Source number estimation based on clustering of speech activity sequences for microphone array processing

In this paper we introduce a novel technique for source number estimation based on the clustering of speech activity sequences. Speech activity sequences, represented by posterior probability time series of speech activity, are modeled as a mixture of Watson distributions. To enable source number estimation, an adaptive Dirichlet prior probability is imposed upon the mixture weights to promote the formation of empty clusters. The proposed source number estimation technique was evaluated on reverberant over-, even- and under-determined settings with accuracy between 75% and 100%.