Effect of frequency oversampling and cascade initialization on permutation control in frequency domain BSS
Christopher Osterwise, Steven L. Grant · 2012
This paper presents a new algorithm for addressing the permutation ambiguity in convolutive blind source separation. The proposed algorithm seeks to prevent permutations by frequency oversampling, and then exploiting the induced correlation between bins. Any remaining permutation is then corrected by beam pattern estimation. Cascade initialization is shown to improve system performance while decreasing processing time, while frequency oversampling is shown to increase performance with slight increases in computational time. The algorithm is proven robust against isotropic noise, producing a SIR improvement of 12.9 dB when the reverberation time is 260 ms, and the signals at the input are only 5 dB above the noise floor.