metadata

license: mit
language:
  - ru
library_name: pyannote-audio
tags:
  - code

Segmentation model

This model was trained on AMI-MixHeadset and my own synthetic dataset of Russian speech.

Training time: 5 hours on GTX 3060

This model can be used for diarization model from pyannote/speaker-diarization

Benchmark	DER%
AMI (headset mix, only_words)	38.8

Usage example

import yaml
from yaml.loader import SafeLoader

import torch
from pyannote.audio import Model
from pyannote.audio.pipelines import SpeakerDiarization


segm_model = torch.load('model/segm_model.pth', map_location=torch.device('cpu'))
embed_model = Model.from_pretrained("pyannote/embedding", use_auth_token='ACCESS_TOKEN_GOES_HERE')
diar_pipeline = SpeakerDiarization(
    segmentation=segm_model,
    segmentation_batch_size=16,
    clustering="AgglomerativeClustering",
    embedding=embed_model
)

with open('model/config.yaml', 'r') as f:
    diar_config = yaml.load(f, Loader=SafeLoader)
diar_pipeline.instantiate(diar_config)

annotation = diar_pipeline('audio.wav')