Codebook Approaches for Single Sensor Speech/Music Separation

Abstract

The work presented is this chapter is an introduction to the subject of single sensor source separation dedicated to the case of speech/music audio mixtures. Approaches related in this study are all based on a full (Bayesian) probabilistic framework for both source modeling and source estimation. We first present a review of several codebook approaches for single sensor source separation as well as several attempts to enhance the algorithms. All these approaches aim at adaptively estimating the optimal time-frequency masks for each audio component within the mixture. Three strategies for source modeling are presented: Gaussian scaled mixture models, codebooks of autoregressive models, and Bayesian non-negative matrix factorization (BNMF). These models are described in details and two estimators for the time-frequency masks are presented, namely the minimum mean-squared error and the maximum a posteriori. We then propose two extensions and improvements on the BNMF method. The first one suggests to enhance discrimination between speech and music through multi-scale analysis. The second one suggests to constrain the estimation of the expansion coefficients with prior information. We finally demonstrate the improved performance of the proposed methods on mixtures of voice and music signals before conclusions and perspectives.