一种简单的噪声鲁棒性语音端点检测方法

语音端点检测是语音识别系统非常重要的组成部分,一种理想的语音端点检测方法,在噪声环境中要具有较强的鲁棒性.为了提高检测方法在噪声环境中的鲁棒性,在短时能量的基础上,结合谱平度和幅度谱的主频率特征,分别进行判决,再采用投票决策机制确定端点检测结果,提出了一种比较理想的语音端点检测方法.实验结果表明,与传统的短时能量法和短时TEO能量法相比,该算法在各种加性噪声下具有良好的鲁棒性,在较低信噪比下仍能准确地区分有用信号和噪声,验证了该算法的有效性.

A Simple Algorithm of Robust Speech Endpoint Detection

Speech endpoint detection is a crucial component in speech recognition.An ideal speech endpoint detector needs to be robust.In order to improve the robustness,on the basis of short-term energy,the proposed method is combined with spectral flatness and the most dominant frequency,which is judged on the corresponding threshold value respectively.Then vote decision-making mechanism is adopted to define the endpoint detection results.Compared with same traditional algorithms such as short-time energy and short-time TEO energy,experiment results show satisfactory performance and the method can still accurately distinguish between voice and noise in various additive noise conditions,so the effectiveness of the proposed algorithm is proved.