中文摘要：

耳语音作为人类发音的一种特殊形式，与正常语音相比具有信噪比低、元音的周期特征不明显等特性，因而耳语音处理比正常语音更为困难。耳语音处理研究的第1个关键步骤就是语音的端点检测，该文利用希尔伯特-黄变换（Hilbert—Huang Transform，HHT）中的经验模态分解（Empirical Mode Decomposition，EMD），首次提出了一种基于EMD拟合特征的耳语音端点检测新方法。利用EMD得到的内禀模态函数（Intrinsic Mode Function，IMF）能量，以其归一化拟合参数为耳语音端点检测的特征，可以准确地划分出耳语音端点。实验表明，该方法在耳语音端点检测中取得了很好的效果，在1200个信噪比为2-10dB的测试样本中，检测准确率为98．25％。

英文摘要：

Whispered speech is the especial form of people＇s pronunciation. There is lower Signal-to-Noise Ratio （SNR） in whispers and unobvious pitch waveform compared with the normal speech, so it is more difficult to process the whispered speech. The endpoint detection of whispers is the first pivotal step of whispered speech signal processing. This paper uses the Empirical Mode Decomposition （EMD） of Hilbert-Huang Transform （HHT） to solve the problem, and firstly proposes a novel algorithm of endpoint detection of whispered speech based on the fitting characteristic of EMD. Normalize the energy of Intrinsic Mode Function （IMF） obtained by EMD, and use the fitting parameters of the energy as the characteristic and then the endpoint of whispers can be easily divided. The results of experiments show that it is very useful in endpoint detection of whispers, and the accurate rate is 98.25% in 1200 samples （SNR=2-10dB）