Speech Recognition Research Of Non-specific Person's Isolated Words Based On DTW Model

In recent years,the technology of speech recognition has been developed,the system identification rate and the performance of the recognition have been improved.Speech recognition of isolated words is widely used on the area of auto-control and Smart home,such as instruments,robots,vehicle driving and control of home appliances,because of its low demand for storage and high activity.This paper studied the three stages:prepossessing,feature extraction and pattern recognition of the speech recognition of isolated words,and mainly accomplished the work as follows:(1)This paper analyzed and summarized the speech recognition.In the research of endpoint detection,this paper took English speech recognition as an example and found that the current detection algorithm is not very accurate to judge a number of polysyllabic words and brings difficulties to the later recognition process.To solve this problem,a detection algorithm of forward multiple search is proposed.It increased a transition zone and set the judgment number of speech frames in this zone,and experiments to verify the validity of the algorithm were did.(2)In the research of feature extraction,this paper made a comparison of the performance of Linear Predictive Cepstral Coefficients(LPCC)and Mel Frequency Cepstrum Coefficient(MFCC).And the MFCC coefficient was chosen as the characteristic parameter.(3)As for the Identification model,this paper discussed the thought of Dynamic Time Warping(DTW),Vector Quantization(VQ),Hidden Markov Model(HMM),and Artificial Neural Network(ANN).Related comparative studies showed that algorithm of DTW is more suitable for non specific people's isolated word recognition.The current DTW algorithm required endpoints of speech to be recognized and in template are aligned.A nonlinear transformation of speech recognition increased the distortion degree of the speech signal.A matching algorithm is proposed to remove this problem.It optimized constrained conditions of searching paths,and by dividing the searching area more reasonably,the signal changes caused by nonlinear transformation were avoided.At last,the algorithm was analyzed.(4)The software part of the speech recognition system was designed and simulation experiments were carried out to verify the work.The experimental results of the recognition rate and the recognition efficiency of the front and improved system aregiven,and the comparison and analysis were carried out.At last,speech recognition experiments were did using the improved endpoint detection algorithm and the traditional endpoint detection algorithm,and the improved endpoint detection algorithm has a better anti noise performance.