Based On The Adsp-21161 Voice Recognition Systems Research And Practice

The Hide Markov Model (HMM) was widely used in modern speech recognition system. In order to simplify the model and reduce the computation, the matrix of covariance was always supposed as diagonal matrix,however, the performance of system will be degraded. A new model for speaker-independent speech recognition is proposed in this paper, called segmental probabilistic model(SPM). This model can eliminate the Vector Quantilization (VQ) distortion and the same performance as continuous density distribution without so much computation.In this paper, the speech recognition system was designed base of AD1836 and ADSP 21161. the LPCC and HMM model was analized detailed, and the proposed CDD-SPM model based on HMM model. Later, the software of the speech recognition was descripted detailed, include the detection of the start-end of the speech signal, detection of eigen-parameters, training and recognition base on the ADSP21161 EZ-KIT-Lite development board. Finally the experimental result shows the model could reduce the space complexity for model storage and the time complecity for model training and recognizing.The main content and creative work in this dissertation include:(1) The new model of training and recognizing, which based of HMM, was proposed. The algorithm was simulated on MATLAB with the recorded speech data., the experimental result shows the model could reduce the space complexity for model storage and the time complexity .(2) The implementation of the CDD-SPM algorithm was completed, include the detection of the start_end of speech signal, the detection of parameter of speech sample and training and recognizing of speech signal on the ADSP21161 EZ-KIT-Lite development board.