In this study, we emphasized sound sources using nonnegative matrix factorization (NMF) with the likelihood of the directions obtained from the estimated localizations. NMF is one of the most commonly used techniques for the separation of music signals. The emphasis of sound sources by NMF has several problems: the uncertainty of the decomposition of the bases corresponding to the sources and the necessity to group the obtained bases according to the sources. These problems have to be solved simultaneously to improve the separation performance. In order to simultaneously solve these problems, we proposed a novel NMF incorporating the cost based on the likelihood of the localizations. According to the experimental results, we could improve the separation performance by using our proposed method.
2014 Research Institute of Signal Processing, Japan