An Improved State Coherence Transform Algorithm for the Location of Dual Microphone with Multiple Sources

This paper proposes a new kernel function in state coherence transform to perform multiple time difference of arrival estimation in order to increase the resolution of location in frequency-domain blind source separation. The state coherence transform associated with each source generalizes the GCC for multiple sources and generates envelopes with clear peaks corresponding to the maximum-likelihood TDOAs. However, the weight allocation of the kernel function is unreasonable for small spacing microphones. We propose an improved kernel function to enhance the resolution of small values, which means that a larger weight allocated to smaller values. Experimental results show that the proposed approach allows to separate four speakers, using very short utterances, in highly reverberant environment even with small-spaced microphones of 2 cm.