TY - GEN
T1 - Infinite sparse factor analysis for blind source separation in reverberant environments
AU - Nagira, Kohei
AU - Otsuka, Takuma
AU - Okuno, Hiroshi G.
PY - 2012
Y1 - 2012
N2 - 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.
AB - 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.
KW - Blind source separation
KW - Infinite sparse factor analysis
KW - Non-parametric Bayes
KW - Reverberant mixtures
UR - http://www.scopus.com/inward/record.url?scp=84868095510&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84868095510&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-34166-3_70
DO - 10.1007/978-3-642-34166-3_70
M3 - Conference contribution
AN - SCOPUS:84868095510
SN - 9783642341656
VL - 7626 LNCS
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 638
EP - 647
BT - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
T2 - Joint IAPR International Workshops on Structural and Syntactic PatternRecognition, SSPR 2012 and Statistical Techniques in Pattern Recognition,SPR 2012
Y2 - 7 November 2012 through 9 November 2012
ER -