论文部分内容阅读
在孤立词语音识别中,从背景噪声中找出语音的开始和终止是很重要的,准确地确定语音的端点,是正确的语音识别的基础,并能使语音处理的计算减少到最小。本文研究了端点识别的二种算法:能量过零率法,多门限过零率法。能量过零率法用得比较普遍:能量检测浊音,过零率检测清音。多门限过零法将能量和过零率结合成一个参数,使端点检测只用一个参数。在此基础上,我们又提出了一种新算法:极值变化率法。极值变化率能更好地区分噪声和清音,而且检测端点时也只用一个参数。
In isolated word speech recognition, it is very important to find out the beginning and the end of speech from background noise. Accurately determining the end point of speech is the basis of correct speech recognition and minimizing the calculation of speech processing. In this paper, two kinds of algorithms for endpoint identification are studied: the energy zero-crossing rate method and the multi-threshold zero-crossing rate method. Energy zero-crossing method is more commonly used: energy detection dullness, zero-rate detection of voiceless. The multi-threshold zero crossing method combines energy and zero-crossing rate into a single parameter that allows endpoint detection to use only one parameter. On this basis, we also proposed a new algorithm: extreme value rate of change method. Extreme rate of change can better distinguish between noise and voiceless, and the detection of end-point only one parameter.