Paper Abstract and Keywords |
Presentation |
2020-03-02 10:10
Multichannel NMF with Joint-Diagonalizable Constraint Based on Generalized Gaussian Distribution for Blind Source Separation Keigo Kamo, Yuki Kubo, Norihiro Takamune (UTokyo), Daichi Kitamura (NIT Kagawa), Hiroshi Saruwatari (UTokyo), Yu Takahashi, Kazunobu Kondo (Yamaha) EA2019-103 SIP2019-105 SP2019-52 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
Multichannel nonnegative matrix factorization (MNMF) is a blind source separation technique, which employs the full-rank spatial covariance matrices and can simulate the situations where the reverberation is strong and the sources are not point sources. Source signals' spectrograms were assumed to follow a multivariate complex Gaussian distribution in MNMF. In this paper, we propose the model extension of MNMF to a multivariate complex generalized Gaussian distribution and derive a new parameter update rule using the auxiliary-function-based method, especially in the sub-Gaussian model. Since the cost function of MNMF of this multivariate complex generalized Gaussian model is hard to minimize, we additionally introduce the joint-diagonalizable constraint, which is the same one of FastMNMF, to MNMF, and transform the cost function to the form to which we can apply the auxiliary functions, deriving the valid parameter update rules. From blind source separation experiments, we show that the proposed method outperforms the conventional methods in source-separation accuracy. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
blind source separation / spatial covariance model / joint diagonalization / multivariate complex sub-Gaussian distribution / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 119, no. 439, EA2019-103, pp. 13-19, March 2020. |
Paper # |
EA2019-103 |
Date of Issue |
2020-02-24 (EA, SIP, SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
EA2019-103 SIP2019-105 SP2019-52 |
|