Realization And Optimization Of Speech Recognition System Based On ARM A8 And Vector Taylor Series Feature Compensation

In recent years, with the development of speech recognition technology and embedded systems, embedded systems based on speech recognition technology has been widely applied to the automobile, smart toys, industrial control and other fields. Voice-based human-computer interaction is more natural and convenient, embedded systems has low cost, small size, low power consumption advantages, thus achieving robust embedded speech recognition system has important application value.This paper studies the optimization of embedded isolated word recognition system based vector Taylor series features compensation. By optimizing accelerate recognition speed of the system, improve system availability. The main work of this paper is as follows:1, Study the structure, evaluation parameter and pretreatment technology of the speech recognition system. This paper uses dual threshold endpoint detection algorithm based on spectral entropy. Mel frequency cepstral coeffficitent and corresponding first-order differemtial cepstral coefficients is used as the chararcteristic parametersin. Hiden markov model is used to model the acoustic parameters. The vector Taylor series is used as feature compensation algorithm.2, Optimizes the program based on the hardware platform features of ARM A8, C language features. Reduce the accuracy of the variables in the program, achieve the purposes of use floating-point co-processor unit NEON to complete the float point arithmetic; adjust the structural of program to improve readability of the program, easy to post-maintenance; optimized code and offline calculation operation of some variables. Improve the efficiency of the program, to accelerate the recognition speed of the system. Then real-time test the error rate and recognition speed of the system on harfware platform after optimized.3, Realize the optimization based on the speech recognition algorithm. Respectively, Fisher ratio criterion is used for reducing the dimension of speech fearure parameters, using the nearest neighbor estimation method to reduce the amount of computation of Viterbi algorithm, and reduce the mixture of gaussian mixture model which is used in the vector taylor series algorithm. First, the MATLAB simulation test is proposed to analysis and verify the algorithm under different noise environment. Then, realize the optimized system on ARM A8 platform, and real-time test the error rate and recognition speed of the system on harfware platform, the test results show the effectiveness and reliability of the optimization method.