Infinite sparse factor analysis for blind source separation in reverberant environments

Kohei Nagira; Takuma Otsuka; Hiroshi G. Okuno

doi:10.1007/978-3-642-34166-3_70

Infinite sparse factor analysis for blind source separation in reverberant environments

Kohei Nagira^*, Takuma Otsuka, Hiroshi G. Okuno

^*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

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.

Original language	English
Title of host publication	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Pages	638-647
Number of pages	10
Volume	7626 LNCS
DOIs	https://doi.org/10.1007/978-3-642-34166-3_70
Publication status	Published - 2012
Externally published	Yes
Event	Joint IAPR International Workshops on Structural and Syntactic PatternRecognition, SSPR 2012 and Statistical Techniques in Pattern Recognition,SPR 2012 - Hiroshima Duration: 2012 Nov 7 → 2012 Nov 9

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	7626 LNCS
ISSN (Print)	03029743
ISSN (Electronic)	16113349

Other

Other	Joint IAPR International Workshops on Structural and Syntactic PatternRecognition, SSPR 2012 and Statistical Techniques in Pattern Recognition,SPR 2012
City	Hiroshima
Period	12/11/7 → 12/11/9

Keywords

Blind source separation
Infinite sparse factor analysis
Non-parametric Bayes
Reverberant mixtures

ASJC Scopus subject areas

Computer Science(all)
Theoretical Computer Science

Access to Document

10.1007/978-3-642-34166-3_70

Cite this

Nagira, K., Otsuka, T., & Okuno, H. G. (2012). Infinite sparse factor analysis for blind source separation in reverberant environments. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 7626 LNCS, pp. 638-647). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7626 LNCS). https://doi.org/10.1007/978-3-642-34166-3_70

Infinite sparse factor analysis for blind source separation in reverberant environments. / Nagira, Kohei; Otsuka, Takuma; Okuno, Hiroshi G.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7626 LNCS 2012. p. 638-647 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7626 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Nagira, K, Otsuka, T & Okuno, HG 2012, Infinite sparse factor analysis for blind source separation in reverberant environments. in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). vol. 7626 LNCS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 7626 LNCS, pp. 638-647, Joint IAPR International Workshops on Structural and Syntactic PatternRecognition, SSPR 2012 and Statistical Techniques in Pattern Recognition,SPR 2012, Hiroshima, 12/11/7. https://doi.org/10.1007/978-3-642-34166-3_70

Nagira K, Otsuka T, Okuno HG. Infinite sparse factor analysis for blind source separation in reverberant environments. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7626 LNCS. 2012. p. 638-647. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-642-34166-3_70

Nagira, Kohei ; Otsuka, Takuma ; Okuno, Hiroshi G. / Infinite sparse factor analysis for blind source separation in reverberant environments. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7626 LNCS 2012. pp. 638-647 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{a57d456d89e44b1e8f6bff64ca76f673,

title = "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.",

keywords = "Blind source separation, Infinite sparse factor analysis, Non-parametric Bayes, Reverberant mixtures",

author = "Kohei Nagira and Takuma Otsuka and Okuno, {Hiroshi G.}",

year = "2012",

doi = "10.1007/978-3-642-34166-3_70",

language = "English",

isbn = "9783642341656",

volume = "7626 LNCS",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

pages = "638--647",

booktitle = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

note = "Joint IAPR International Workshops on Structural and Syntactic PatternRecognition, SSPR 2012 and Statistical Techniques in Pattern Recognition,SPR 2012 ; Conference date: 07-11-2012 Through 09-11-2012",

}

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 -

Infinite sparse factor analysis for blind source separation in reverberant environments

Abstract

Publication series

Other

Keywords

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this