pyannote.audio
The open-source standard for speaker diarization: pretrained pipelines that answer who spoke when.
About
pyannote.audio is the open-source toolkit behind most serious speaker diarization work. Built on PyTorch by Hervé Bredin and the pyannote team, it ships neural building blocks for voice activity detection, speaker change detection, overlapped speech detection, and speaker embeddings, plus pretrained pipelines that segment a recording by speaker.
It is usually paired with a transcription engine: Whisper or faster-whisper produces the words, pyannote assigns the speakers, which is why services like Gladia build on it. The speaker-diarization-community-1 pipeline runs locally and improved diarization error rates across AMI, DIHARD 3, and VoxConverse versus the older 3.1 pipeline. Expect to fine-tune on your own data for the best results on far-field or noisy audio.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- AMI IHM 17.0%, AliMeeting 20.3%, DIHARD3 20.2%, AMI SDM 19.9% DER (speaker-diarization-community-1, provider-reported)
- ~10% DER with optimized configurations, ~2.5% real-time factor on GPU (third-party comparison, AssemblyAI)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: MIT-licensed library with ~9,700 GitHub stars; community-1 DER figures (AMI IHM 17.0%, DIHARD3 20.2%) cross-checked across the pyannote GitHub org and independent 2026 upgrade write-ups; ~2.5% real-time factor from AssemblyAI's 2026 comparison. Strengths: the de-facto open diarization standard, foundation for several commercial services, keeps improving (3.1 to community-1 cut speaker confusion ~50%). Docked for: (1) pretrained pipeline weights are HF-gated with their own licenses (community-1 is CC-BY-4.0, check commercial terms); (2) DER is reported in isolation from transcription, which flatters production accuracy. License: MIT (library); pipeline weights per-model (community-1: CC-BY-4.0).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · community-1 DER: AMI IHM 17.0%, AliMeeting 20.3%, DIHARD3 20.2% (provider-reported); ~2.5% RTF on GPU (AssemblyAI comparison)
- · Foundation for commercial services incl. Gladia; ~9,700 GitHub stars