中文摘要：

回声消除是提高通信中语音信号质量的关键技术。其主要难题是回声路径估计的自适应算法的控制逻辑。为了达到较好的回声消除效果，自适应滤波器需要在双端发音模式下缓慢更新或停止更新，而在其他模式快速学习。现有的双端发音检测算法没有考虑检测延时问题，使得滤波器在停止更新前已经发散，严重影响了回声消除的效果。针对该问题，在滤波器收敛时回声消除至少达到10dB的假设前提下，对传统的能量比较法进行改进，提出低延时的解决方案。实验结果表明，该方法比相关比较法的检测延时减少了35毫秒以上。

英文摘要：

Echo cancellation is a key part of speech communication. One of the main design issues is the control logic for the adaptive filter which estimates echo path. To avoid divergence of the adaptive filter, the filter coeffi- cients should be frozen in the presence of double-talk and adapt as quickly as possible once double-talk is absent. However the filter may have diverged before double-talk is detected due to detection delay issue. A low detection de- lay algorithm is presented based on energy comparison approach with the assumption that echo suppression should be more than 10 dB once adaptive filter has been converged. The experimental results show that, compared with cross-correlation approach, detection delay of the algorithm is reduced about 35 ms.