Research And Implementation Of Voice Recognition System Based On Embedded Platform

In recent years, the speaker voice recognition technology is maturing gradually, the speaker voice recognition as a biological authentication technology, has its unique advantages, such as the sound is non-contact, natural, users can easily accept. Because of the advantage of language media, through the voice identification technology is rapidly applied to the actual, highlight the enormous market potential, speaker voice print recognition technology has become a new and high technology industry. Along with the computer hardware and software technology, semiconductor technology, electronic technology, communication technology and network technology development, as well as the embedded technology is continuous development and updated, its performance is greatly improved robust and portable. Real time data acquisition, filtering, can be at a low power consumption, small volume of embedded equipment. Today, the processor command, because of its special structure and high efficiency to enable it to quickly compile the realization of speech recognition algorithm, meet the needs of today’s digital signal processing and high real-time requirements. High performance embedded voice recognition system for speech recognition, because of convenience, economy, accuracy and embedded system’s portability, mobility and other advantages, are widely used in people’s daily life, has the broad prospects of development.Based on the analysis of the common relevant principles and techniques of voiceprint recognition, the paper mainly does a research on the Mel-frequency cepstral coefficients (MFCC) feature extraction and DTW algorithm was improved. Besides, this paper makes the corresponding improvements to the deficiencies of MFCC and DTW. this paper builds a small-capacity voiceprint recognition system based on ARM11and WinCE embedded platform. On the basis of precursor’s study, the mainly relevant improvements includes the following three aspects:1. Feature extraction:As for the shortcomings of the widely used standard MFCC, this paper introduces a weighted difference combined MFCC feature parameters. Specifically, the improved MFCC uses the short-time frame energy and short-time weighted Zero-crossed rate to take the place of the negative role of1,2-order component of MFCC, then weights the remaining MFCC different components according to their different contributions of each component. At last, a new feature parameters was composed of weighted MFCC and its first differential orders.2. DTW algorithm:Using the improvement DTW algorithm, an alternative to standard DTW algorithm, this paper puts forward the overall path constraint, the algorithm has a good robustness, and improve the efficiency of the algorithm and code quality.3. Embedded system realization:based on the ARM11ok6410embedded system,Some optimization processes are utilized to overcome the limitation of the embedded system’s resources. And those optimization processes include optimizing and transplanting the WinCE operation system,and through the cross-platform software development, embedded development successfully build a good platform for a large vocabulary speech recognition system. According to the project requirements, research and analysis of the improved DTW algorithm and the traditional DTW algorithm performance difference between the operation and embedded in the analysis of the circumstances.At last, some relative experiments are executed to test the voiceprint recognition system. The results of the experiment show that the system has a higher recognition rate in the same text content test, and it is necessary to improve the recognition rate in the text-independent test. In addition, the average recognition rate of the system is improved by4%or so by using the modified characteristic parameters and improved algorithm.