Performance evaluation of single channel speech separation using non-negative matrix factorization

Abstract

Blind Source Separation (BSS) of underdetermined mixture has acquired a huge attention in signal processing environment, even though it is very much difficult to separate the underlying sources. The difficulty in source separation arise due to the mixing of large number of source signals in time and frequency, and propagation of it to one or more sensors through air. The objective in BSS is to identify the underlying source signals based on measurements of the mixed sources. Among many of the techniques used in BSS, due to its direct, easy to code and intuitive interpretability of basis and activation components, NMF provide an accurate form of parts-based representation of underlying data. Even though both supervised and unsupervised modes of operations are used in NMF, supervised mode performs well due to the use of pre-learned basis vectors corresponding to each underlying source. In this paper two of the multiplicative algorithms, Regularized Expectation Minimization Maximum Likelihood Algorithm (REMML) and Regularized Image Space Reconstruction Algorithm (RISRA) with sparseness constraint are taken to evaluate the performance of BSS. By the use of speech and music mixtures, Signal to Distortion Ratio (SDR), Signal to Interference Ratio (SIR) and Signal to Artifact Ratio (SAR) are evaluated.