Publication Details

Bayesian HMM based x-vector clustering for Speaker Diarization

DIEZ SÁNCHEZ, M.; BURGET, L.; WANG, S.; ROHDIN, J.; ČERNOCKÝ, J. Bayesian HMM based x-vector clustering for Speaker Diarization. In Proceedings of Interspeech. Proceedings of Interspeech. Graz: International Speech Communication Association, 2019. p. 346-350. ISSN: 1990-9772.
Czech title
Bayesovské shlukování x-vektorů založené na HMM pro diarizaci
Type
conference paper
Language
English
Authors
URL
Keywords

Speaker Diarization, Variational Bayes, HMM, x-vector, DIHARD

Abstract

This paper presents a simplified version of the previously proposed diarization algorithm based on Bayesian Hidden Markov Models, which uses Variational Bayesian inference for very fast and robust clustering of x-vector (neural network based speaker embeddings). The presented results show that this clustering algorithm provides significant improvements in diarization performance as compared to the previously used Agglomerative Hierarchical Clustering. The output of this system can be further employed as an initialization for a second stage VB diarization system, using frame-wise MFCC features as input, to obtain optimal results.

Published
2019
Pages
346–350
Journal
Proceedings of Interspeech, vol. 2019, no. 9, ISSN 1990-9772
Proceedings
Proceedings of Interspeech
Publisher
International Speech Communication Association
Place
Graz
DOI
UT WoS
000831796400070
EID Scopus
BibTeX
@inproceedings{BUT159992,
  author="Mireia {Diez Sánchez} and Lukáš {Burget} and Shuai {Wang} and Johan Andréas {Rohdin} and Jan {Černocký}",
  title="Bayesian HMM based x-vector clustering for Speaker Diarization",
  booktitle="Proceedings of Interspeech",
  year="2019",
  journal="Proceedings of Interspeech",
  volume="2019",
  number="9",
  pages="346--350",
  publisher="International Speech Communication Association",
  address="Graz",
  doi="10.21437/Interspeech.2019-2813",
  issn="1990-9772",
  url="https://www.isca-speech.org/archive/Interspeech_2019/pdfs/2813.pdf"
}
Files
Back to top