tags:
- pyannote
- pyannote-audio
- pyannote-audio-pipeline
- audio
- voice
- speech
- speaker
- speaker-diarization
- speaker-change-detection
- voice-activity-detection
- overlapped-speech-detection
- automatic-speech-recognition
datasets:
- ami
- dihard
- voxconverse
- aishell
- repere
- voxceleb
license: mit
🎹 Speaker diarization
Relies on pyannote.audio 2.0: see installation instructions.
TL;DR
# load the pipeline from Hugginface Hub
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization@2022.07")
# apply the pipeline to an audio file
diarization = pipeline("audio.wav")
# dump the diarization output to disk using RTTM format
with open("audio.rttm", "w") as rttm:
diarization.write_rttm(rttm)
Advanced usage
In case the number of speakers is known in advance, one can use the num_speakers
option:
diarization = pipeline("audio.wav", num_speakers=2)
One can also provide lower and/or upper bounds on the number of speakers using min_speakers
and max_speakers
options:
diarization = pipeline("audio.wav", min_speakers=2, max_speakers=5)
If you feel adventurous, you can try and play with the various pipeline hyper-parameters.
For instance, one can use a more aggressive voice activity detection by increasing the value of segmentation_onset
threshold:
hparams = pipeline.parameters(instantiated=True)
hparams["segmentation_onset"] += 0.1
pipeline.instantiate(hparams)
Benchmark
Real-time factor
Real-time factor is around 5% using one Nvidia Tesla V100 SXM2 GPU (for the neural inference part) and one Intel Cascade Lake 6248 CPU (for the clustering part).
In other words, it takes approximately 3 minutes to process a one hour conversation.
Accuracy
This pipeline is benchmarked on a growing collection of datasets.
Processing is fully automatic:
- no manual voice activity detection (as is sometimes the case in the literature)
- no manual number of speakers (though it is possible to provide it to the pipeline)
- no fine-tuning of the internal models nor tuning of the pipeline hyper-parameters to each dataset
... with the least forgiving diarization error rate (DER) setup (named "Full" in this paper):
- no forgiveness collar
- evaluation of overlapped speech
Benchmark | DER% | FA% | Miss% | Conf% | Expected output | File-level evaluation |
---|---|---|---|---|---|---|
AISHELL-4 | 14.61 | 3.31 | 4.35 | 6.95 | RTTM | eval |
AMI Mix-Headset only_words | 18.21 | 3.28 | 11.07 | 3.87 | RTTM | eval |
AMI Array1-01 only_words | 29.00 | 2.71 | 21.61 | 4.68 | RTTM | eval |
CALLHOME Part2 | 30.24 | 3.71 | 16.86 | 9.66 | RTTM | eval |
DIHARD 3 Full | 20.99 | 4.25 | 10.74 | 6.00 | RTTM | eval |
REPERE Phase 2 | 12.62 | 1.55 | 3.30 | 7.76 | RTTM | eval |
VoxConverse v0.0.2 | 12.76 | 3.45 | 3.85 | 5.46 | RTTM | eval |
Support
For commercial enquiries and scientific consulting, please contact me.
For technical questions and bug reports, please check pyannote.audio Github repository.
Citations
@inproceedings{Bredin2021,
Title = {{End-to-end speaker segmentation for overlap-aware resegmentation}},
Author = {{Bredin}, Herv{\'e} and {Laurent}, Antoine},
Booktitle = {Proc. Interspeech 2021},
Address = {Brno, Czech Republic},
Month = {August},
Year = {2021},
}
@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},
Address = {Barcelona, Spain},
Month = {May},
Year = {2020},
}