基于鲁棒音阶特征和测度学习SVM的音乐和弦识别  

Musical Chord Recognition Based on Robust Pitch Class Profile and Metric Learning Support Vector Machine

作　　者：王蒙蒙[1] 关欣[1] 李锵[1] WANG Meng-meng GUAN Xin LI Qiang(School of Electronic Information Engineer, Tianjin University, Tianjin 300072, Chin)

机构地区：[1]天津大学电子信息工程学院,天津300072

出　　处：《信号处理》2017年第7期943-952,共10页Journal of Signal Processing

基　　金：国家自然科学基金资助项目(60802049;61471263);天津市自然科学基金重点项目(16JCZDJC31100)

摘　　要：和弦识别是音乐信息检索领域重要的研究内容之一,在信息处理、音乐结构分析以及推荐系统等方面具有重要的作用。为了降低人声对和弦进程的影响且恢复和弦所对应的谐波信息,文章分别对频谱中和弦所对应的谐波信息和人声信息进行建模,构建双目标优化问题,对和弦所对应的谐波信息进行有效重建,同时去除人声;其次,对谐波信息进行降维处理得到鲁棒性的音阶轮廓特征;最后为了提高支持向量机性能,文章采用测度学习的方法得到马氏距离,并使用马氏距离替换支持向量机的高斯核函数的欧氏距离,使得支持向量机的判别函数包含有数据的空间分布信息。最终实验结果表明,同基于现今流行的和弦识别算法相比,提出的和弦识别算法识别正确率提高3.5%~12.2%。Chord recognition is an important aspect of Music Information Retrieval, which plays an important role in infor- mation processing, musical structure analysis and recommender system. In order to reduce the influence of voice on chord progression and recovery harmonic information of chord, harmonic and voice component were modeled and a two-target opti- mal problem was constructed. Solving the optimal problem, harmonic structure was reconstructed and sparse voice was re- moved. Then, through performing a pitch mapping step, robust pitch class profile was obtained. At last, a Mahalanobis Metric was obtained from feature space of train samples through metric learning, then Euclidean Metric was replaced by Ma- halanobis Metric in Radius Basis Function of SVM. Mahalanobis Metric contains distribution information of specified real dataset, so the classification result is more robust, Compared with currently popular chord estimation algorithm, Results show proposed system improves the accuracy ratio of 3.5% - 12. 2% on chord recognition.

关 键 词：和弦识别 音阶轮廓特征 核范数 测度学习 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]