Speech Recognition System Based On HMM/ANN Hybrid Model

This paper firstly analyzes the situation of speech recognition and concludes that noise-robust speech recognition has become an important research area in recent years. Then the basic theories of speech recognition are introduced, such as the framework and categories of recognition system, preprocessing, feature extracting, recognition methods. Hidden Markov Model (HMM) and Artificial Neural Network (ANN) have been widely used in speech recognition. Their respective advantages and disadvantages are analyzed and the combination methods of them are summarized. According to human aural perception, a novel approach based on sub-band HMM/ANN hybrid model is presented, which is composed of several sub-band HMM models and a full-band HMM model, then a RBF neural network is used to make the final decision by the fusion of all the HMM models. In this paper, both sub-band MFCC and full-band MFCC are used as recognition features, which is different from the traditional sub-band recognition systems and the combining method of ANN and HMM is also different. In order to testify this new method, the sub-band HMM/ANN hybrid model is simulated with the MATLAB 7.0. The influences on recognition results of some factors as feature parameter, the definition of sub-bands, noise intensity are discussed. By comparing with the traditional techniques, results show that this sub-band hybrid model can improve the recognition performance for both noisy and clean speech. The hardware development platform is designed and realized lastly, which is based on fix-point DSP (TMS320VC5409) of TI Corporation. This system is portable and low-cost. The method in this paper has a great importance on robustness for speech recognition and its commercialization.