Sequential Initialization Method with Increasing the Number of Channels for Blind Source Separation by Using Multichannel Nonnegative Matrix Factorization

Takanobu URAMOTO  Yuuki TACHIOKA  Tomohiro NARITA  Iori MIURA  Shingo UENOHARA  Ken'ichi FURUYA  

Publication
D - Abstracts of IEICE TRANSACTIONS on Information and Systems (Japanese Edition)   Vol.J101-D   No.3   pp.569-577
Publication Date: 2018/03/01
Online ISSN: 1881-0225
DOI: 
Type of Manuscript: Special Section PAPER (Special Section on Student Research)
Category: 
Keyword: 
sound source separation,  noise reduction,  nonnegative matrix factorization (NMF),  multichannel NMF,  initialization,  

Full Text(in Japanese): PDF(1.4MB)
>>Buy this Article


Summary: 
Nonnegative Matrix Factorization (NMF) factorizes a non-negative matrix into two non-negative matrices. In the field of acoustics, Multi-channel NMF (MNMF) has been proposed. MNMF utilizes spatial information by multichannel extension of NMF and can perform high-accurate sound source separation. However, the conventional MNMF tends to be trapped by local minima because their models have too many free parameters and this causes initial value dependencies of the separation performance. Prior studies mainly focused on two channels, but this paper applies MNMF for the case of three or more channels. Our music separation experiments showed that an increase in the number of channels did not improve the separation performance when their initial values were randomly set. Prior studies show that the separation performance of MNMF are highly dependent on the initial values of the spatial correlation matrix among four matrices that are used for a separation. Therefore, we proposed a method that sequentially sets a sub matrix of the initial spatial correlation matrix from that obtained at the fewer channels, when the number of channels increased. The separation performance was improved against the random setting, which shows the effectiveness of the proposed method.