This paper presents a robust sound source separation method for binaural sound mixture in a reverberant environment. The separation method that uses harmonic structure and sound source direction as clues, suffers severe degradation of performance as the reverberation increases. This is because fundamental frequency (F0) and sound source direction is made vague by the reverberation. To cope with these problems, we first investigate how the reverberation affects the estimation of these attributes, and propose a method to robustly estimate them in a reverberant environment. Experimental results showed that the present method could detect more than 90% of sound sources in correct directions from sound mixtures composed of three simultaneous utterances when the reverberation time is 0.5 of a second. The F{sub}0 estimation accuracy as well as the spectrum distortion of separated sound were also shown to be improved.
展开▼