Abstract: A Toeplitz de-noising method using the maximum eigenvalue is proposed for the voice activity detection at low SNR scenarios. This method uses the self-correlation sequence of speech bandwidth spectrum to construct a new symmetric Toeplitz matrix and to compute the largest eigenvalue, and the double decision thresholds in the largest eigenvalue are  applied in the decision framewok. Simulation results show that the presented algorithm is more effective in distinguishing speech from noise  and  has better robustness under various noisy environments. Compared with novel method of recurrence rate analysis, this algorithm shows lower wrong decision rate. The algorithm is of low computational complexity and is simple in real-time realization.

Key words: voice activity detection, speech bandwidth spectrum, maximum eigenvalue, robustness

摘要： 针对在低信噪比条件下语音端点检测问题，提出了一种基于Toeplitz最大特征值的去噪语音端点检测方法。该方法用语带频谱自相关序列构造一个对称Toeplitz矩阵，利用该矩阵最大特征值的信息量对语音信号进行双门限端点检测。新算法经过实验，能够有效地区分语音和噪声，在不同的低噪声环境条件下具有良好的鲁棒性。与新近的信号递归度分析方法比较，准确率较高。该算法计算代价小，实时性好，简洁易实现。

关键词: 语音端点检测, 语带频谱, 最大特征值, 鲁棒性