Chinese Speech Recognition Based On Vowel Length Adjustment

Speech recognition is a major research topic in the world.. Speech recognition system has certain capacity of self-adaptation to speech velocity. However, when it comes to different person, the capability of self-adaptation is not satisfied. The recognition rate is often comparatively low when it distinguish those pronunciation which is either too fast or too slow.This paper starts with some basic theory of speech recognition, adjusting the differents between the length of pronunciation result in decline of recognition rate. The major point of this paper is developing the vowel grouping algorithm and adjustment of the length of vowel, which are both based on HMM and DTW. Moreover, the author thoroughly simulates the algorithm and compares the results. The main contents of this paper are described below:1,First of all, the author explicitly analyzes the Hidden Markov Model (HMM) and Dynamic Time Warping (DTW). Meanwhile, some simulation experiments is accomplished based on the model and algorithm. The results of simulation based on DTW were the foundation recognition rate for further research, clearly showed the improved algorithm had risen up the recognition rate.2,In Chinese pronunciation, each syllable contains the vowel. The vowel's length is the main part in the syllable, but the vowel doesn't contain the important information. According to these characteristics, we simulate a method to adjust the speech velocity by using similar waveform, which is found by correlative coefficient in vowel part to insert or delete the similar waveform.When it comes to manipulate the vowel part, the author proposed a brand-new method. Compared with the original "modulation-judgment-modulation again-judgment again" method, this innovate method rose up the arithmetic speed better.3,Because of the old method, the threshold value of Lmax and Lmin cannot consider most of isolate words. This article offered an improving algorithm, vowel grouping. This algorithm used three kinds of grouping algorithm:grouping based on the type of vowel, grouping based on similarity and grouping based on length. Different vowel groupings had different threshold values, that helped accomplish the goal. The result of this simulation shows that comparing with the old algorithm, the new method has the remarkable improvement on the recognition rate of the system.