What is pyannote speaker-diarization-community-1?
speaker-diarization-community-1 is the open pipeline from pyannoteAI, the Paris company founded in 2024 by the authors of the pyannote.audio toolkit. It does speaker diarisation: given a recording, it returns time stamps for each stretch of speech and a label for the speaker (SPEAKER_00, SPEAKER_01 and so on). It does not transcribe words. Teams pair it with a speech-to-text model such as Whisper and merge the two outputs.
It shipped with pyannote.audio 4.0 on 29 September 2025 and replaces the 3.1 pipeline that many open tools still use. It chains a segmentation model, a WeSpeaker speaker-embedding model and VBx clustering, takes 16 kHz mono audio, and adds an "exclusive" output in which only one speaker is active at a time, which makes lining up speaker turns with transcript words simpler.
The weights are CC BY 4.0 and the library is MIT. Downloading requires a Hugging Face token and sharing contact details with pyannoteAI. The model recorded about 5.1 million downloads in the past month.
How much does pyannote speaker-diarization-community-1 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | CC BY 4.0; MIT library |
| Hosted community-1 | €0.035 per hour | |
| Precision-3 (hosted) | €0.112 per hour (Developer) or €0.096 (Starter) | Plans €19/month and €99/month, credit included |
Free under CC BY 4.0. Hosted on pyannoteAI at €0.035 per hour of audio; its paid Precision-3 model is €0.112 per hour (Developer, €19/month) or €0.096 (Starter, €99/month).
Is pyannote speaker-diarization-community-1 worth it?
community-1 is the safe open choice for diarisation because almost everything else is built around pyannote. On the card's twelve benchmarks it lowers diarisation error rate (DER, the share of time attributed to the wrong speaker, missed or falsely detected) on ten of them against 3.1: 22.7% to 19.9% on single-microphone AMI meetings, 28.5% to 26.7% on CALLHOME phone calls and 24.5% to 20.3% on AliMeeting. It ties on VoxConverse and gets worse on REPERE, 7.9% to 8.9%. The scoring is strict, with no forgiveness collar and overlapping speech included.
The gap to pyannoteAI's own paid model is the honest benchmark. The commercial pipeline in the same table scores 16.6% on CALLHOME against community-1's 26.7%, and 10.5% against 20.8% on the RAMC Mandarin conversation set. The company has since moved its API to Precision-3, so the open model is now two generations behind the paid one, by design.
It is a component, not a transcriber. Teams still need a speech-to-text model and code to merge word times with speaker turns; the exclusive-diarisation output helps, but single-model systems such as MOSS-Transcribe-Diarize do both jobs in one pass. The card publishes no speed figures, runs on CPU unless moved to a GPU, and wild audio remains hard: DER is 44.6% on AVA-AVD film clips and 46.8% on Ego4D head-camera recordings.
The terms are easy for a business. CC BY 4.0 needs only attribution, the library is MIT, and a cloned copy runs fully offline. The gate is a marketing opt-in rather than a licence restriction, and with about 5.1 million monthly downloads and 10,750 GitHub stars, problems tend to be documented before you meet them.
Where it earns its place
A backend engineer at a 30-person meeting-notes start-up
Label who said what in recorded calls before summarising them
Run community-1 on a GPU next to Whisper and merge exclusive speaker turns with word time stamps. Passing min_speakers and max_speakers from the calendar invite reduces counting errors; there is no per-hour fee.
A call-centre analytics lead processing 10,000 hours of calls a month
Separate agent and customer speech without running GPUs
Switch to the hosted community-1 endpoint at €0.035 an hour, about €350 a month. Moving the same code to Precision-3 costs about €960 a month on the Starter rate for markedly lower error.
A university research team analysing interview recordings
Diarise sensitive audio without it leaving the lab
Clone the repository once with a token, then load it from disk with no internet connection. CC BY 4.0 requires only attribution in published work.
Before you commit
- Setup
- pip install pyannote.audio (4.x), accept the gate on Hugging Face, create a token, then Pipeline.from_pretrained and pipeline.to(cuda); audio is downmixed and resampled to 16 kHz mono on load.
- Integrations
- pyannote.audio Python library; widely used under Whisper-based tools; hosted community-1-cloud and Precision-3 on the pyannoteAI API with 100 requests a minute on Developer and Starter plans.
- Data and privacy
- Self-hosted copies run fully offline. Weights CC BY 4.0, code MIT; the gate shares your contact details for occasional marketing email. pyannoteAI advertises GDPR compliance and EU data residency for its API.
- Cost at scale
- Self-hosted is GPU time only; hosted community-1 is €0.035 per hour of audio, so 10,000 hours is about €350 a month, and Precision-3 is €0.096 to €0.112 per hour.
- Limits
- No transcription; DER from 8.9% on broadcast news to 46.8% on Ego4D; no published speed figures; regresses against 3.1 on REPERE.
What’s good
- Lower error than 3.1 on 10 of 12 benchmarks
- CC BY 4.0 weights and MIT library; runs offline
- Exclusive mode simplifies merging with transcripts
- Same code can switch to hosted models by changing one line
What’s not
- Diarisation only; needs a separate speech-to-text model
- About 10 points worse DER than pyannoteAI's paid model on CALLHOME
- No published speed figures
- Download is gated behind sharing contact details
Best for
Engineers building self-hosted meeting, call or podcast transcription who need speaker labels under a permissive licence.
pyannote speaker-diarization-community-1 alternatives
pyannote speaker-diarization-community-1: common questions
- Is pyannote speaker-diarization-community-1 worth it?
- Still the open default for telling speakers apart, and a clear step up from 3.1, but its own makers' paid model makes about a third fewer errors on hard audio. We rate it 4.1 out of 5.
- How much does pyannote speaker-diarization-community-1 cost?
- Free under CC BY 4.0. Hosted on pyannoteAI at €0.035 per hour of audio; its paid Precision-3 model is €0.112 per hour (Developer, €19/month) or €0.096 (Starter, €99/month).
- What are the downsides of pyannote speaker-diarization-community-1?
- Diarisation only; needs a separate speech-to-text model; about 10 points worse DER than pyannoteAI's paid model on CALLHOME; no published speed figures; download is gated behind sharing contact details.
- Who is pyannote speaker-diarization-community-1 best for?
- Engineers building self-hosted meeting, call or podcast transcription who need speaker labels under a permissive licence.
- How much does pyannote speaker-diarization-community-1 cost for a team or in production?
- Self-hosted is GPU time only; hosted community-1 is €0.035 per hour of audio, so 10,000 hours is about €350 a month, and Precision-3 is €0.096 to €0.112 per hour.
- Is pyannote speaker-diarization-community-1 safe for company data?
- Self-hosted copies run fully offline. Weights CC BY 4.0, code MIT; the gate shares your contact details for occasional marketing email. pyannoteAI advertises GDPR compliance and EU data residency for its API.
- What are the best pyannote speaker-diarization-community-1 alternatives?
- The closest alternatives we have reviewed are MOSS-Transcribe-Diarize 0.9B, Whisper large-v3-turbo, ElevenLabs Scribe v2 and Deepgram Nova-3.
Answers checked Nov 2, 2026.
Make pyannote speaker-diarization-community-1? Show this rating on your site
