Speech Enhancement For Environment Of Lecture Hall

Speech signal in our report and lecture which recordered by ourself is usually weak and often mixed with various disturbance.It is necessary to process the speech signal effectively to obtain the report clearly.So speech enhancement is the basis of speech processing.The thesis analyses characteristics and shortcomings of some common speech enhancement methods,especially focus on the spectral subtraction method, and experimental analysis of this method with the existing residual noise.To address this problem, using an improved wavelet hierarchical threshold approach, a new threshold determination algorithm is introduced, using the selected threshold value after the wavelet decomposition of noisy speech by the layers of wavelet coefficients vector threshold processing, and then use the endpoints detection algorithm to detect the wavelet coefficients vector precessed by threshold, the detected voice signal components extracted from the wavelet coefficients to form a new wavelet coefficients vector, and then only those wavelet coefficients vector reconstruction. Experimental results show that our method can effectively avoid the spectral subtraction of the residual noise and achieved ideal speech enhancement. The result of wavelet hierarchical threshold processing in the condition of low signal to noise ratio is not ideal, this paper uses neural network method for low SNR noisy signal for further processing, through simulation research and record.signal analysis shows: using neural networks for noisy speech enhancement processing, can significantly improve the signal to noise ratio, achieved good results.