中文摘要：

在频域应用高阶统计量（High order statistics，HOS），提出一种基于幅度谱HOS新特征的语音端点检测（Voice activity detection，VAD）算法。算法利用相邻帧获取当前帧的统计信息，并用幅度谱构造独立零均值高斯随机序列，通过计算此序列的归一化偏度来得到HOS特征。新特征利用了噪声的长时平稳特性和无序性的先验信息，借用语音生成模型来分析噪声模型，并通过合理的假定，提取潜藏在幅度谱中的高斯信息。因此相比传统HOS特征只能用于高斯或准高斯白噪声检测，幅度谱HOS适用范围扩展到包括有色噪声在内的所有平稳随机噪声。同时新特征表现出许多优异的特性，如：平稳噪声的特征值趋近于零；语音间隙噪声段和语音结束时呈现出负峰特性等。利用这些特性可以建立适用于不同类型、不同信噪比、且具有随机切入点的强鲁棒性能的VAD算法。文章详细阐述了新特征的原理以及特性，并结合判决准则构造了一个简单的VAD算法。实验结果表明，对于平稳噪声基于幅度谱HOS的VAD算法，在检测的准确性和算法鲁棒性的综合性能上优于基于传统特征的算法。

英文摘要：

By applying high order statistics （HOS） in spectrum, an algorithm for a new feature, called the spectral HOS, is developed. Firstly, the statistics information of current frame from nearby frames is derived, then from the spectrum, the hidden Gaussian sequence with zero mean is abstracted, and finally the proposed feature is constructed by calculating the normalized skewness of the sequence. In the new feature, the priori information, such as noise character istic of long time stationarity and nature of Gaussian signal, is considered, and the speech production model is introduced to analyze and find out the hidden Gaussian information. Compared with traditional features, the new feature can be applied to all stationary noise detection unlimited to Gaussian or Gaussian-like types. Furthermore, some good characteristics are exhibited, e.g. the value of feature is approached to zero in stationary noise segments ; and negative peaks appear in the speech pause and end segments, ect. With the feature, a VAD algorithm is constructed. As a result, the proposed algorithm can be adapted in complex environments when the noise type, SNR or the starting point is varied. Experimental results demonstrate the better performance in accuracy and robustness compared with the traditional featurebased algorithms.