嵌入式语音识别的前后端处理关键技术研究  

Research on Key Technologies of Front-end and Back-end for Embedded Automatic Speech Recognition

作　　者：何东之 黄樟钦 侯义斌 丁志浩 

机构地区：[1]北京工大学软件学院

出　　处：《计算机仿真》2010年第2期192-195,共4页Computer Simulation

基　　金：北京市优秀人才培养资助个人项目(20081D0501500169)

摘　　要：在语音识别技术的研究中,语音端点检测和拒识是语音前后端处理的关键技术。在噪声环境下,传统的过零率和短时能量的端点检测效果会变得很差;频域的端点检测方法虽然较时域的端点检测方法鲁棒性更高,但是它需要进行大量的计算不能很好地满足嵌入式系统。针对嵌入式系统的特点,为提高语音识别能力,提出了基于统计理论的孤立词的端点检测算法,在一个相对较长的时间段内语音信号服从正态分布,而噪音信号主要存在于信号均值的一定方差范围之内。方法既满足了嵌入式系统的计算要求,又有一定鲁棒性。Speech endpoint detection and out - of - vocabulary rejection are two important parts of the whole automatic speech recognition process. The performance of traditional speech endpoint detection based on short - term energy and zero - crossing rate becomes very poor in noisy environments, and sometimes even unable to work. Methods based on frequency - domain need complex computing, and it can not meet embedded systems well. In this paper a new endpeint detection algorithm is presented, which is based on statistical theory for isolated - word. In order to distinguish speech and noise, speech signals are looked as submitting to normal distribution in a relative long duration, and noise signals exist only within 0. 62 standard deviation. Using this method, successful endpoint detection is carried out.

关 键 词：语音处理 语音识别 拒识 端点检测 支持向量机 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]