Using pyannote.audio open-source toolkit in production?
Consider switching to
pyannoteAI for better and faster options.
pyannote speaker diarization toolkit
pyannote.audio is an open-source toolkit written in Python for speaker diarization. Based on PyTorch machine learning framework, it comes with state-of-the-art pretrained models and pipelines, that can be further finetuned to your own data for even better performance.
Highlights
- 🤯 state-of-the-art performance (see Benchmark)
- 🤗 pretrained pipelines (and models) on 🤗 model hub
- 🚀 built-in support for pyannoteAI premium speaker diarization
- 🐍 Python-first API
- ⚡ multi-GPU training with pytorch-lightning
pyannote.audio open-source speaker diarization pipeline
- Make sure
ffmpegis installed on your machine (needed bytorchcodecaudio decoding library) - Install with
uvadd pyannote.audio(recommended) orpip install pyannote.audio - Accept
pyannote/speaker-diarization-4.0user conditions - Create Huggingface access token at
hf.co/settings/tokens
import torch
from pyannote.audio import Pipeline
from pyannote.audio.pipelines.utils.hook import ProgressHook
# Open-source pyannote speaker diarization pipeline
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-4.0",
token="HUGGINGFACE_ACCESS_TOKEN")
# send pipeline to GPU (when available)
pipeline.to(torch.device("cuda"))
# apply pretrained pipeline (with optional progress hook)
with ProgressHook() as hook:
diarization = pipeline("audio.wav", hook=hook) # runs locally
# print the result
for turn, _, speaker in diarization.itertracks(yield_label=True):
print(f"start={turn.start:.1f}s stop={turn.end:.1f}s speaker_{speaker}")
# start=0.2s stop=1.5s speaker_0
# start=1.8s stop=3.9s speaker_1
# start=4.2s stop=5.7s speaker_0
# ...
pyannoteAI premium speaker diarization pipeline
- Create pyannoteAI API key at
dashboard.pyannote.ai - Enjoy 150 hours of free credits!
from pyannote.audio import Pipeline
# Premium pyannoteAI speaker diarization service
pipeline = Pipeline.from_pretrained(
"pyannoteAI/speaker-diarization-precision", token="PYANNOTEAI_API_KEY")
diarization = pipeline("audio.wav") # runs on pyannoteAI servers
# print the result
for turn, _, speaker in diarization.itertracks(yield_label=True):
print(f"start={turn.start:.1f}s stop={turn.end:.1f}s {speaker}")
# start=0.2s stop=1.6s SPEAKER_00
# start=1.8s stop=4.0s SPEAKER_01
# start=4.2s stop=5.6s SPEAKER_00
# ...
Visit docs.pyannote.ai to learn about other pyannoteAI features (voiceprinting, confidence scores, ...)
Benchmark
Out of the box,
pyannote.audio speaker diarization pipeline v4.0 is expected to be much better than v3.1.
pyannoteAI premium models are even better (and also 2x faster).
labs model is currently in private beta.
| Benchmark (last updated in 2025-08) | v3.1 |
v4.0 |
API |
labs |
|---|---|---|---|---|
| AISHELL-4 | 12.2 | 11.7 | 11.8 | 11.4 |
| AliMeeting (channel 1) | 24.5 | 20.3 | 16.3 | 15.2 |
| AMI (IHM) | 18.8 | 17.0 | 13.2 | 12.9 |
| AMI (SDM) | 22.7 | 19.9 | 15.8 | 15.6 |
| AVA-AVD | 49.7 | 44.6 | 40.7 | 37.1 |
| CALLHOME (part 2) | 28.5 | 26.7 | 17.6 | 16.6 |
| DIHARD 3 (full) | 21.4 | 20.2 | 15.7 | 14.7 |
| Ego4D (dev.) | 51.2 | 46.8 | 44.7 | 39.0 |
| MSDWild | 25.4 | 22.8 | 17.9 | 17.3 |
| RAMC | 22.2 | 20.8 | 10.6 | 10.5 |
| REPERE (phase2) | 7.9 | 8.9 | 7.3 | 7.4 |
| VoxConverse (v0.3) | 11.2 | 11.2 | 9.0 | 8.5 |
Diarization error rate (in %, the lower, the better)
| Benchmark (last updated in 2025-08) | Speed up | ||
|---|---|---|---|
| AMI (IHM), ~1h files | 31s per hour of audio | 14s per hour of audio | 2.2x faster |
| DIHARD 3 (full), ~5min files | 37s per hour of audio | 14s per hour of audio | 2.6x faster |
Processing speed on a NVIDIA H100 80GB HBM3
Telemetry
With the optional telemetry feature in pyannote.audio, you can choose to send anonymous usage metrics to help the pyannote team improve the library.
