Time-frequency blind signal separation: extended methods, performance evaluation for speech sources

Abstract

Most reported blind source separation (BSS) methods are based on independent component analysis (ICA), which esp. requires the sources to be stationary (and non-Gaussian). Time-frequency (TF) BSS methods avoid these restrictions and are therefore e.g. attractive for speech signals. We first introduce extensions of three types of TF-BSS methods that we recently proposed, and we analyze the relationships between these methods. We then provide a detailed benchmarking of these methods, based on a large number of tests performed with linear instantaneous mixtures of speech signals. This demonstrates the good performance of these methods (SNR typically above 60 dB) and their low sensitivity to the values of their TF parameters.