中文摘要：

关键音频检测是指从音频库中检索出查询样例,是音频检索的一种重要形式。该文针对传统关键音频检测方法在效率和鲁棒性上的不足分别在预处理、指纹提取以及检索部分进行了优化。在预处理阶段采用基于子带能量比的语音端点检测算法,并在窗函数选择和子带划分方法上进行了改善;在指纹提取阶段采用种子片段选取的方法,并将指纹提取方法改进为子带频谱质心法;在检索阶段通过设定命中次数门限以提高效率。实验结果表明：该文提出的改进系统在查全率、查准率以及抗噪能力提升的同时提高了检索效率,有效地提升了检索性能。

英文摘要：

Key audio detection, an important form of audio retrieval, uses a query audio sample to search in an audio database but such searches are not very efficient or robust. This paper optimizes the pre-proeessing, fingerprint extraction and retrieval of the audio retrieval. The pre-proeessing uses endpoint detection based on the sub-band energy ratio with a modified window function and measurements of the sub-hand divisions. The fingerprint extraction uses seed fragments and spectral sub-band centroids. The retrieval part uses a threshold for the hit counts to improve the efficiency. This system improves the precision and reduces the recall rate with good noise suppression. The retrieval efficiency and performance are effectively improved.