ONLINE BLIND SIGNAL SEPARATION OF SPEECH SIGNALS

Abstract

Separation of sources consists of recovering a set of signals of which only linear instantaneous mixture is observed is of prime importance for audio quality enhancement and speech recognition. In this thesis, analysis of various time-domain BSS (Blind Signal Separation) algorithms are performed and their performance compared. AMUSE (Algorithm for Multiple Unknown Source Extraction) algorithm based online BSS is implemented and its performance studied. The permutation problem of online BSS is tackled with a novel approach. The proposed method‟s performance parameters has been optimized with the help of GA (Genetic Algorithm) performed on polynomial regression, neural network and ANFIS (Adaptive Neuro Fuzzy Inference System) model for the algorithm. The scope of the thesis is limited to instantaneous mixtures for over determined case (the number of sensors recording signals greater than or equal to the number of source signals) with no noise.