Estimation for the Number of Sound Sources Using Directional Histogram of Observed Mixture Signals

Abstract

This paper proposes a number estimation method for sound source signals without information of sources and transfer functions. The proposed method can estimate the number of sound sources using only observed mixture signals with two microphones. A joint distribution of mixture signals has the same number of linear components as the number of sound sources. Therefore, we can estimate the source number using directional histogram of mixture signals. Furthermore, in order to be robust against outliers, the proposed method removes very small amplitude components without speech. The method works well from several simulations in the case of simultaneous utterances of multiple speakers.