Single channel subband blind source separation using temporal dependency of speech via viterbi algorithm

Abstract

This dissertation presents a method for achieving blind source separation when given only a single channel mixture. The main concept is base on assuming that the sources are quantized and we know the quantization levels previously, and the following quantization errors are considered as channel noise. The object of the proposed method is to separate sources up to the quantized signals. We define the transition probabilities between assumed quantized levels using the temporal dependency of speech signals by the fact that amplitudes of speech don\``t change suddenly. The sources are separated up to quantized signals by searching the best path in Viterbi algorithm with the cost function which is a function of quantization errors and transition probabilities. The proposed method is applied using subband decomposition to improve the separated speech quality. Experimental results exhibit a separation performance for simulated mixtures of two different sources both fullband and subbands.