Research And Implementation Of Speech Recognition Algorithm Based On DSP

Modern society has entered the era of rapid development of information, the previous means of information dissemination and storage have been unable to meet the needs of people.In order to make our work and daily life faster and more convenient, we eager to resolve some matters by talking, such as voice dialing, voice navigation, voice input and so on.Meanwhile, with the development of voice recognition algorithms and digital signal processing devices, the practical speech recognition system has been applied and will be applied to broader areas.Therefore, the speech recognition system has a very high value of research and application.This paper introduces the theoretical knowledge of speech recognition systems, software and hardware design and related algorithms in detal.Above all, we describe the basis of speech recognition theory briefly,and discuss the front-end speech signal preprocessing.Then, it introduce features of DSP and structure of TMS320VC5402, and propose the speech recognition system based on DSP. We reserch the hardware system design, describe the composition of the system, and analyze the working process. The core circuit of the entire system is TMS320VC5402, and A/D conversion completed by TLC320AD50C, and the speech signal is trained and recognized by TMS320VC5402. Finally, MCU controls the LCD display of recognition results.According to the characteristics of Chinese phonetics, we design software system using existing algorithms,discuss the design process of software system, which its main process is the pretreatment, endpoint detection, feature extraction and pattern matching. Preprocessing inculdes voice signal analog/digital (A/D) conversion, pre-emphasis and windowing processin. Endpoint detection use extraction methods with better anti-noise performance, which is based on spectral analysis; Considering the human ear hearing characteristics, feature extraction used the MFCC. Considering the requirements is to design a speech recognition system which combine the three character, specific person, small vocabulary and isolated word, We select dynamic time neat (DTW) algorithm as identification algorithms of the system, and gives robustness training methods of voice template library. From the view of improving system rate and the speed of recognition, we study the improvement technology of the dynamic time neat improved algorithms, and put forward calculation method of the template threshold.The program loaded on the DSP are described in detail, including bootloader mode of DSP, bootloader mode of parallel memory and curing process and so on. Finally, the whole system performance was verified on the DSP evaluation board.