Gain estimation in model-based single channel speech separation

Abstract

In most current model based single channel separation techniques, it is assumed that the recording conditions are identical in the training phase and application phase. In this paper, we consider a general case in which training data and application data have different levels of energy and a technique is proposed to estimate the sources' gains which are required for the separation process. We use the periodogram of the speech signal as the selected feature for separation such that the sources' gains are estimated in terms of normalized periodograms of the sources and the mixture. The proposed technique is compared with a state-of-the-art technique which uses AR modeling of the speech signal and maximum likelihood for estimating gain and separating the sources. Experimental results show that our technique not only outperforms this technique in terms of SNR results and gain estimation accuracy but also reduces computational complexity.