基于指数门限的语音端点检测方法

Exponential Threshold Based Speech Endpoint Detection Method

提出了一种基于指数门限(ET)的端点检测方法.ET法为短时能量的概密函数(PDF)建立起统一的语音和噪声模型,根据当前语音数据的信噪比估计出最优的检测门限,并给出了最优检测门限的指数型公式.在'八六三'大词汇量连续语音数据库上的实验结果表明,ET法具有较好的检测性能,在噪声环境中表现出较好的稳健性,信噪比为0 dB时,检测正确率可达89.5%.在信噪比为0～15dB时,检测正确率要明显高于基本能量法、对数能量聚类法(LEC)以及χ2法等语音检测(VAD)方法.

A new endpoint detection method based on the exponential threshold （ET） is proposed. The probability density function （PDF） of the short-time energy is investigated, and a new consistent mathematical model is built up for the speech and the noise. Then, an optimal detection threshold exponential formula is concluded. Experimental evaluations on “863” program of China across a wide range of SNRs are given. Results show that when the SNR is 0 dB, the accuracy achieves 89.5%, and when the SNR is 0-15 dB, the ET method has good performances than the basic energy, the logarithmic energy clustering （LEC） and χ^2methods.