Source separation using multiple directivity patterns produced by ICA-based BSS

In this paper, we propose a multistage source separation method constructed by combining blind source separation (BSS) based on independent component analysis (ICA) and segregation using multiple directivity patterns (SMDP) introduced in our previous paper. We obtain the directivity patterns needed in SMDP by ICAbased BSS. In the SMDP, simultaneous equations of amplitudes of sound sources are generated by using these multiple directivities. The solution of these equations gives good disturbance estimates. We apply spectral subtraction using these disturbance estimates and the speech enhancement of the target source is performed. We conducted experimentation in a real room in the source-number-given condition where there is no priori information about the sound sources and the characteristics of room acoustics. The experimental results of double talk recognition show that the proposed technique is effective in reducing the error rate by 30% compared to frequency domain BSS.