基于单类支持向量机的异常声音检测

Abnormal sound detection based on one class support vector machine

提出基于单类支持向量机的异常声音在线检测算法。该算法针对公共场合正常的环境声音,训练一个单类支持向量机模型,用来判断声音是否属于正常的环境声音,若不是则属于需要进一步识别的异常声音。采用窗长2秒的滑动窗对声音进行分窗,对每一个窗内的声音分帧并提取梅尔倒谱系数,短时能量,频谱质心,短时平均过零率等特征。采用基于帧之间互相关系数的方法对声音自动分段。最后对分段声音的判别结果进行中值滤波。当有连续多个帧被判别为异常时判定有异常声音出现。最后检验了算法在地铁背景条件下六类异常声音的漏检率和每小时误检次数,结果表明算法能有效检测到异常声音的发生而且误检次数较低。

This paper proposes an abnormal sound detection method based on one-class support vector machine. For the normal sound in the public environment, we build an one-class support vector machine model to detect whether a piece of audio belong to the normal environment, if not, it is an abnormal sound needed to recognize. We apply a slipping window of two seconds length on the sound, frame the sound and extract features including Mel Frequency Cepstrum Coefficients（MFCC）, Spectral Centroid（SC）, Short term energy（STE）. We calculate the Cross Correlation Coefficient of the adjacent two frames, and use the Cross Correlation Coefficients to implement the segmentation automatically. Then distinguish whether one frame is abnormal, finally we apply median filter on the result. If there are more than certain successive frames are abnormal,we think the sound is a abnormal sound. Then we examine our algorithm on the background sound of subway, the result shows that our method can detect abnormal sounds effectively and the false alarm time is low.