Study On The Method Of Speaker-Dependent Isolated Word Speech Recognition

With the coming of the information era based on digital techniques, human- machine interaction is more and more used. Natural communication between people and machines without using keyboard or mouse is the goal pursued for a long time. Today since computers are widely used, communication by speech has become an important research problem of human-machine communication. Speech recognition is a compound science, it has very wide application fields. Today, speaker-dependent isolated speech recognition system is the most application of speech recognition.Based on the summary and research of the existing speech recognition method, proposed a new short time energy-permutation entropy endpoint detection algorithm. Constructed a speech recognition system based on DTW, implemented this new endpoint algorithm in this system.Endpoint detection is one of the important parts in speech recognition. Efficient endpoint detection can reduce the processing time, eliminate the disturbance of non-speech segment noise, and improve the system's recognition performance. Short time energy algorithm is sensitive to fluctuation of energy's magnitude but it performances is poor in low signal-to-noise. Permutation entropy can illustrate the tendency of signal's fluctuation, but it is very difficult to judge the endpoint by itself. This article proposed a new short time energy-permutation entropy endpoint detection algorithm. It used short time and permutation entropy these two characters, adopted two thresholds judge methods, and improves the accuracy of endpoint detection. The experiment result showed that short time energy-permutation entropy algorithm is better than either short time energy algorithm or permutation entropy algorithm.Constructing a speech recognition system, includes implementing pre-processing, endpoint detection,feature extraction and speech recognition algorithm which is the core problem of speech recognition. This article used Matlab platform,used hamming window to frame-division , extracted MFCC parameter, used short time energy-permutation entropy endpoint detection algorithms we proposed, choosed DTW algorithm, and then carried out a speech recognition system of speaker-dependent isolated small vocabulary.