Infinite sparse factor analysis for blind source separation in reverberant environments

Abstract

Sound source separation in a real-world indoor environment is an ill-formed problem because sound source mixing is affected by the number of sounds, sound source activities, and reverberation. In addition, blind source separation (BSS) suffers from a permutation ambiguity in a frequency domain processing. Conventional methods have two problems: (1) impractical assumptions that the number of sound sources is given, and (2) permutation resolution as a post processing. This paper presents a non-parametric Bayesian BBS called permutation-free infinite sparse factor analysis (PF-ISFA) that solves the two problems simultaneously. Experimental results show that PF-ISFA outperforms conventional complex ISFA in all measures of BSS-EVAL criteria. In particular, PF-ISFA improves Signal-to-Interference Ratio by 14.45 dB and 5.46 dB under RT 60∈=∈30 ms and RT 60∈=∈460 ms conditions, respectively. © 2012 Springer-Verlag Berlin Heidelberg.