A Complex Matrix Factorization approach to Joint Modeling of Magnitude and Phase for Source Separation

TL;DR

This paper introduces a complex matrix factorization method that jointly models magnitude and phase for improved source separation, leading to better quality speech reconstruction.

Contribution

It presents a novel approach that incorporates spectral phase into matrix factorization, transforming the complex problem into an NMF framework for enhanced separation.

Findings

Improved source separation quality demonstrated on GRID corpus.

Joint modeling of magnitude and phase reduces interference artifacts.

Objective evaluations show significant performance gains.

Abstract

Conventional NMF methods for source separation factorize the matrix of spectral magnitudes. Spectral Phase is not included in the decomposition process of these methods. However, phase of the speech mixture is generally used in reconstructing the target speech signal. This results in undesired traces of interfering sources in the target signal. In this paper the spectral phase is incorporated in the decomposition process itself. Additionally, the complex matrix factorization problem is reduced to an NMF problem using simple transformations. This results in effective separation of speech mixtures since both magnitude and phase are utilized jointly in the separation process. Improvement in source separation results are demonstrated using objective quality evaluations on the GRID corpus.