Presentation 2018-03-19
Stable Estimation Method of Spatial Correlation Matrices for Multi-channel NMF
Yuuki Tachioka,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Multi-channel non-negative matrix factorization (MNMF) achieves a high sound source separation performance but its initial value dependency is a problem. To address this, some studies proposed that initial spatial correlation matrices are obtained from enhanced speech after binary masking or from rank 1 approximation. Rank 1 approximation is effective because the number of parameters to be estimated is relatively small. However, the experimental comparison between initialization from full-rank spatial correlation matrices and rank 1 approximated ones is necessary. In addition, in the case of rank 1 approximation, when the number of channels is much greater than that of sound sources, the update of spatial correlation matrices becomes unstable. To avoid this, this paper proposes an utterance division for averaging spatial correlation matrices of each division with rank 1 approximation. Experiments show that the proposed method improved the performance of MNMF and stabilized the update of spatial correlation matrices.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) multi-channel NMF / spatial correlation matrix / steering vector / automatic speech recognition
Paper # EA2017-103,SIP2017-112,SP2017-86
Date of Issue 2018-03-12 (EA, SIP, SP)

Conference Information
Committee SIP / EA / SP / MI
Conference Date 2018/3/19(2days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English) Speech, Engineering/Electro Acoustics, Signal Processing, and Related Topics [SIP, EA, SP]/ Medical Image Engineering, Analysis, Recognition, etc. [MI]
Chair Masahiro Okuda(Univ. of Kitakyushu) / Suehiro Shimauchi(NTT) / Yoichi Yamashita(Ritsumeikan Univ.) / Kensaku Mori(Nagoya Univ.)
Vice Chair Shogo Muramatsu(Niigata Univ.) / Naoyuki Aikawa(TUS) / Mitsunori Mizumachi(Kyutech) / Hiroki Mori(Utsunomiya Univ.) / Yoshiki Kawata(Tokushima Univ.) / Yuichi Kimura(Kinki Univ.)
Secretary Shogo Muramatsu(Chiba Inst. of Tech.) / Naoyuki Aikawa(Takushoku Univ.) / Mitsunori Mizumachi(Akita Pref. Univ.) / Hiroki Mori(Shizuoka Inst. of Science and Tech.) / Yoshiki Kawata(Shizuoka Univ.) / Yuichi Kimura(Meijo Univ.)
Assistant Masayoshi Nakamoto(Hiroshima Univ.ひろ) / TREVINO Jorge(Tohoku Univ.) / Nobutaka Ito(NTT) / Kei Hashimoto(Nagoya Inst. of Tech.) / Satoshi Kobashikawa(NTT) / Ryo Haraguchi(Univ. of Hyogo) / Yasushi Hirano(Yamaguchi Univ.)

Paper Information
Registration To Technical Committee on Signal Processing / Technical Committee on Engineering Acoustics / Technical Committee on Speech / Technical Committee on Medical Imaging
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Stable Estimation Method of Spatial Correlation Matrices for Multi-channel NMF
Sub Title (in English)
Keyword(1) multi-channel NMF
Keyword(2) spatial correlation matrix
Keyword(3) steering vector
Keyword(4) automatic speech recognition
1st Author's Name Yuuki Tachioka
1st Author's Affiliation Denso IT Laboratory(Denso IT Lab)
Date 2018-03-19
Paper # EA2017-103,SIP2017-112,SP2017-86
Volume (vol) vol.117
Number (no) EA-515,SIP-516,SP-517
Page pp.pp.7-12(EA), pp.7-12(SIP), pp.7-12(SP),
#Pages 6
Date of Issue 2018-03-12 (EA, SIP, SP)