Abstract:
Index Terms single channel speech enhancement; adverse environment; additive noise; noise newlineestimation; short-time Fourier transform; stationary wavelet packet filterbank; critical band rate scale; newlinespectral subtractive-type algorithms; iterative processing; spectrogram; objective measure; subjective newlinemeasure. newlineThe degradation of speech due to the presence of additive background noise causes severe newlineproblems in a variety of communication environments. The spectral subtraction method is a newlineclassical approach and is widely used for enhancement of degraded speech. The drawback of this newlineclassical approach is that it uses a fixed set of values for the subtraction parameters. Thus, its newlineapplications are limited only for the noises that degrade the speech signal uniformly. However, newlinereal-world noise is mostly non-stationary or colored in nature and a multi-band approach is found newlineto be more efficient than the classical approach. newlineThis thesis mainly addresses the problem of single channel speech enhancement, where newlinethe signal is derived from a single microphone, in adverse environment and proposes transform newlinebased multi-band speech enhancement algorithms. The aim of the proposed research work is to newlineexplore the transform based multi-band algorithms for the augmentation of the overall quality and newlineintelligibility of the processed speech by suppressing the background noise as well as the remnant newlinecomponent of the noise. The algorithms proposed in the thesis are based on short-time Fourier newlinetransform and stationary wavelet packet transform and the noise reduction is performed on the newlinetransform coefficients by using adaptive noise estimation approach. The performance of the newlinealgorithms has been evaluated comprehensively and their comparative study has been done. newlineThe thesis presents a detailed study of the spectral subtractive-type algorithms with a newlineunified view of the single channel speech enhancement algorithms in the frequency domain, newlinewhich provides the necessary algorithmic framework required for the development of the newlineproposed transform.
