Simultaneous speech source separation and noise reduction via clustering and MMSE-based filtering

We propose a new multichannel approach for simultaneous blind source separation (BSS) and acoustic noise reduction. In this approach, we first cluster the mixtures of sounds into N + 1 clusters representing the N (≥ 1) speech sources of interest in addition to the acoustic noise. We estimate the posterior probabilities of these clusters using the expectation maximization (EM) algorithm. Then, we use these probabilities to calculate the noise and speech source statistics separately. Finally, we formulate the multichannel minimum-mean-square-error-based (MMSE) filter that extracts each of the N speech source signals and reduces the additive noise based on these statistics. Experimental results with two speech sources and additive background noise are provided to demonstrate the effectiveness of the proposed approach.