一种基于改进信息熵的语音端点检测方法研究

Voice Activity Detection Method based on Modified Speech Information Entropy

端点检测是指从一段包含语音的信号中检测出语音的起始点,是语音处理中的重要部分。常用的语音端点检测方法有短时能量和信息熵端点检测方法。但是,短时能量法抗噪声性能较差,而信息熵算法检测精度与稳定性波动较大,端点处的检测曲线较平缓。针对以上问题,结合两种方法,提出了一种基于改进信息熵的端点检测算法,即信能比算法。实验结果表明,该算法较短时能量和信息熵算法,具有更好的检测性能和检测精度,且抗噪声性能更好。

Endpoint detection, as an important part of speech processing, detects the starting point of speech from a signal containing speech. The commonly-used speech endpoint detection method involves short-time energy and information entropy endpoint detection method. However, the short-time energy method is poor in noise resistance, while the information entropy algorithm fairly fluctuated in detection accuracy and stability, and the detection curve at the endpoint relatively smooth. In combination with these two methods, a new endpoint detection algorithm based on modified information entropy（SNR） is proposed, and the experimental results indicate that as compared with short-time energy and information entropy algorithm, this proposed algorithm has better detection performance and detection precision, and also better anti-noise performance.