Pyannote Audio
PyTorch toolkit that labels who spoke when, with local pretrained pipelines and trainable building blocks.
About
Speaker diarization, working out who spoke when in a recording, is the central job of this PyTorch library created by Herve Bredin. Setup needs ffmpeg on the system for audio decoding, a pip or uv install, and a Hugging Face token from an account that has agreed to the gated community-1 terms; after that, one Pipeline.from_pretrained call yields timestamped turns tagged by speaker, computed on your own CPU or GPU. Beneath that pipeline sit reusable pieces for voice activity, speaker change and overlapped speech detection and for speaker embeddings, plus Lightning-based training to adapt models to your recordings. The same API can also call precision-2, a paid pyannoteAI pipeline that runs on the company's servers and posts clearly lower error rates on the published benchmarks. Library code is MIT and community-1 weights are CC-BY-4.0, so commercial use is fine with attribution. Telemetry can be switched off with the PYANNOTE_METRICS_ENABLED variable. It is the standard building block for meeting notes and call analytics, and pairs well with Whisper-style recognizers that transcribe words but never say who spoke.
Should you use Pyannote Audio?
Pick it when
Pick Pyannote when you need speaker diarization, overlap detection, or speaker embeddings to add who-spoke-when to transcripts from any ASR engine, and may fine-tune the pipeline on your own recordings.
Look elsewhere when
Skip it if you need the text itself, since it outputs no transcript; MOSS-Transcribe or WhisperX return speaker-labeled text directly. It also wants a GPU with about 4 GB, where Silero VAD runs on CPU for plain speech detection.
Alternatives to Pyannote Audio
- MOSS-Transcribe
Transcribes and labels speakers in one end-to-end pass, simpler than chaining ASR plus diarization, but you cannot swap in your own recognizer.
- WhisperX
Wraps Whisper with alignment and external diarization into a ready transcript pipeline, so you do not wire the parts together yourself.
- Silero VAD
Answers only whether speech is present, in under 1 ms on CPU, the lighter choice when you need a speech gate, not speaker identity.
- Reverb
Ships Rev's own Pyannote-based diarization tuned with its English ASR, but the weights are restricted to non-production use.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- MIT
- Minimum VRAM
- 4 GB
- Added
- Apr 3, 2026