Hmm And Wavelet Neural Network-based Speech Recognition System

Speech recognition becomes more and more important due to its significant theoretical meaning and practical value. Nowadays, most speech recognition is based on conventional linear system theory, such as Hidden Markov Model(HMM) and Dynamic Time Warping (DTW), and so on. In order to solve the limitation of the linear system theory of speech recognition, nonlinear system theory method must be introduced.In this paper, preprocessing, feature parameters extraction and recognition algorithm of the speech recognition system is discussed. This includes the following parts:1. The mixed feature extraction method based on the LP Mel Cepstrum Coeficient and Mel-Frequency Cepstrum Coeficient is discussed. It shows that using mixed feature parameters achieves better effects than other feature parameters extraction method under the different signal and noise ratio(SNR) by the experimental results.2. Considering the phases of speech preprocessing, the results of speech recognition may be affected by the noise background when we directly do the end detect and feature parameters extraction. We put forward a wavelet-based denoising method by Modified threshold function which is self-adaptive on multiscale. It shows that the algorithm can remove the noise effectively and the denoised signals follow the original signal very closely by the experiments.3. We study a new speech recognition method based on HMM and WNN by using the abilities of the modeling ability in time domain of Hidden Markov Model and the powerful classification and decision ability of Wavelet Neural Network. Compared with the traditional method(HMM), It shows us that this method has good performance at speech recognition specially in noisy speech condition by the experiment. At the training method of WNN, a new method of WNN based on gradient PID is proposed. Compared with the traditional gradient method, It shows that this method has good performance by the experiments. Not only convergence faster than the traditional one but also not easy to fall into local minimum.4. The robot control system based on speech recognition at the AS-R robot platform is designed. This system include three parts: the speech recognition, wireless network communication and the robot motion control. We finish the software of the system design and do the experiment on the robot.