An efficient VQ-based method for monaural speech separation

We present a new model-based monaural speech separation technique for separating two speech signals when only a single recording of their linear mixture is available. Two important aspects of model-based monaural speech separation are the applied modeling technique and the estimation technique. In this approach, we introduce sub-section vector quantization technique and use it as the modeling technique instead of conventional vector quantization method. Then, separated speech signals are estimated using a simple soft mask filter whose states are controlled by the components of the codevectors. In the speech separation experiments, the proposed method is shown to improve SNR by 1.8 dB compared to the system using conventional VQ technique as the modeling technique and binary mask filter as the estimator.