Speech Enhancement Techniques

Abstract

Speech enhancement is concerned with improving some perceptual aspect of speech that has been degraded by additive noise. In most application, the aim of speech enhancement is to improve the quality and intelligibility of degraded speech. The NOIZEUS dataset is used to apply different speech enhancement algorithm. There are various speech enhancement methods available in literatures which are applica- tion specific. Speech enhancement algorithms are divided manly into three categories such as statistical model based, spectral subtractive and Subspace algorithm based methods. The objective of the thesis is to identify limitations of these algorithms and com- pare results of standard speech enhancement techniques available in literature which includes wiener a priori SNR method, wavelet thresholding method using multitapper spectrum, log MMSE estimator method, spectral subtraction method and multiband spectral subtraction method for speech enhancement. It has been observed from the simulation results that log MMSE and multiband spectral subtraction algorithm outperforms than other compared algorithm. The comparative results of these algo- rithms in terms of different objective and subjective quality evaluation parameter is presented in this thesis. And from all these result, we see that log MMSE and multi- band spectral subtraction algorithms perform best in almost all noise environments at different input SNR.