What we track
For each call to Pipeline.from_pretrained({origin}) (or Model.from_pretrained({origin})), we track information about {origin} in the following privacy-preserving way:
- If
{origin}is an officialpyannoteorpyannoteAIpipeline (or model) hosted onHuggingface, we track it as{origin}. - If
{origin}is a pipeline (or model) hosted onHuggingfacefrom any other organization, we track it ashuggingface. - If
{origin}is a path to a local file or directory, we track it aslocal.
We also track the pipeline Python class (e.g. pyannote.audio.pipelines.SpeakerDiarization).
For each file processed with a pipeline, we track
- the file duration in seconds
- the value of
num_speakers,min_speakers, andmax_speakersfor speaker diarization pipelines
We do not track any information that could identify who the user is.
Configuring telemetry
Telemetry can be configured in three ways:
- Using an environment variable
- Within the current Python session only
- Globally across sessions
All of these options will modify the value of the environment variable for consistency.
If the environment variable is not set, pyannote.audio will read the default value in the telemetry config.
The default config can also be changed from Python.
Using environment variable
You can control telemetry by setting the PYANNOTE_METRICS_ENABLED environment variable:
# enable metrics
export PYANNOTE_METRICS_ENABLED=1
# disable metrics
export PYANNOTE_METRICS_ENABLED=0
For current session
To control telemetry for your current Python kernel session:
from pyannote.audio.telemetry import set_telemetry_metrics
# enable metrics for current session
set_telemetry_metrics(True)
# disable metrics for current session
set_telemetry_metrics(False)
Global configuration
To set telemetry preferences that persist across sessions:
from pyannote.audio.telemetry import set_telemetry_metrics
# enable metrics globally
set_telemetry_metrics(True, save_choice_as_default=True)
# disable metrics globally
set_telemetry_metrics(False, save_choice_as_default=True)
Documentation
- Changelog
- Videos
- Speaker diarization, a
loveloss story / JSALT 2025 plenary talk / 60 min - Introduction to speaker diarization / JSALT 2023 summer school / 90 min
- Speaker segmentation model / Interspeech 2021 / 3 min
- First release of pyannote.audio / ICASSP 2020 / 8 min
- Speaker diarization, a
- Blog
- Community contributions (not maintained by the core team)
- 2024-04-05 > Offline speaker diarization (speaker-diarization-3.1) by Simon Ottenhaus
- 2024-09-24 > Evaluating
pyannotepretrained speech separation pipelines by Clément Pagés
- Tutorials
Those tutorials were written for older versions of pyannote.audio and should be updated. Interested in working for pyannoteAI as a community manager or developer advocate? This might be a nice place to start!
Citations
If you use pyannote.audio please use the following citations:
@inproceedings{Plaquet23,
author={Alexis Plaquet and Hervé Bredin},
title={{Powerset multi-class cross entropy loss for neural speaker diarization}},
year=2023,
booktitle={Proc. INTERSPEECH 2023},
}
@inproceedings{Bredin23,
author={Hervé Bredin},
title={{pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe}},
year=2023,
booktitle={Proc. INTERSPEECH 2023},
}
Development
The commands below will setup pre-commit hooks and packages needed for developing the pyannote.audio library.
pip install -e .[dev,testing]
pre-commit install
Test
pytest
