Speech Recognition Used For Language Healing System Of Deaf Children

The Language Healing System of Deaf Children is mainly used for extracting the speech parameters after processed the training speech data. And then make the speech recognition, let the trainer contrast his pronunciation with the standard pronunciation and correct his pronunciation step by step.The speech recognition is the center theory of the system. This paper mainly dissertate the speech recognition theory, include the preprocess, the endpoint detection of the speech , extracting the parameters,(such as the pith ,LPCC and the MFCC), HMM theory ,at last we use the HTK to have the speech recognition experiment.During the speech preprocess, we dissertate the speech de-noise and the endpoint detection. In the speech de-noise, we introduce a new de-noise method of wavelet threshold based on new threshold function, this method is between the soft threshold and the hard threshold, it has nice continuities and high-order derivatives resulting in mathematics disposals convenient., it can reduce the noise better. In the endpoint detection, we improve the traditional double-threshold method. We set three points of the training speech, and then use the time domain parameters to detect the endpoint of the speech successfully.In the extracting the speech parameters, at first, we extract the pitch of the speech. We also improve the AMDF method, we make the speech have the magnitude difference square function and then normalize it, at last during the late process, it can restrain the isolate point successfully, enhance the precision. We also particular introduce the LPCC and MFCC , and contrast their advantages and disadvantages.The mainstream technology of speech recognition acoustic modeling should be the HMM which models the system based on statistic method. The primary principles of HMM and its applications in speech recognition is introduced in the paper.Finally, several experiments are performed to test the steps and principles of speech recognition. In the experiments the performances of different feature parameters and different HMM states displayed. And we choose the best kind of feature parameter and HMM model's amount according to the experiments results.