# WhisperX

> WhisperX wraps a faster-whisper backend with batched inference (README: 70x realtime with large-v2), wav2vec2 forced alignment for word timestamps, and pyannote speaker diarization. Code is BSD-2-Clause. Diarization uses the gated pyannote speaker-diarization-community-1 model (CC-BY-4.0) and needs a free Hugging Face token. README states under 8 GB GPU memory for large-v2; CPU mode is supported.

- Repository: https://github.com/m-bain/whisperX
- Official docs: https://github.com/m-bain/whisperX#readme
- Checked: 2026-10-15

## What it replaces
- ElevenLabs: Speech recognition (full). Local transcription with word timestamps and speaker labels; README notes diarization is far from perfect.
- Otter.ai: Meeting transcription, Speaker identification (partial). Transcripts with generic speaker labels from recordings; no live meeting capture, names or summaries.

## Facts, with sources

| Question | Status | Answer | Source | Checked |
|---|---|---|---|---|
| Runs on my machine? | Confirmed | Yes, CLI and Python package on GPU or CPU | https://github.com/m-bain/whisperX#readme | 2026-10-15 |
| Code licence | Confirmed | BSD-2-Clause | https://github.com/m-bain/whisperX/blob/main/LICENSE | 2026-10-15 |
| Model weights licence | Partial | Mixed: Whisper weights via faster-whisper (Systran conversions tagged MIT); diarization model pyannote/speaker-diarization-community-1 is CC-BY-4.0 and gated; language-specific alignment model licences are not stated in the README | https://github.com/m-bain/whisperX#acknowledgements | 2026-10-15 |
| Commercial use | Partial | BSD-2-Clause code and MIT Whisper weights allow it; diarization adds CC-BY-4.0 attribution terms, and alignment model licences vary by language and are not stated | https://github.com/m-bain/whisperX#acknowledgements | 2026-10-15 |
| Paid services needed? | Confirmed | None required; diarization needs a free Hugging Face read token and accepting the pyannote model's user agreement | https://github.com/m-bain/whisperX#speaker-diarization | 2026-10-15 |
| Hardware | Confirmed | Under 8 GB GPU memory for large-v2 with beam\_size=5; CPU mode supported (int8); CUDA toolkit 12.8 for GPU | https://github.com/m-bain/whisperX#readme | 2026-10-15 |
| Operating systems | Confirmed | Linux and Windows CUDA setup documented; macOS via CPU mode | https://github.com/m-bain/whisperX#setup | 2026-10-15 |
| Setup effort | Confirmed | Medium; pip install whisperx, plus CUDA toolkit 12.8 for GPU and a Hugging Face token for diarization; ffmpeg may be needed | https://github.com/m-bain/whisperX#setup | 2026-10-15 |

GitHub: 24599 stars, 2483 forks, last push 2026-09-26, archived: no.
---
Source: RepoQuarry (https://repoquarry.dev/projects/whisperx). Data licence: CC BY 4.0. Quotes are from project documentation and belong to their authors.