Research On The Grouping Speech Recognition System Based On HMM

Speech recognition technology is also called Automatic Speech Recognition (ASR), its aim is to convert the human words into signals which the computer can identify. In recent years, because of the rapid development of the electronic information and the internet, the speech recognition technology has achieved periodical progress. Along with the function of the embedded products becoming increasingly powerful, the speech recognition technology based on embedded equipments becomes a hot spot on research and application at present. A speech recognition system with large identifiable vocabulary, good real-time performance, low demand for the system resources and high recognition rate, has very practical value and great market prospects.In this paper, I introduced the historical background of speech recognition technology and the current research status at home and abroad in the beginning. I theoretically studied and analyzed each link in the process of speech recognition. Firstly, the pre-emphasis, framing, windowing and endpoint detection by short-time energy and zero crossing rate in the pre-process were analyzed. Secondly, the implementation of the LPC parameters, the LPCC parameters and the MFCC parameters in the progress of feature extraction were described. The advantages of the MFCC parameters compared with the LPCC parameters and the LPC parameters were illustrated as the reasons to choose the MFCC parameters in the process. Finally, several commonly used recognition algorithms in the process of recognition were analyzed, emphatically for the structure and the principle of the Hidden Markov Model (HMM). For isolated words, speaker-independent and large vocabulary in the recognition process, an improved grouping speech recognition algorithm was proposed in this paper to solve the problem of high demand for the system hardware resources, which was caused by long recognition time, large amount of calculation and memory. To begin with, the piecewise average dimension reduction method was used to calculate the MFCC parameters, which can ably avoid the time neat problem caused by the different lengths (frames) of speech and ensure the high stability of grouping. Secondly, in order to eliminate the effect on packets stability that came from the randomness of initial value selection in the K-means clustering and the problem of reducing differentiation degree to templates caused by too many speech templates, the experience adjustment algorithm was proposed to further increase the stability of grouping. Moreover, in order to offer the theory guarantee for the stability of grouping, combined with the knowledge on mathematical statistics, the confidence test algorithm was proposed to prove the grouping was stable in math. Last but not least, according to different groupings, the rate of recognition and the time of recognition would were different. So the maximum algorithm was proposed to determine the best group, and so as to greatly meet the requirements of users both on the recognition rate and the recognition time.In order to examine the effects of the methods, it was tested on the PC platform by Matlab. I analyzed the experimental data by comparing with the traditional HMM algorithm and the traditional clustering algorithm. The results showed that the grouping of stability for the improved grouping speech recognition system was very high, which needed less recognition time, less amount of calculation (reduce more than50%). The memory footprint was decreased significantly, and the demand for the system hardware resources was also reduced. The only shortage was that the recognition rate decreased slightly. Overall, the methods were effective.