Speech segregation using constrained ICA

Abstract

In natural environment, speech often occurs concurrently with acoustic interference. How to effectively extract speech remains a great challenge. This paper describes a novel constrained Independent Component Analysis (ICA) approach, the ICA with reference (ICA-R), to speech segregation. Different from the traditional ICA which recovers simultaneously all the source signals, the ICA-R extracts only some desired source signals from the mixtures of source signals by incorporating some a priori information into the separation process. We show how the ICA-R can be applied to separate a target speech signal from interfering sounds by exploiting a proper reference signal, which is based on the different characteristic between speech signal and its environmental noises, i.e., the speech signal has pitch and its harmonic frequencies whereas the noises usually do not. Results of computer experiments demonstrate the efficiency of the proposed method. © Springer-Verlag 2004.