Certain investigations on speech enhancement algorithms with performance analysis

Abstract

The interest in the field of speech enhancement emerges from newlinethe increased usage of digital speech processing applications like speech newlinerecognition mobile telephony digital hearing aids and humanmachine newlinecommunication systems in our daily life The trend to make these newlineapplications mobile increases the variety of potential sources for quality newlinedegradation Speech enhancement methods are used to increase the quality newlineof these speech processing devices and make them more robust under noisy newlineconditions newlineIn this research speech enhancement algorithms under dual newlinemicrophone and single microphone conditions are reviewed and algorithms newlineare proposed to improve the performance of the existing methods Further newlinethe application of these methods as a preprocessing stage in digital hearing newlineaid is studied and verified that the performance of the system is improved newlineIn dual channel speech enhancement algorithms reference newlinesignal is available in addition to the noisy input signal For these methods newlineadaptive filters are used where a noisy input signal and a reference signal newlineare given as inputs and the output is the enhanced signal using the weight newlineupdation algorithms In this work a conventional adaptive filter is taken newlineand it is identified that the performance of the filter is improved when it is newlineoperated in frequency domain inputs rather than time domain inputs Least newlineMean Square LMS and Recursive Least Square RLS algorithms are in newlineuse for the proposed adaptive filters for the weight updation In this work a newlinedual transformation known as discrete cosine transform and Hadamard newlinetransform are applied to the input signals before they are given to the filter newlineThe performance of the proposed dual channel algorithms are newlineanalyzed based on the objective and subjective quality measures The proposed dual transformed LMS and RLS adaptive algorithms provide newlinebetter results in contrast to the signals operated in time domain Compared newlineto the conventional LMS and RLS algorithm with DCT alone as a newlinepreprocessing on an average of 15 to 24 dB raise in Signal to N