基于短时能量的语音端点检测算法研究

研究了噪声环境下，利用短时能量为特征进行语音端点检测的问题。在采用短时全带能量为特征的基础上，提出的算法将短时高频能量作为辅助特征，同时使用了最优边沿检测滤波以及双门限-三态转换判决机制，从而保证了算法在噪声环境下的端点检测准确性和对信号绝对幅度变化的稳健性。实验结果表明，与传统的能量闻值法以及G．729中使用的VAD算法相比，提出的算法在噪声环境下具有更好的性能，是一个简单、高效和稳健的语音端点检测算法。

Speech Endpoint Detection Algorithm Analyses Based on Short-term Energy

This paper analyzes speech endpoint detection based on short-term energy feature in the presence of noise. Besides short-term full band energy feature, short-term high band energy is employed as an accessorial feature in the proposed algorithm. It also uses an optimal edge detection filter plus a three-state transition and judgment mechanism based on double thresholds, which ensure the accuracy in noisy environment and the robustness to changes in absolute levels. Experiments show that the proposed algorithm outperforms traditional energy threshold and G.729 VAD for speech endpoint detection in noisy environments and proves its accuracy, simplicity and robustness.