A blind source separation approach based on IVA for convolutive speech mixtures

Here we present a new algorithm for the separation of convolutive speech observations using recordings from 2 microphones. This method is the union of independent vector analysis (IVA) and ideal binary mask (IBM), in conjunction with a post-filtering process in the cepstral domain. The proposed algorithm comprises of 3 steps. In the first step, an IVA algorithm is applied for the separation of the source signals from 2-microphone recordings. Second step is the estimation of IBM by the comparison of the energy of corresponding time-frequency (T-F) units of the segregated sources that are achieved using the IVA technique. Final step is the reduction of the musical noise by employing cepstral smoothing and such a noise is generated due to T-F masking. The signal to noise ratio (SNR) measurement has been used to evaluate the overall performance of the proposed method by employing the reverberant mixtures that are produced via simulated room model. The evaluation shows that it is more efficient and speech quality has been improved while generating similar segregation performance compared to a state-of-the-art approach.