Design and Analysis of Single Channel Speech Enhancement Techniques

Abstract

Speech is the only way of communication. When we communicate speech passes through a medium and it may be corrupted by many unwanted noises which degrade the actual speech. So because of this a person can not listen properly the background noises also create problem in many automatic speech processing tasks such as speech recognition hearing aid systems etc. Therefore denoising of corrupted single channel speech has become a very necessary and important aspect for research. There are many single channel speech improvement methods are available in past. Some of these perform better for one particular types of noise whereas others are suitable for other types of noise and also on large data set these method do not work well. So based on these limitations this research work has some algorithms. Firstly we designed the algorithm using different wavelet and decompose them in different frequency bin and using wiener filter for noise filtering in different frequency bin the given method shows that coiflet5 and Daudechies40 gives best results for speech enhancement as compare to other mother wavelets. In Chapter 4 single channel speech is enhanced by the application of spectral subtraction method using minimal statics Minimal statistic estimates the power spectrum of noise signal by avoiding the problem of detecting speech activity by finding the smallest value for smooth power spectrum of noisy signal and further we have designed the two machine learning algorithms VoSE and G Cocktail to solve the problem of speech enhancement and separation from a mixed signal. These algorithm can work for any language with or without a large Dataset . Result shows the the good improvement in the accuracy of the single channel speech signal. newline