A novel approach for blind separation dereverberation of speech mixtures using multiplestep linear predictive coding

Abstract

A new method for the combination of blind separation and dereverberation of speech signals using linear convolutive mixing model is presented. The proposed algorithm consists of two parts. In the first part pre-filtering process is applied on speech mixtures to predict late reverberations by employing long-term multiple-step linear prediction (MSLP) and then these late reverberations are mitigated by using spectral subtraction (SS) technique. In the second part, a source separation technique has been applied consisting of various steps. Here in this part, first Independent component analysis (ICA) algorithm is used to separate target speech sources from sensor readings using the assumptions that sources involved in the mixing process are independent. Then by differentiating energy of individual time-frequency signatures of the separated target speech signals we compute ideal binary mask (IBM). Finally artifacts are suppressed which are normally the basis of time varying nature of IBM by means of cepstral smoothing. Simulated environment for reverberant mixtures is used to analyse the efficiency of our proposed algorithm. Simulations results evaluated in terms of signal to noise ratio (SNR) indicate a considerably enhanced quality of segregated speech as compared to a previous method.