Binary mask estimation for voiced speech segregation using Bayesian method

Abstract

The ideal binary mask (IBM) estimation has been set as the computational goal of Computational auditory scene analysis (CASA). A lot of effort has been made in the IBM estimation via statistical learning method. The current Bayesian methods usually estimate the mask value of each time-frequency (T-F) unit independently with only local auditory features. In this paper, we propose a new Bayesian approach. First, a set of pitch-based auditory features are summarized to exploit the inherent characteristics of the reliable and unreliable time-frequency (T-F) units. A rough estimation is obtained according to Maximum Likelihood (ML) rule. Then, we propose a prior model which is derived from onset/offset segmentation to improve the estimation. Finally, an efficient Markov Chain Monte Carlo (MCMC) procedure is applied to approach the maximum a posterior (MAP) estimation. Proposed method is evaluated on Cooke's 100 mixtures and compared with previous model. Experiments show that our method performs better.