论文部分内容阅读
本文研究了一种利用短时能频值(Energy-Frequency-Value)进行语音端点检测的方法,它区别于传统的分别用短时能量和短时平均过零率作是否超过阈值判断,再通过“与”和“或”运算判定语音端点的方法,而是把两者有机结合起来成为短时能频值。为提高该方法对噪声的适应性,进一步引入了相对阈值的概念,它是两个时刻的语音采样的比值关系,具有相对意义。为检验这种方法的性能,用Burg法求取了线性预测倒频谱(LPC-CEP)并以它为主要参数,短时能频值作端点检测,建立了一个基于离散隐马尔可夫模型(DHMM)的语音识别系统,经过实验验证,平均识别率达到了91.4％,证明了这种时域参数的良好性能.
This paper studies a method of voice endpoint detection using Energy-Frequency-Value, which is different from the traditional method of using short-term energy and short-term average zero-crossing rate to judge whether it exceeds the threshold, The “and” and “or” operations determine the end point of speech, but combine the two organically to become short-term frequency values. In order to improve the adaptability of the method to noise, the concept of relative threshold is further introduced. It is the relative ratio of the two sampling points. To test the performance of this method, the LPC-CEP was obtained by the Burg method and taken as the main parameter. The short-time frequency value was used as the endpoint detection. A discrete Hidden Markov Model DHMM) speech recognition system, the experimental verification, the average recognition rate reached 91.4%, proving the good performance of this time-domain parameters.