Neural speaker diarization with `pyannote.audio`

pyannote.audio is an open-source toolkit written in Python for speaker diarization. Based on PyTorch machine learning framework, it provides a set of trainable end-to-end neural building blocks that can be combined and jointly optimized to build speaker diarization pipelines.

TL;DR

# 1. visit hf.co/pyannote/speaker-diarization and hf.co/pyannote/segmentation and accept user conditions (only if requested)
# 2. visit hf.co/settings/tokens to create an access token (only if you had to go through 1.)
# 3. instantiate pretrained speaker diarization pipeline
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization",
                                    use_auth_token="ACCESS_TOKEN_GOES_HERE")

# 4. apply pretrained pipeline
diarization = pipeline("audio.wav")

# 5. 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
# ...

What's new in `pyannote.audio` 2.x?

For version 2.x of pyannote.audio, I decided to rewrite almost everything from scratch. Highlights of this release are:

🤯 much better performance (see Benchmark)
🐍 Python-first API
🤗 pretrained pipelines (and models) on 🤗 model hub
⚡ multi-GPU training with pytorch-lightning
🎛️ data augmentation with torch-audiomentations
💥 Prodigy recipes for model-assisted audio annotation

Installation

Only Python 3.8+ is officially supported (though it might work with Python 3.7)

conda create -n pyannote python=3.8
conda activate pyannote

# pytorch 1.11 is required for speechbrain compatibility
# (see https://pytorch.org/get-started/previous-versions/#v1110)
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 -c pytorch

pip install pyannote.audio

Documentation

Changelog
Models
- Available tasks explained
- Applying a pretrained model
- Training, fine-tuning, and transfer learning
Pipelines
- Available pipelines explained
- Applying a pretrained pipeline
- Training a pipeline
Contributing
- Adding a new model
- Adding a new task
- Adding a new pipeline
- Sharing pretrained models and pipelines
Blog
- 2022-10-23 > "One speaker segmentation model to rule them all"
- 2021-08-05 > "Streaming voice activity detection with pyannote.audio"
Miscellaneous
- Training with pyannote-audio-train command line tool
- Annotating your own data with Prodigy
- Speaker verification
- Visualization and debugging

Frequently asked questions

How does one capitalize and pronounce the name of this awesome library?

📝 Written in lower case: pyannote.audio (or pyannote if you are lazy). Not PyAnnote nor PyAnnotate (sic). 📢 Pronounced like the french verb pianoter. pi like in piano, not py like in python. 🎹 pianoter means to play the piano (hence the logo 🤯).

Pretrained pipelines do not produce good results on my data. What can I do?

Annotate dozens of conversations manually and separate them into development and test subsets in pyannote.database.
Optimize the hyper-parameters of the pretained pipeline using the development set. If performance is still not good enough, go to step 3.
Annotate hundreds of conversations manually and set them up as training subset in pyannote.database.
Fine-tune the models (on which the pipeline relies) using the training set.
Optimize the hyper-parameters of the pipeline using the fine-tuned models using the development set. If performance is still not good enough, go back to step 3.

Benchmark

Out of the box, pyannote.audio default speaker diarization pipeline is expected to be much better (and faster) in v2.x than in v1.1. Those numbers are diarization error rates (in %)

Dataset \ Version	v1.1	v2.0	v2.1.1 (finetuned)
AISHELL-4	-	14.6	14.1 (14.5)
AliMeeting (channel 1)	-	-	27.4 (23.8)
AMI (IHM)	29.7	18.2	18.9 (18.5)
AMI (SDM)	-	29.0	27.1 (22.2)
CALLHOME (part2)	-	30.2	32.4 (29.3)
DIHARD 3 (full)	29.2	21.0	26.9 (21.9)
VoxConverse (v0.3)	21.5	12.6	11.2 (10.7)
REPERE (phase2)	-	12.6	8.2 ( 8.3)
This American Life	-	-	20.8 (15.2)

Citations

If you use pyannote.audio please use the following citations:

@inproceedings{Bredin2020,
  Title = {{pyannote.audio: neural building blocks for speaker diarization}},
  Author = {{Bredin}, Herv{\'e} and {Yin}, Ruiqing and {Coria}, Juan Manuel and {Gelly}, Gregory and {Korshunov}, Pavel and {Lavechin}, Marvin and {Fustes}, Diego and {Titeux}, Hadrien and {Bouaziz}, Wassim and {Gill}, Marie-Philippe},
  Booktitle = {ICASSP 2020, IEEE International Conference on Acoustics, Speech, and Signal Processing},
  Year = {2020},
}

@inproceedings{Bredin2021,
  Title = {{End-to-end speaker segmentation for overlap-aware resegmentation}},
  Author = {{Bredin}, Herv{\'e} and {Laurent}, Antoine},
  Booktitle = {Proc. Interspeech 2021},
  Year = {2021},
}

Support

For commercial enquiries and scientific consulting, please contact me.

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

Tests rely on a set of debugging files available in test/data directory. Set PYANNOTE_DATABASE_CONFIG environment variable to test/data/database.yml before running tests:

PYANNOTE_DATABASE_CONFIG=tests/data/database.yml pytest

farisalasmary/pyannote-audio

Neural speaker diarization with pyannote.audio