Abstract: In order to improve correctness of Voice Activity Detection（VAD） system under low Signal Noise Rate（SNR）, an improved approach of VAD based on Hilbert-Huang Transformation（HHT） is proposed. Every frame of signal is decomposed into finite Intrinsic Mode Functions（IMFs） by Empirical Mode Decomposition （EMD）. Then all IMFs except the first one are filtered by a bandpass filter whose passband is from 250 Hz to 3 500 Hz to choose the parts of an IMF. And then useful IMFs are weighted with different weights and transformed by Hilbert Transformation（HT） to get energy value. After that, noise threshold is evaluated through the analysis of noise feature. On the basis of threshold, the starting points and ending points are seeked out on Hilbert energy specturm. Simulation results show that the presented method not only can improve correctness of VAD, but also have strong robustness under low SNR.

Key words: voice activity detection, hilbert-huang transformation, empirical mode decompositon, hilbert energy

摘要： 为提高语音端点检测系统在低信噪比环境下检测的正确率，提出一种强噪声环境下基于改进的希尔伯特-黄变换语音端点检测方法。对每帧信号进行经验模态分解，得到有限个固有模态函数，去掉第一个固有模态函数，其他的都让其通过一个带宽为250~3 500 Hz的带通滤波器，消除部分噪声。对所选固有模态函数加权，再进行希尔伯特变换得到能量特征值。通过分析噪声特性，估计噪声阈值。在希尔伯特能量谱上，根据阈值搜索语音起点以及终点。仿真实验表明，在低信噪比的情况下，方法的准确率有明显的提高，并具有很强的鲁棒性。

关键词: 语音端点检测, 希尔伯特-黄变换, 经验模态分解, 希尔伯特能量