基于近似子乐句的翻唱歌曲识别研究

翻唱歌曲识别是音乐相似计算研究中的问题之一。通过对表征音乐和声的Chroma特征的分布分析,进行了基于近似子乐句的翻唱歌曲识别研究;在已收集的翻唱歌曲数据集上,完成了歌曲调的归一化处理、结构分割和特征抽取;比较了特征降维后,在EMD距离矩阵、欧式距离矩阵和余弦距离矩阵等不同测度下的音乐相似度计算。通过与MIREX中翻唱歌曲识别结果的比较表明,本研究中基于近似子乐句的研究方法在准确率上有一定程度的提升。

Study on cover songs identification based on phrase content

Cover songs identification has been a major topic in the field of music similarity computing for a long time. A new idea of using the music phrase content to identify coves songs was proposed by analyzing the music harmony feature-Chroma and its distribution. The key normalization, structure segmentation and feature extraction were used to manipulate the cover songs of the collected data sets. Then several distance matrixes, including EMD distance matrix, Euclidean distance matrix and Cosine distance matrix, etc, were compared to compute music similarities after feature dimension reduction. The result is better than MIREX2011′s cover songs identification results.