Catalog-based single-channel speech-music separation with the Itakura-Saito divergence

Abstract

In this study, we introduce a catalog-based single-channel speech-music separation method with the Itakura-Saito (IS) divergence measure. Previously, we have developed the catalog-based separation method with the Kullback-Leibler (KL) divergence. In the probabilistic point of view, IS divergence corresponds to a complex Gaussian observation model. Comparison of divergence measures or observation models in speech-music separation task is carried out with both of catalog-based and traditional Non-Negative Matrix Factorization (NMF) methods. The separation performance is compared using Speech-to-Music Ratio (SMR), Speech-to-Artifact Ratio (SAR) and speech recognition performance measure via the Word Error Rate (WER). We showed that, using IS divergence in both of catalog-based or NMF based speech-music separation methods yields better separation performance than KL divergence. Moreover, in this study, it is shown that catalog-based approaches with both divergence measures outperform traditional NMF based approaches in speech recognition experiments.