基于非负张量分解的音频分类方法  

Audio Classification Method Based on Non-Negative Tensor Factorization

作　　者：杨立东[1,2] 谢湘[1] 王晶[1] 匡镜明[1] 

机构地区：[1]北京理工大学信息与电子学院,北京100081 [2]内蒙古科技大学信息工程学院,包头014010

出　　处：《天津大学学报（自然科学与工程技术版）》2015年第9期761-764,共4页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(61473041);内蒙古高校科研基金资助项目(NJZY13139)

摘　　要：为了提高音频数据分类正确率,提出一种通过非负张量分解(NTF)的分类方法.音频信号经过预处理后,提取声学特征和感知特征参数,然后构建非负的3阶音频张量,其各阶分别对应特征、帧、样本;其次,通过NTF得到每一类音频的核张量与因子矩阵,让测试样本构建的张量与各类型音频的因子矩阵的转置进行张量乘,得到对核张量的近似;最后,通过Frobenius范数进行相似性度量,完成分类.使用古典音乐、流行音乐、语音、噪声4种类型的音频数据测试分类效果.结果表明,平均分类正确率在85%,以上,说明该方法可以有效地完成音频分类.To improve the accuracy of audio classification, a classification method based on non-negative tensor factorization（NTF） was proposed. Firstly, acoustics features and perceptual features were extracted after pre- processing of audio signal. Then, a 3-order non-negative tensor was constructed, the orders being features, frames and samples, respectively. Secondly, core tensor and factor matrixes of each class of audio were obtained by using NTF. Next, test tensor was multiplied by transpose of factor matrixes of each class to obtain approximate tensor of core tensor. Finally, audio samples were classed by using Frobenius norm similarity measure. Experiments including classical music, popular music, speech and noise were provided to demonstrate the performance of audio classifica- tion. Results showed that the mean classification accuracy rate is above 85%, which proves that the proposed method can class audio effectively.

关 键 词：音频分类 非负张量分解 特征提取 因子矩阵 

分 类 号：TN912.3[电子电信—通信与信息系统]