###
DOI:
本文二维码信息
基于深度学习和手工设计特征融合的翻唱歌曲识别模型
(华东理工大学 信息科学与工程学院)
Cover Song Identification Based on Fusion of Deep Learning and Manual Design Features
Yang Mei1, Chen Ning2
(1.School of Information Science and Engineering,East China University #$NLof Science and Technology,Shanghai,200237;2.China)
摘要
相似文献
本文已被:浏览 207次   下载 41
投稿时间:2017-07-04    修订日期:2017-08-28
中文摘要: 在翻唱歌曲识别中,手工设计的特征虽然具有高可定制性,但其采用的浅层线性结构难以表现音乐的非线性长效结构;而基于深度学习的特征提取算法分析音乐的非线性动力学特性可以弥补这一缺陷。本文在研究两者互补性的基础上,提出一种融合手工特征和深度特征的翻唱歌曲识别算法。该算法分别利用深度学习模型和手工设计算法提取歌曲的音级轮廓特征和旋律特征(Melody, MLD),然后将基于这两种特征的相似度组合成相似度向量输入到改进的SVM模型中,并将输入歌曲属于翻唱组合的概率作为融合相似度。为了验证算法性能,以两个公开的数据库(covers80, covers1212)作为测试对象对算法性能进行测试。实验结果表明该算法比基于单个特征的算法和基于相似度融合的算法取得了更高的识别率和分类准确率。
中文关键词: 特征融合  深度学习  翻唱歌曲识别  SVM
Abstract:The hand-engineered features used in cover song identification are highly customizable, but shallow processing can not express the dynamic characteristics of music. While the features extracted by deep learning algorithm can express the nonlinear structure and long-term feature of music. In that case, we propose a method to fuse these two features after studying the complementary. The proposal method trains a deep learning model to extract the Deep Pitch Class Profile (DPCP) feature; and extracts the Melody (MLD) feature by a hand-engineered method. Then, we put these two features together as the input of an improved SVM model. Finally we calculate the probability of whether the test song is a cover song as the fused similarity. For experiment, we test the effect on two public databases (covers80, covers1212) and compare ours with other one-feature and multi-feature method, the result shows that the proposal method has a higher recognition rate and classification accuracy.
文章编号:20170704003     中图分类号:TP391    文献标志码:
基金项目:国家自然科学基金资助项目(61271349)
引用本文:
杨妹,陈宁.基于深度学习和手工设计特征融合的翻唱歌曲识别模型[J].华东理工大学学报(自然科学版),DOI:.

用微信扫一扫

用微信扫一扫