Dereverberation and Signal Separation of Speech Signal Mixtures

Abstract

This paper contributes to the dereverberation and signal separation problem of speech signal mixtures in reverberant environments by comparing the performance of different subband transform techniques, namely PolyPhase Filter Banks (PPFB) and weighted overlap-add short-term Fourier transform (WOLA STFT). The subband techniques allow large deconvolution problem to be subdivided into many smaller problems, which are feasible to solve. The critical finding is that the PPFB has better performance measures when compared with the WOLA STFT while having a lower computational complexity due to a lower subsampling rate. From the evaluation study, it can be seen that both the signal separation and the de-reverberation perform well even for high levels of background noise (SNR=0dB).