A Study On The Identification Of Isolated Words In Yi Language

Speech recognition is one of a research area of pattern recognition and artificial intelligence. Yi speech recognition is a new direction for research of speech recognition. Currently, speech recognition products for English, Chinese and so on have been put into commercial. In the domestic, the research for Chinese speech recognition has been developed to large vocabulary continuous speech and robust speech recognition, especially for Yi language which has about 871 million users is little speech recognition research. So it has important significance to research the Yi speech recognition. On the one hand, it can make the Yi compatriots enjoy the convenience brought by technology and promote communication and development of all ethnic groups. On the other hand, it could be better to inherit the splendid culture of Yi and enhance our rich cultural deposits.In this paper, after deeply analyzed the characteristics of Yi language and preprocessing methods, characteristic parameters especially linear prediction cepstrum and Mel frequency cepstrum coefficients of Yi speech signal are discussed further. The influence of the recognition rate of Yi isolated word speech recognition system that is studied in this paper caused by the different order of linear prediction cepstrum coefficients and Mel frequency cepstrum coefficients and its differential parameters are compared. The experimental results show that the 12 order of Mel frequency cepstrum has the highest recognition rate and most suitable for the characteristics parameters of the system that studied in this paper.Support vector machine has better classification accuracy and generalization ability than hidden Markov model, artificial neural network and the other speech recognition algorithms that is commonly used with condition of small sample and it can solve linearly inseparable problems by introducing the kernel function. Least squares support vector machine is a development of the standard support vector machine.It turns a solution of convex quadratic programming into a solution of linear equations, then, it has a faster convergence speed and reduces the resource utilization when solving the question, and the experiment for Chinese speech recognition that uses the least squares support vector machine for the recognition algorithm verify its effectiveness in the speech recognition system. So the least squares support vector machine is used as recognition algorithm in this paper. In addition, it is compared the performance of different kernel functions for the system that studied in the paper using recognition rate, recognition time and model training time as indicators. The result shows that Gaussian kernel function has a better performance, on this basis, Gaussian kernel function is improved by the dynamic time warping algorithm and the experiment result shows that the improved Gaussian kernel has a better performance in the certain experimental environment.