基于多尺度样本熵与阈值的语音端点检测

Speech Endpoint Detection Based on Multi-scale Sample Entropy and Threshold

针对样本熵对突变噪声敏感导致的误检问题,提出一种改进的语音端点检测算法。该算法在时域采用尺度因子对语音信号进行多尺度变换,计算各尺度下的样本熵和阈值,统计样本熵大于门限阈值的尺度个数并与总尺度个数进行比较,实现语音端点检测。实验结果表明,该算法能够较好地消除样本熵对突变噪声的敏感性,并且与近似熵和样本熵检测算法相比,在低信噪比条件下具有更高的检测准确率。

In order to overcome the defect that sample entropy can be falsely detected due to its sensitivity to the suddenly changing noise,this paper proposes a speech endpoint detection algorithm. This algorithm does the multi-scale transform for the speech signal in the time domain. The sample entropy and threshold of different scales can be calculated. The number of the sample entropy which is greater than the threshold of corresponding scale is counted and compared with the number of total scale to realize speech endpoint detection. Experimental results show that this algorithm can eliminate the mutation noise sensitivity of the sample entropy, and the detection accuracy is well improved in the low Signal Noise Ratio （SNR） conditions, compared with approximate entropy and sample entropy detection algorithms.