摘要
端点检测在语音识别中占有十分重要的地位，端点检测的准确性将直接影响整个语音识别系统的性能。已往的自动端点检测方绝大多数都是利用帧平均能量ＥＮ，帧平均跨零数ＺＮ，帧平均跨零积Ａ和帧平均零比Ｂ等参数来确定语音段的始点和终点。这些方法的缺点是难以设置对各次实验都合适的固定阈值，这给实际应用带来了很多不便。本文提出了一种基于迟滞编码的自动端点检测方法──在对语音信号进行迟滞编码的基础上，利用各帧的码字和来判断语音段的起点和终点。该方法充分利用了噪声和信号的统计特性，克服了已往端点检测方法的不足。实验结果表明，该方法具有良好的性能。
In this paper, we present a new effective technique for automatic segmentation of speech. The approach to the problem canbe divided into three steps: Firstly, code the speech signal with sluggish-coding techniqUe proposed by us. Then, Calculate the sum ofcodes in each frame. Finally, compare the sum with threshold given, and thus determine the beginning and the end of a speech period. Ex periment has been Performed, and the result shows that the method has superior performance. Compared with traditional segmentation methcds by using the short-time average energy level (EN),the zero-crossing number (ZN) and parameters related to EN and ZN, this methodhas better performance and is insensitive to the SNR, and thus need not establish different thresholds according to the variations of theSNR.
出处
《电路与系统学报》
CSCD
1996年第4期29-32,共4页
Journal of Circuits and Systems
