Most speech enhancement systems suffer from musical residual noise when a system enhances a noisy speech signal corrupted by various kinds of colored noise. The musical residual noise sounds very annoying to the human ears, and is caused by the spectral peaks of residual noise that randomly appear over subbands of a frame and in successive frames. This project aims to reduce the effect of musical residual noise. In order to analyze the properties of residual noise in enhanced speech, we have to collect the residual noise data. Therefore, some novel speech enhancement algorithms should be implemented to achieve residual noise. The enhancement algorithms include spectral domain, wavelet domain and signal subspace et al. Hence, we try to design a post processing system to reduce the amount of residual noise for a speech enhancement system by analyzing the spectral variation contour of each subband in successive frames. The enhanced spectra are categorized as either speech dominant or noise dominant by analyzing the motion property for each frequency bin. In the case of noise-dominant spectral bin, this bin should be suppressed to reduce the residual noise. Conversely, this spectral bin is speech dominated and should be preserved to maintain speech quality. Moreover, the speech quality major depends on the vowel signal in an utterance for mandarin Chinese spoken language. We will also try to reconstruct the harmonics of a vowel signal based on the inter-frame spectral variation being smooth and the spectral correlation being high among the neighbors of a spectral bin. The improved speech quality can be obtained by reconstructing the harmonics which were removed by a speech enhancement system or destroyed by a corrupting noise.
