Independent Component Analysis for Separation of Speech Mixtures: A Comparison Among Thirty Algorithms

Abstract

Vast number of researches deliberated the separation of speech mixtures due to the importance of this fieldof research. Whereas its applications became widely used in our daily life; such as mobile conversation, videoconferences, and other distant communications. These sorts of applications may suffer from what is well known thecocktail party problem. Independent component analysis (ICA) has been extensively used to overcome this problem andmany ICA algorithms based on different techniques have been developed in this context. Still coming up with somesuitable algorithms to separate speech mixed signals into their original ones is of great importance. Hence, this paperutilizes thirty ICA algorithms for estimating the original speech signals from mixed ones, the estimation process iscarried out with the purpose of testing the robustness of the algorithms once against a different number of mixed signalsand another against different lengths of mixed signals. Three criteria namely Spearman correlation coefficient, signalto interference ratio, and computational demand have been used for comparing the obtained results. The results of thecomparison were sufficient to signify some algorithms which are appropriate for the separation of speech mixtures.