Abstract:
A multichannel extension of non-negative matrix factorization (NMF) associates the spatial property of the sources with each of the NMF bases. An initial-value selection method based on log-likelihood for multichannel non-negative matrix factorization (MNMF) is introduced to reduce the variation of the source separation performance. Experimental results showed selecting initial values that provide high log-likelihood would improve the source separation performance of MNMF depending on the sources.