Abstract: Speech/non-speech detection is a key factor for speech enhancement and recognition.This paper proposes a robust speech/non-speech detection method based on the forth order statistics.Taking advantage of the super-Gaussian property of the speech’s amplitude spectrum，the proposed method calculates the forth order statistics to measure the non-Gaussianity of the amplitude spectrum of each speech frame.Combining with the energy，a simple threshold classifier is used to distinguish whether the speech frame is speech or non-speech.Computer simulations show that the proposed method is effective.

Key words: speech/non-speech detection, speech enhancement, the forth order statistics

摘要： 语音有声/无声检测是影响语音增强和识别性能的一个关键因素，提出一种鲁棒的基于四阶统计量的语音有声/无声检测技术。利用语音信号的振幅谱是超高斯分布的特性，对每一帧语音信号的振幅谱，计算其四阶统计量，用来度量其超高斯性。结合该帧语音信号的能量，使用一个简单的阈值分类器，实现语音“无声”和“有声”期的检测。所提出的语音有声/无声检测技术，经实验证明具有很好的效果。

关键词: 有声/无声检测, 语音增强, 四阶统计量