基于耳蜗谱图纹理特征的声音事件识别

Sound event recognition based on texture features of cochleagram

针对在各种环境下声音事件的识别问题,提出了一种基于谱图纹理特征的声音事件识别方法。首先,将声音信号通过伽马通(Gammatone)滤波器组,使原始声音样本转化为灰度耳蜗谱图;然后,对谱图进行曲波(Curvelet)变换,得到不同尺度、不同方向的Curvelet子带;再采用改进完全局部二值模式(Improved Completed Local Binary Pattern,ICLBP)提取Curvelet子带的纹理特征,并生成分块统计直方图,将统计直方图级联作为一种新的声音事件特征;最后,使用支持向量机作为分类器对16种声音事件在不同噪声和不同信噪比下进行识别。实验结果表明,所提特征与其他声音特征相比,可以有效识别各种噪声环境下不同种类的声音事件。

A sound event recognition method based on texture features of cochleagram is proposed for improving sound event recognition in various environments.Firstly,the original sound sample is converted into a grayscale cochleagram by Gammatone filter bank.Then,the cochleagram is processed by Curvelet transform to obtain Curvelet sub-bands with different scales and directions.The texture features of Curvelet sub-bands are extracted by using the improved completed local binary pattern(ICLBP)to generate the block statistical histograms which are cascaded as a new sound event feature for recognition.Finally,the support vector machine is used as a classifier to identify 16 kinds of sound events under different noise environments and different signal-to-noise ratios.The experimental results show that the proposed algorithm can effectively identify different kinds of sound events in various noise environments compared with other sound features.