| Summary: | 碩士 === 國立中央大學 === 電機工程研究所 === 88 === Hidden Markov model (HMM) was widely used for speech recognition and has been proved useful in dealing with the statistical and sequential aspects of the speech signal. However, their discriminative properties are weak if they are trained with the maximum likelihood. On the other hand, neural networks (NN) have powerful classification capability but are not well-suited for dealing with time-varying input patterns. In this study, a hybrid HMM-NN speech recognition system that combines the advantages of both models is presented. Three neural net state models, HMM-NN-Net, HMM-HMM-Net and NN-NN-Net, are developed for the proposed hybrid HMM-NN system. All the experimental results are compared with the one obtained from HMM.
In the speaker-dependent experiment, the recognition rates of all the three models are above the level of 90 percent. Furthermore, in spite of the results of HMM-HMM-Net models, all error rates approach to zero after adjusting the criterion.
In the speaker-independent case, HMM-NN-Net model achieves a recognition rate of 94.25 percent and has the best performance compared with other models. Besides, NN-NN-Net model requires less training time than HMM-NN-Net model although its recognition capability cannot compete with HMM-NN-Net model.
The experimental results indicate that the hybrid HMM-NN recognition system based on HMM-NN-Net model improves the performance of traditional HMM system. It is also found that the criterion of neural net state models was related to the recognition capability.
|
|---|
