Abstract: The process of speech signal recognition and de-noising is usually used only for speech signal segments, and in some situations, different approaches are need to be adopted to deal with speech segment and noise segment. Correlation function is used to describe the associate degree of the signal value at different times. Due to the randomness of noise, there is a great difference between the correlation function of noise and the correlation function of speech, which can be employed to the endpoint detection of speech. Based on this idea, a method is presented in this paper which uses auto-correlation function. The comparison is also made with the classical method of speech endpoint detection which is the double-threshold decision method based on the short-time average energy and short-time average zero-crossing rate. Experimental results show that the presented methods has higher accuracy, and can achieve better results than the short-time average energy and short-time average zero-crossing rate in low SNR.

Key words: endpoint detection, auto-correlation function, zero-crossing rate, double-threshold

摘要： 在语音信号的识别、去噪等过程中通常只需对有声段进行处理，并且对语音段和噪声段可能需要采取不同的处理方法。相关函数描述的是随机信号在不同时刻取值的关联程度，由于噪声的随机性，噪声的相关函数和语音的相关函数有很大的不同，利用此不同点可以进行语音的端点检测。基于此提出了自相关函数的语音端点检测方法，并对比了经典的双门限法——基于短时平均能量和短时平均过零率的双门限判决法。实验表明该方法具有较高的准确性，并且在较低信噪比下能取得比短时平均能量和短时平均过零率的判决法更好的效果。

关键词: 端点检测, 自相关函数, 过零率, 双门限法