A Study on Speech Endpoint Detection and Applications
碩士 === 大葉大學 === 電機工程學系 === 103 === In this research we discuss the effect between endpoint detection and speech recognition. First we refer to HTK’s endpoint detection. We use MATLAB to design a program to change the threshold constants for five kinds of SNR wave files. Each SNR wave files can obt...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | zh-TW |
| Published: |
2015
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/38763613104682958616 |
| Summary: | 碩士 === 大葉大學 === 電機工程學系 === 103 === In this research we discuss the effect between endpoint detection and speech recognition. First we refer to HTK’s endpoint detection. We use MATLAB to design a program to change the threshold constants for five kinds of SNR wave files. Each SNR wave files can obtain a best threshold constant. This optimization of threshold constant have 1.2%~3.9% endpoint detection error improvement than HTK’s endpoint detection. But SNR5 and SNR0 wave files have 39.5% and 37.0% endpoint detection error improvement. Then we improve the endpoint detection by adding zero crossing rate endpoint detection. We get 20.8%~46.29% endpoint detection error improvement for all kinds of SNR wave files. Especially on SNR20 wave files have 50.86% endpoint detection error improvement. Finally, we put both HTK’s endpoint detection and our endpoint detection results into a speech recognition system. On the least noise wave files. The recognition rate of HTK’s endpoint detection and our endpoint detection are 98.64% and 99.01%. Each SNR wave files have 0.37%~6.57% recognition rate improvement. Even on SNR10 and SNR5 wave files have 31.94% and 21.49% improvement of recognition rate.
|
|---|