's error field, not the service.

References

• whisperX-FastAPI — the server wrapper
• pyannote.audio — the diarization models
• Recording Transcription — the automated, diarized pipeline