A block-based blind source separation approach with equilateral triangular microphone array

In this paper we describe a method for multiple speech sources separation using an equilateral triangular microphone array. Firstly, the azimuths of horizontal plane are divided into many units and the spatial features of some directions observed by the microphone array are modeled precisely. Secondly, the input mixing signals are segmented into blocks, and then the number of active speakers and their directions are estimated in each block. Thirdly, the pre-trained model with the nearest azimuth to each speaker is adapted to obtain a precise model, which is then used for time-frequency binary mask estimation. Finally, we separate every source appeared in each block and concatenate those sounds from same unit to reproduce the whole stream. The experiments are set up in a real meeting room. The results show that our method can separate multiple speech sources correctly with low distortion, and are competitive with the total un-blind separation results.