嵌入式平台的复杂环境下端点检测

An Endpoint Detection of Complex Environment Base on Embedded System

在嵌入式平台下的语音识别系统中,端点检测是非常重要的,语音起始点的位置直接影响识别率。然而在复杂环境下,语音的输入波形千差万别,端点检测的精确性很难保证。本文研究了一种基于零能积的区域概率统计法,通过训练得到语音零能积的概念分布,并结合一阶线性差分及谱减法进行端点检测。进一步提高了信噪比,解决了复杂环境下端点检测精确度低的问题。

For speech recognition systems in embedded environment, endpoint detection is extremely important since location of speech endpoint works on recognition rate directly. However, great differences of input speech signal are shown while low accuracy of endpoint detection in complexity environment. In this paper, an area probability statistical method based on the energy-zero-product is presented. So the conceptual distribution of the speech energy-zero-product is obtained through training, with which first order linear difference method and spectral subtraction are used in endpoint detection. Simulation experiment results illustrate that this method raises SNR further as well as arrives at the solution of low accuracy of endpoint detection in complexity environment.