一种改进的基于倒谱特征的带噪端点检测方法

A Modified Endpoint Detection Method of Noisy Speech Based on Cepstral

影响语音识别性能的一个关键因素是端点检测的准确性。实际应用中的信噪比较低,使得某些高信噪比下性能好的检测算法不能有效地工作,影响系统的识别率。该文针对基于倒谱特征的带噪端点检测算法提出了3点改进:(1) 将语音信号经滤波后分成高低频两子带,分别进行分析;(2) 用LPC美尔倒谱特征LPCCMCC代替常规倒谱特征作为特征参数;(3) 改进噪声估计,使其具有自适应性。实验结果表明本方法在低信噪比下有较好的检测性能。

A major factor influencing the capability of speech recognition(SR) systems is the accuracy of endpoint detection .Some good detection algorithms with a high SNR can not take effect in practical use because of small SNR ,which may reduce the recognition rates. In this paper, three points are proposed to modify the endpoint detection method of noisy speech based on cepstral. The first is to filter the speech signal into two frequency bands of high and low, then analyze them respectively; The second is to take the LPC Mel cepstral coefficient (LPCCMCC) as feature parameters instead of normal cepstral coefficient ; The last point is to make it more adaptive by improving the estimation of noise. The experiments show good detection capability with a low SNR.