A Blind Separation of Monaural Sound Based on Peak Tracking of Frequency Spectra

This paper describes a blind separation algorithm of monaural sound based on peak tracking of frequency spectra. We have already reported a blind separation method based on the change ratio of frequency components. However, it cannot handle a signal with frequency fluctuation such as a speech signal or a vibrato tone, because such type of signal is regarded as the mixture of different sounds. Our new method proposed in this paper can handle a sound with frequency fluctuation by tracking frequency peaks along time axis. The effectiveness of the proposed method is evaluated with some experiments on real voice data.