Using this open-source model in production?
Consider switching to pyannoteAI for better and faster options.

🎹 Speaker diarization 3.0

This pipeline has been trained by Séverin Baroudi with pyannote.audio 3.0.0 using a combination of the training sets of AISHELL, AliMeeting, AMI, AVA-AVD, DIHARD, Ego4D, MSDWild, REPERE, and VoxConverse.

It ingests mono audio sampled at 16kHz and outputs speaker diarization as an Annotation instance:

stereo or multi-channel audio files are automatically downmixed to mono by averaging the channels.
audio files sampled at a different rate are resampled to 16kHz automatically upon loading.

Requirements

Install pyannote.audio 3.0 with pip install pyannote.audio
Accept pyannote/segmentation-3.0 user conditions
Accept pyannote/speaker-diarization-3.0 user conditions
Create access token at hf.co/settings/tokens.

Usage

# instantiate the pipeline
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained(
  "pyannote/speaker-diarization-3.0",
  use_auth_token="HUGGINGFACE_ACCESS_TOKEN_GOES_HERE")

# run the pipeline on 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)

Processing on GPU

pyannote.audio pipelines run on CPU by default. You can send them to GPU with the following lines:

import torch
pipeline.to(torch.device("cuda"))

Real-time factor is around 2.5% 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 1.5 minutes to process a one hour conversation.

Processing from memory

Pre-loading audio files in memory may result in faster processing:

waveform, sample_rate = torchaudio.load("audio.wav")
diarization = pipeline({"waveform": waveform, "sample_rate": sample_rate})

Monitoring progress

Hooks are available to monitor the progress of the pipeline:

from pyannote.audio.pipelines.utils.hook import ProgressHook
with ProgressHook() as hook:
    diarization = pipeline("audio.wav", hook=hook)

Controlling the number of speakers

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)

Benchmark

This pipeline has been benchmarked on a large 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	12.3	3.8	4.4	4.1	RTTM	eval
AliMeeting (channel 1)	24.3	4.4	10.0	9.9	RTTM	eval
AMI (headset mix, only_words)	19.0	3.6	9.5	5.9	RTTM	eval
AMI (array1, channel 1, only_words)	22.2	3.8	11.2	7.3	RTTM	eval
AVA-AVD	49.1	10.8	15.7	22.5	RTTM	eval
DIHARD 3 (Full)	21.7	6.2	8.1	7.3	RTTM	eval
MSDWild	24.6	5.8	8.0	10.7	RTTM	eval
REPERE (phase 2)	7.8	1.8	2.6	3.5	RTTM	eval
VoxConverse (v0.3)	11.3	4.1	3.4	3.8	RTTM	eval

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},
}

pyannote
/

speaker-diarization-3.0

You need to agree to share your contact information to access this model