Optimizing Bayesian Hmm Based X-Vector Clustering for the Second Dihard Speech Diarization Challenge

Optimizing Bayesian Hmm Based X-Vector Clustering for the Second Dihard Speech Diarization Challenge

Paper in proceedings (conference paper)

This paper presents an analysis of our diarization systemwinning the second DIHARD speech diarization challenge,track 1. This system is based on clustering x-vector speakerembeddings extracted every 0.25s from short segments of theinput recording. In this paper, we focus on the two x-vectorclustering methods employed, namely Agglomerative HierarchicalClustering followed by a clustering based on BayesianHidden Markov Model (BHMM). Even though the systemsubmitted to the challenge had further post-processing steps,we will show that using this BHMM solely is enough toachieve the best performance in the challenge. The analysiswill show improvements achieved by optimizing individualprocessing steps, including a simple procedure to effectivelyperform "domain adaptation" by Probabilistic LinearDiscriminant Analysis model interpolation. All experimentsare performed in the DIHARD II evaluation framework.

This paper presents an analysis of our diarization systemwinning the second DIHARD speech diarization challenge,track 1. This system is based on clustering x-vector speakerembeddings extracted every 0.25s from short segments of theinput recording. In this paper, we focus on the two x-vectorclustering methods employed, namely Agglomerative HierarchicalClustering followed by a clustering based on BayesianHidden Markov Model (BHMM). Even though the systemsubmitted to the challenge had further post-processing steps,we will show that using this BHMM solely is enough toachieve the best performance in the challenge. The analysiswill show improvements achieved by optimizing individualprocessing steps, including a simple procedure to effectivelyperform "domain adaptation" by Probabilistic LinearDiscriminant Analysis model interpolation. All experimentsare performed in the DIHARD II evaluation framework.

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

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

DIEZ SÁNCHEZ, M.; BURGET, L.; LANDINI, F.; WANG, S.; ČERNOCKÝ, J.

2021

04.05.2020

IEEE Signal Processing Society

Barcelona

978-1-5090-6631-5

ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings

6519

6523

5

https://ieeexplore.ieee.org/document/9053982

diez_icassp2020_09053982