Performance of quadratic time-frequency distributions in blind source separation of speech signals

Abstract

Blind source separation (BSS) based on time-frequency distributions (TFD) has certain advantages over other methods, such as allowing the separation of Gaussian sources with identical spectral shape but different t-f localization properties. However, quadratic TFD, which are the most important class of TFD suitable for this approach, differ widely in resolution and ability to reduce the cross-terms that disturb the signal interpretation. This paper investigates the performance of different time-frequency distributions on the blind separation of speech signals and nonlinear FM signals. We also discuss the possibility and the suitability of the solution to convolutive mixture of the speech signals using time-frequency based blind source separation because of its ability to process complex-valued signals.