# pyannote (pyannoteAI)

> pyannote is the leading open-source speaker diarization toolkit, and pyannoteAI the Paris company behind it: community-1, Precision-3 and Live-1.

Source: https://metavert.io/pyannote  
Published: 2026-10-21  
Updated: 2026-10-21

**pyannote** is the most widely used open-source toolkit for [speaker diarization](https://metavert.io/speaker-diarization): working out who spoke when in a recording. pyannoteAI is the Paris company built around it. The library, pyannote.audio, was started on GitHub in March 2016 by the researcher Hervé Bredin. It's written in Python on PyTorch, its code is MIT-licensed, and it has about 10,700 GitHub stars. Podcast tools, meeting recorders, transcription services and research projects use it to split audio into speaker turns that a [speech recognizer](https://metavert.io/speech-recognition) can then label. Its models are distributed through [Hugging Face](https://metavert.io/hugging-face), and downloading them requires a Hugging Face access token.

### The Open-Source Pipeline

pyannote.audio 4.0 arrived on September 29, 2025 with a new default pipeline, community-1, released under CC-BY-4.0. The latest version is 4.0.7 (June 30, 2026), which needs Python 3.10 or later and FFmpeg. The pipeline has three stages: segmentation finds speech and speaker changes, an embedding model turns each segment into a voice vector (see [embeddings](https://metavert.io/embeddings)), and clustering groups the vectors into speakers. Community-1 added an "exclusive" mode that assigns one speaker at a time, which makes the output easier to align with transcript timestamps.

| Benchmark (DER, %) | pyannote 3.1 | community-1 | Precision-2 (paid) |
| --- | --- | --- | --- |
| AMI (IHM) | 18.8 | 17.0 | 12.9 |
| DIHARD 3 | 21.4 | 20.2 | 14.7 |
| VoxConverse | 11.2 | 11.2 | 8.5 |
| AISHELL-4 | 12.2 | 11.7 | 11.4 |
| AliMeeting | 24.5 | 20.3 | 15.2 |

The figures are the project's own, from the community-1 model card; lower diarization error rate is better. pyannote's models also run outside its library. The sherpa-onnx toolkit, for example, ships pyannote's segmentation 3.0 model for on-device diarization paired with NVIDIA NeMo TitaNet or 3D-Speaker embeddings, which is one way to run pyannote's models on laptops and [edge devices](https://metavert.io/edge-ai).

### pyannoteAI

Bredin co-founded pyannoteAI with Vincent Molina. In April 2025 it raised a $9 million seed round led by Crane Venture Partners and Serena, and it said the open-source project was used by more than 100,000 developers. The company sells hosted and on-premises models with higher accuracy than the open pipeline:

- **Precision-2** (September 2, 2025): about 14% more accurate than Precision-1, with better speaker counting.
- **Precision-3** (September 17, 2026): average DER falls 10.4% against Precision-2, from 16.02 to 14.35 across its benchmarks, with new controls for voice-activity and crosstalk sensitivity.
- **Live-1** (GA July 7, 2026): streaming diarization with under 300 milliseconds of latency.
- **STT orchestration** (December 11, 2025): diarization and transcription returned together, extended to OpenAI's Whisper large-v3-turbo in January 2026.

Premium features include voiceprints and confidence scores. Plans start at €19 a month for developers and €99 for startups, with custom enterprise pricing.

### How It Compares

pyannote's open alternatives come mainly from NVIDIA: the end-to-end Streaming Sortformer models and the TitaNet speaker-embedding model, which sit in the same NeMo toolkit as its [Parakeet](https://metavert.io/parakeet) recognizers. Hosted speech APIs bundle diarization too. [AssemblyAI](https://metavert.io/assemblyai) builds it into its Universal-3.5 Pro recognizer, [Deepgram](https://metavert.io/deepgram) offers it as an add-on, and ElevenLabs' Scribe v2 handles up to 32 speakers. Published DER figures are hard to compare across vendors, because benchmarks differ on whether overlapping speech is scored and whether a tolerance collar is allowed at turn boundaries. Voiceprints are biometric data, so identifying named speakers raises consent questions that anonymous diarization doesn't (see [speaker diarization vs speaker identification](https://metavert.io/compare/speaker-diarization-vs-speaker-identification) and [data privacy](https://metavert.io/data-privacy)).

## Related Topics

- [Speaker Diarization](https://metavert.io/speaker-diarization) — Who spoke when, and how it's measured
- [Speaker Diarization vs Speaker Identification](https://metavert.io/compare/speaker-diarization-vs-speaker-identification) — Anonymous turns versus named voices
- [Speech Recognition (ASR)](https://metavert.io/speech-recognition) — The transcription that diarization labels
- [Deepgram vs AssemblyAI](https://metavert.io/compare/deepgram-vs-assemblyai) — Hosted APIs with diarization built in
- [Transcript-Based Editing](https://metavert.io/transcript-based-editing) — Editing by words, with speakers
- [Speech & Voice AI](https://metavert.io/speech-voice-ai) — The broader field
- [Data Privacy](https://metavert.io/data-privacy) — Rules for voice and biometric data
- [Open Source AI](https://metavert.io/open-source-ai) — The model of development behind pyannote

## Further Reading

- [pyannote.audio on GitHub](https://github.com/pyannote/pyannote-audio) — source, documentation and MIT license
- [pyannote.audio on PyPI](https://pypi.org/project/pyannote.audio/) — version 4.0.7, June 2026
- [community-1 model card](https://huggingface.co/pyannote/speaker-diarization-community-1) — CC-BY-4.0 pipeline and benchmark table
- [pyannoteAI changelog](https://www.pyannote.ai/changelog) — Precision-3, Live-1 and STT orchestration
- [sherpa-onnx speaker diarization](https://k2-fsa.github.io/sherpa/onnx/speaker-diarization/index.html) — running pyannote segmentation on-device