Stable and Low-Distortion Algorithm Based on Overdetermined Blind Separation for Convolutive Mixtures of Speech

We propose a new algorithm with a stable learning and low­ distortion based on overdetermined blind separation for the convolutive mixture of the speech. To improve the separation performance， we have proposed multistage ICA， in which frequency-domain ICA and time­ domain ICA (TDICA) are cascaded. For temporally correlated signals， we must use TDICA with a nonholonomic constraint to avoid the decor­ relation effect. However， the stability cannot be guaranteed in the non­ holonomic c回e. Also， in the holonomic case， the sound quality of the separated signal is distorted by the decorrelation effect. To solve the problem of the stability， we perform TDICA with the holonomic con­ straint. To avoid the distortions， we estimate the distortion components by TDICA with the holonomic constraint ai1d we compensate the sound qualities by using the estimated components. The stability of the pro­ posed algorithm can be guaranteed by the holonomic constraint， and the proposed compensation work prevents the distortion. The experiments in a reverberant room reveal that the algorithm results in higher stability and higher separation performance. 1 Introduction Blind source separation (BSS) is an approach for estimating original source sig­ nals only from the information of the mixed signals observed in each input chan­ nel. This technique is applicable to high-quality hand-free speech recognition systems. Many BSS methods based on independent component analysis (ICA) [lJ have been proposed [2，3J for the acoustic signal separation. However， the performances of these methods degrade seriously， especially under heavily re­ verberant conditions. In order to improve the separation performance， we have proposed multistage ICA (MSICA) involving subarray processing [4]， in which frequency-domain ICA (FDICA) [3，5] and time-domain ICA (TDICA) [2，6] are cascaded (see Fig. 1). In this method，自rst， we divide the observed signals in a microphone array into the observed signals in the subarrays. In every subarray ，