Self-Tuning Spectral Clustering for Speaker Diarization

About

This repository contains the implementation for the paper:

• "Nikhil Raghav, Avisek Gupta, Md Sahidullah, and Swagatam Das, Self-Tuning Spectral Clustering for Speaker Diarization, to appear in Proc. of ICASSP 2025.

The details of the technique can be found here.

The poster can be accessed here.

🎥 YouTube Presentation: Watch the presentation here

Dependencies

Our implementation is based on a modified version of the AMI recipe provided in the SpeechBrain toolkit.

• Follow the installation guidelines for the SpeechBrain toolkit provided here

Method

The following three files were modified from the exisitng AMI recipe, and were adapted for the experiments on the DIHARD-III dataset. It contains, the scripts for the proposed SC-pNA technique:

• experiment.py located at /speechbrain/recipes/AMI/Diarization/experiment.py
• ecapa_tdnn.yaml located at /speechbrain/recipes/AMI/Diarization/ecapa_tdnn.yaml
• diarization.py located at /speechbrain/speechbrain/processing/diarization.py

Citation

If you find our approach useful in your research, please consider citing:

@INPROCEEDINGS{10890194,
  author={Raghav, Nikhil and Gupta, Avisek and Sahidullah, Md and Das, Swagatam},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Self-Tuning Spectral Clustering for Speaker Diarization}, 
  year={2025},
  volume={},
  number={},
  pages={1-5},
  keywords={Laplace equations;Costs;Clustering algorithms;Signal processing;Gaussian distribution;Acoustics;Computational efficiency;Sparse matrices;Speech processing;Tuning;speaker diarization;spectral clustering;matrix sparsification;eigengap;DIHARD-III},
  doi={10.1109/ICASSP49660.2025.10890194}}

License

This project is licensed under the MIT License. The full terms of the MIT License can be found in the LICENSE.md file at the root of this project.