In this paper we present a new approach for estimating the number of active speech sources in the presence of interfering noise sources and reverberation. First, a binaural front-end is used to detect the spatial positions of all active sound sources, resulting in a binary mask for each candidate position. Then, each candidate position is characterized by a set of features. In addition to exploiting the overall spectral shape, a new set of mask-based features is proposed which aims at characterizing the pattern of the estimated binary mask. The decision stage for detecting a speech source is based on a support vector machine (SVM) classifier. A systematic analysis shows that the proposed algorithm is able to blindly determine the number and the corresponding spatial positions of speech sources in multisource scenarios and generalizes well to unknown acoustic conditions
展开▼
