【正文】 aved as a template eigenvector. For the identification process, the vector of these characteristics pared with the template to calculate distance, the minmum distance is the recognition results.3 Hard system designThe design of the hardware structure is Sunplus SCM SPCE061A as the core, the extemal expansion of the corresponding functional hardware. Sunplus 16bit microcontroller SPCE061Achip hardware in the mand structure and system are very applicable to the voice signal processing, its main features are: faster, more disruption, a number of A/D converter, particularly, it has s builtin audio amplifier and the AGC function of singlechannel A/D converter, and with audio output of 10 dualchannel D/A converter. The instruction adds DSP, it is very convenient for the plex digital signal processing applications, and much cheaper than the DSP chip. The whole system hardware structure diagram shown in Figure 12.SensorAntialiasing filiterSPCE061A ReproducerControl output3 keystokes32KB FLASH store Figure 12Sunplus 16 SCM is the core, we divide the entire hardware system into the following parts:(1) Voice feature extraction, learning, discriminant function We can use the DSP functions of the SPCE061A to preemphasis and depositor for the enter voice digital signal, extraction feature vector。 P is the order of the feature model。 Imax、Imin for the maximum and minimum energy.Chinese characters are constituted by Qingyin and Zhuoyin or only Zhuoyin directly , there is not a Qingyin at the end, the wordterm isolation gap between is very short. The way of determine the starting point is: the first frame which consecutive 10 frames En are more than IIL —directly address the situation posed by the Zhuoyin。These will enable large the whole system and cumbersome design. This design uses a DSP function and builtin voice of A/D converter of Sunplus microcontroller, the function of DSP and the integration of control makes the system greatly simplified. Now there are many methods on the features of voice recognition and extraction. For example, Hidden Markov law on nonspecific person has a hign continuous speech recognition rate, but has a cumbersome process . In real life, people often use some short order to control the conduct of the object, in response to the Chinese word recognition of the isolation of this particular, the voice features is about the linear prediction for the design of the voice, the sublinear matching method, which based on the timing of the characteristics of poor, is the identifying method, it has a high rate and simple operation and can well meet the design requirements. 2 A voice recognition system The basic speech recognition system, including a major pretreatment, A/D conversion, recognition of the beginning and end point, feature extraction and recognition judgement, and other parts of the structure, the diagram shown in Figure 11. Recognition of the beginning and end point The voice signal processing bases on the shortterm stability of the voice, when the sampling frequency is about 8kHz, the desire of a frame is about 128point and the long is 16ms. Ziyin includes Zhuoyin and Qingyin, pared with the noise, Zhuoyin performance for the highenergy, Qingyin performance for the high rate of zero. For a voice data, energy and zero rates were short periods of energy En and the characterization of short time with zero rate ZCRn. Through the different of the background noise, we can determine the voice beginning and end point. For 10 frame consecutive background noise datas, we can calculate the IZCT (zerorate threshold, recorded as tIZCT ) and the IIL (energy threshold, recorded as tIIL) : tIZCT = min(IF,ZC +2IZC)tIIL = min ((Imax Imin) + Imin,4Imin)ZC、IZC for the zero rate and the average standard deviation。(3) it needs an external A/D converter chips 。外文翻譯（原文）Design of Voice Recogn ition System Base