Abstract: It is a challenge to detect voice endpoints in the condition that includes stationary noise and instantaneous noise. This paper presents a method that uses self-adaptation detection thresholds and multi layer perceptron to recognize noise and voice based on the selected anti noise features. Experimental results show that the correct voice endpoints detection rate is 27% higher by using the selected features than using conventional frame energy and cross zero rate in 0 dB SNR, and the use of multi layer perceptron achieves 94.47% isolated instantaneous noise recognition rate in normal condition, those types of noise include the sound of opening and closing door, cough sound, sound of turning pages and sound of breath, etc.

Key words: voice endpoint detection, speech processing, anti noise feature, multi layer perceptron

摘要： 在既有平稳噪音又有突发噪声的环境下进行语音端点检测是一项挑战。在选择抗噪特征的基础上，提出了自适应判定阈值和用多层感知器进行语噪鉴别的语音端点检测办法。实验结果表明，选择的语音参数比传统的帧能量和过零率在信噪比为0 dB时，正确的语音端点检出率高出27%，而多层感知器在正常环境下，检出94.47%的开关门声、咳嗽声、翻书声和呼吸声等孤立突发噪声。

关键词: 语音端点检测, 语音处理, 抗噪特征, 多层感知器