Blind Source Counting and Separation with Relative Harmonic Coefficients

Blind source separation aims to recover a series of unmixed signals from only a mixture recording of source signals. For the mixture of speech signals where sources have unique spatial positions, the relative transfer function has been found useful as it contains the spatial information of each source regardless of its signal. In this paper, we transfer a previous relative transfer function based separation method into the wave domain by utilizing a higher-order microphone for the mixture recording. Consequently, the higher-order microphone enables the processing with relative harmonic coefficients, which provides better independency between different coefficient components than the conventional relative transfer functions in each time frame. We demonstrate that the proposed method using relative harmonic coefficients outperforms the relative transfer function based method in highly reverberant rooms through simulation with both simulated and real measured room impulse response data.