Algorithm Research Of Speech Endpoint Detection And Hardware Implementation

Content：With the rapid development of communication technology and digitalsignal processing technology, speech endpoint detection which can separate thespeech and noise has had an important position in speech processing system.This paper introduced and analyzed kinds of typical speech endpoint detectionalgorithms. It summarized the respective advantages and disadvantages. Then, thepaper improved two algorithms from the research or application direction so that theycould be adapted to broader environment. Finally, it used micro-controller implementthe hardware circuit which used to verify algorithms. The main research results are asfollows:1The paper improved spectral entropy algorithm and average magnitudedifference function algorithm after the basic research: first, the paper showed a newcharacteristic parameter based on short-time energy and spectral entropy algorithm.The results showed that this method had better robustness and precision, especially inno noise environment. Second, In order to distinguish voice and impact noiseaccurately, the paper improved algorithm with a new parameter combined with thevoice periodicity and short-time average magnitude difference function to detectspeech and car impact noise. The results showed that the new method could make adistinction between speech and impact noise in a quiet environment.2The paper finished the hardware circuit which was verified improved spectralentropy algorithms. The system is mainly made of enhanced MCU STM32F103RBchip and a voice system about ISD1760. It implemented voice signal recording, ADtransform, speech endpoint detection and showed the output with LCD which wouldbe saved in the RAM. The main process is as follows: first, the MCU made ISD1760record voice through SPI. Second, after AD transform, the MCU made the processabout speech endpoint detection. At last, the different results might be saved weredisplayed by LCD with different color.