This repository contains tools and scripts for fine-tuning pyannote speaker segmentation models for improved speaker diarization performance.

This project provides a workflow for:

• Creating custom datasets for speaker segmentation
• Training pyannote segmentation models on your data
• Integrating fine-tuned models into existing pyannote pipelines

Create your custom dataset following the guidelines from the diarizers repository:

# Follow the dataset creation guide
# https://github.com/huggingface/diarizers/tree/main/datasets

Train your pyannote segmentation model using the diarizers training pipeline:

# Follow the training guide
# https://github.com/huggingface/diarizers/tree/main

Replace the segmentation model in your pyannote pipeline with your fine-tuned model:

from pyannote.audio import Pipeline
from diarizers import SegmentationModel

# Load your fine-tuned model
model = SegmentationModel().from_pretrained("diarizers-community/speaker-segmentation-fine-tuned-callhome-jpn")
model = model.to_pyannote_model()

# Initialize pipeline and replace segmentation model
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
pipeline._segmentation.model = model.to(device)

# Use the pipeline with your custom model
diarization = pipeline("audio.wav")

• Dataset Creation Guide
• Model Training Guide
• pyannote.audio Documentation