Developing optimized algorithm for speech dereverberation and source separation

Abstract

Blind source separation is a significant problem since the signal to newlinebe separated is unknown. The observed signal is a mixture of signals such as newlinemusic, interferences, reverberation, and noises without any prior knowledge newlineabout the original source signals to be separated. To achieve this objective, newlinepreviously the researchers performed the blind source separation and blind newlineDereverberation separately and hence the target speech signal is not accurate. newlineMoreover, several researchers mutually combined the BSS and BD newlinetechniques, however, the optimized blind source signal separation is yet to be newlineobtained. The effectiveness of the work is measured by some parameters such newlineas signal to interference ratio (SIR), Double Data Rate (DRR), target signal newlinepreservation, global quality, quality in terms of other (interfering) signal newlinesuppression (OSS). To accomplish effective measurement of those values, newlinethis thesis proposes two novel approaches in this thesis which are specified newlinebelow. newlineIn Phase I, a novel mutually optimal approach is proposed newlinecombining two techniques: Principal Component Analysis (PCA) based on newlineLocally Weighted Projection Regression (LWPR) and Weighted Prediction newlineError (WPE) based on Deep Neural Network (DNN) (WPE). The sample newlineobserved signals are first pre-processed by utilizing FFT and whitening newlineapproaches. Then the reverberation signals are removed with the DNN-WPE newlineapproach and the resultant signal is subjected to the LWPR-PCA approach to newlineremove the existing noises. This method effectively separates the blind signal newlinefrom the observed signal. newline newline