Improved wavelet pre-enhancement and hybrid model applied in speech recognition system

After study on the robust optimization of speech recognition system, we propose an improved wavelet thresholds de-noising method and combine it with the temporal filter to pre-enhance the noisy speech signals before recognition, which leads to good results. Then a hybrid model of hidden Markov and BP neural network is proposed, using BP to get the HMM (hidden Markov model) observation probability, which effectively combines the temporal model HMM and the acoustics model ANN. During the process of BPNN modeling, the selection of the hidden layer´s node number and the training arithmetic is optimized. Experiments conducted on dasiaten digitpsila speech recognition demonstrate the superiority of our approaches over the predominant approaches.