Evaluation of Multichannel Speech Signal Separation using Independent Component Analysis

Multichannel speech signal separation using Independent Component Analysis (ICA) is not a now field in speech signal processing. However, maximization of output entropy as one of the measure of ICA is not well researched. Therefore, this paper uses maximum output entropy based basic Bell-Sejnowski's infomax theorem, embedded with gradient ascent algorithm for separating Malay language supported speech signal from two-talker competing speech. In addition to that maximum entropy is used as a convergence criterion. Results show that separated speech signals have high correlation with the original signals before mixing. The final r-value is high, r > 0.99 and this is tested by p-value p < 0.05 for four word and five word speech signals. Moreover, as the number of samples are increased for various speech signals, gradient ascent algorithm takes more time to compute thus indicating increased computational complexity.