Research Of Speech Recognition Technologies Based On HMM-ANN Model

Although Hidden Markov Model(HMM) and Artificial Neural Network(ANN)are regarded as the most widely used and the most popular algorithms in the area of automatic speech recognition(ASR), both of them have shortcomings. HMM has strong capabilities in dynamic modeling and processing the dynamic speech signals,but its performance of classifications should be improved. ANN has extremely strong abilities in classification and input/output mapping among signals, however, its performance of processing dynamic signal is still unsatisfactory. Therefore, in order to foster strengths and circumvent weaknesses, it is necessary to combine the two models into a hybrid model. In this paper, a hybrid speech recognition model is proposed based on both HMM and ANN(HMM-ANN) models, and Probabilistic Neural Network(PNN) is introduced as the network learning algorithms in the section of ANN. This hybrid model shows good abilities in processing signals and improvement of speech recognition rate.In this paper, some achievements have been made like below:(1) Made comparisons of advantages and disadvantages of ASR methods between HMM and ANN, and then proposed a ASR system based on HMM-ANN hybrid model which can recognize small and isolated Chinese vocabularies. Finally,simulations for each function module of the model by using MATLAB software were finished.(2) Made comparisons of speech recognition rate among the three different ASR models. The experimental results showed that the speech recognition rate of the hybrid ASR model achieved the highest.(3) Studied the robustness of speech recognition. On one hand, this paper proposed an improved method of endpoint detection algorithm. The noisy audio signal was de-noised first in order to improve the Signal-to-Noise Ratio(SNR) environment,then the traditional algorithm of dual-threshold speech endpoint detection was introduced to detect the endpoints of the de-noised speech signal. The experimental results indicated that this proposed algorithm enhanced the improvement of recognition rate and achieved better robustness when adapting the noisy environment.On the other hand, this paper also proposed another improved algorithm during feature extraction. Compared with traditional MFCC, the experimental resultsdemonstrated that the proposed SS-MFCC parameters, which was extracted based on the approximate estimation method of non-zero entries in spectral subtraction estimation, were more closer to pure speech feature parameters than MFCC, and the speech recognition rate achieved higher.