Time-Domain Blind Audio Source Separation Using Advanced Component Clustering and Reconstruction

We present a novel time-domain method for blind separation of convolutive mixture of audio sources (the cocktail party problem). The method allows efficient separation with good signal-to-interference ratio (SIR) and signal-to-distortion ratio (SDR) using short data segments only. In practice, we are able to separate 2-4 speakers from audio recording of the length less than 6000 samples, which is less than 1 s in the 8 kHz sampling. The average time needed to process the data with filter of the length 20 was 2.2 seconds in Matlab v. 7.2 on an ordinary PC with 3 GHz processor.