一种基于检测元音的孤立词端点检测算法  

A Robust Endpoint Detection Based on Vowel Detection in Isolated Word Recognition

作　　者：邝航宇[1] 张军[1] 韦岗[1] 

机构地区：[1]华南理工大学电子与通信工程系,广东广州510640

出　　处：《电声技术》2005年第3期40-43,48,共5页Audio Engineering

摘　　要：提出了基于检测元音的端点检测算法。首先检测到语音中的元音的端点,然后利用元音的端点作为参考端点检测出语音真实的端点。将新方法应用在TI46数据包在NoiseX-92的5种噪声下的端点检测和识别实验,并和基于能量的端点检测算法比较。2种不同的实验表明,提出的基于检测元音的孤立词端点检测算法可以在不同信噪比下提高端点检测的准确率,并在低信噪比的环境下能明显提高语音识别系统的识别率。A new algorithm for endpoint detection which based on the different characteristics of speech and noise is presented. The proposed algorithm relies on detecting the endpoints of the vowels. The starting-point of the first vowel and the ending-point of the last voiced vowel are located to be used as reference endpoints for further determination of the actual endpoint. The new algorithm is evaluated by a task on TI46 database and NoiseX-92 that added to original speech at different SNRs. Experimental results show that the performance of endpoint detection and speech recognition can be improved by this new algorithm in various noisy environments.

关 键 词：语音识别 端点检测 自相关函数 

分 类 号：TN912.34[电子电信—通信与信息系统]