Study On Distant Hybrid Speech Recognition Method

Conventional speech recognition method has higher recognition rate in no interference noisy and quiet environment. However, traditional speech recognition system performance degrades because the speech signals can be seriously distorted by noise and reverberation. Especially with the distance between the user and the microphone increasing, speech signal becomes significantly deteriorated. Besides, a single language has been failed to meet people’s needs. Therefore, studying on distant hybrid speech recognition method becomes a hot point of machine learning and speech processing.The thesis takes microphone array as a front-end of speech recognition system in order to use its spatial selectivity. Therefore, the thesis carries out the research on distant hybrid speech recognition method. The thesis introduces traditional beamforming vs parameter optimization method, GMM-HMM vs DNN-HMM acoustic model, respectively. In addition, this thesis does research about the distant hybrid speech recognition method of array parameters optimization.According to different characteristics of Chinese-English pronunciation and modeling unit, the thesis has a hybrid acoustic model. On the basis of it, the method designs decision tree in order to obtain the acoustic model parameter. Besides, the thesis builds distant hybrid speech recognition system which based on array parameters optimization by HTK Toolkit. And the thesis uses multi-frames MFCC feature parameters and builds DNN-HMM acoustic model. This thesis builds distant hybrid large vocabulary speech recognition system by Kaldi toolkit. Experimental results show that the recognition rate of DNN-HMM acoustic model is more than that of GMM-HMM.