Research On Noisy Speech Enhancement Techniques

At present, speech recognition (SR) technology has become more and more mature, and has been applied in more and more areas. But the SR systems that are well trained in laboratory environment, if deployed in a different acoustic environment, especially in intensive noisy environment, its performance would be rapidly deteriorating. Speech enhancement may be the primary answer to this problem. If speech signal being processed by some speech enhancement techniques before sent to SR system, the system could retain nearly as high recognition rate as in laboratory. This thesis focuses researches on noisy speech enhancement techniques.Firstly, in this thesis, the principles of some typical classic speech enhancement methods are introduced. These methods include spectral subtraction, Wiener filtering, adaptive filtering and auditory masking effects method. Of these methods, the emphases are spectral subtraction and wiener filtering, these two methods and their improved versions are carefully analyzed and experimented. These experiment results are evaluated by both of subjective and objective criteria. Their performances and enhancement effects are contrasted with each other. Their respective advantages and disadvantages are also listed.Secondly, independent component analysis(ICA) and its applications in speech denoising are heavily investigated. The theory of ICA, popular fixed-point algorithm and selection of criterion functions are reviewed. The fixed-point algorithms based on kurtosis and negative entropy are labored. By simulation experiment and calculation, these two algorithms are carried out to get results of respective separation effects and overall performances. The results are evaluated from both of subjective and objective perspectives. Single channel ICAs are also applied to speech enhancement. One of these Single channel ICAs, signal dichotomy method is implemented and validated. Single channel ICAs extend the applications of ICA and also can reduce the cost of ICA algorithm implementation.Thirdly, we propose a new algorithm based on spectral subtraction and ICA which being applicable in intensive noisy environments: (1)Apply ICA to separate single-channel speech signal and noise signal which embedded in measured noisy speech signal for increasing SNR; (2)Subtracted noise signal which got at previous step from measured noisy speech signal by spectral subtraction method for higher SNR. Based upon above techniques, the separations of multiple-channels speech signals and noise signal are also experimented. The results show these techniques are effectual with obvious noise elimination effects, however, speech separation of multiple-channels speech signals is far from perfection.Finally, some conclusions are derived, and also some proposals for further researches are given.