Estimation of the Number of Sources in Measured Speech Mixtures with Collapsed Gibbs Sampling

In blind source separation (BSS), the number of sources present in the measured speech mixtures is unknown. The focus of this work is therefore to automatically estimate the number of sources from binaural speech mixtures. Collapsed Gibbs sampling (CGS), a Markov chain Monte Carlo (MCMC) technique,...

Full description

Saved in:
Bibliographic Details
Published in2017 Sensor Signal Processing for Defence Conference (SSPD) pp. 1 - 5
Main Authors Yang Sun, Yang Xian, Pengming Feng, Chambers, Jonathon, Naqvi, Syed Mohsen
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.12.2017
Subjects
Bayes methods
Blind source separation
Estimation
Microphones
Spectrogram
Speech
Online AccessGet full text

Cover

More Information
Summary:In blind source separation (BSS), the number of sources present in the measured speech mixtures is unknown. The focus of this work is therefore to automatically estimate the number of sources from binaural speech mixtures. Collapsed Gibbs sampling (CGS), a Markov chain Monte Carlo (MCMC) technique, is used to obtain samples from the joint distribution of the speech mixtures. Then the Chinese Restaurant Process (CRP) within the framework of the Dirichlet Process (DP) is exploited to cluster samples into different components to finally estimate the number of speakers. The accuracy of the proposed method, under different reverberant environments, is evaluated with real binaural room impulse responses (BRIRs) and speech signals from the TIMIT database. The experimental results confirm the accuracy and robustness of the proposed method.
DOI:10.1109/SSPD.2017.8233232