Abstract
We explore the permutation problem of frequency domain blind source separation (BSS). Based on performance analysis of three approaches: exploiting spectral continuity, exploiting time envelope structure and beamforming alignment; we present a new hybrid method which incorporates a psychoacoustic filtering process for the misaligned permutations unable to be delt with by these approaches. We use a sub-space based method (MUSIC) rather than conventional beamforming for the accurate estimation of the direction of arrivals (DOAs) of the source components, and a frequency dependent distance for the correlation of time envelopes. The proposed methods are compared with other approaches by signal to interference ratio (SIR) evaluation, and the new hybrid approach is shown to have the best performance. © Springer-Verlag 2004.