基于交叉递归图和局部匹配的翻唱歌曲识别

基于Qmax算法,提出了一种新的序列局部匹配算法,用于翻唱歌曲识别。该算法通过改变所使用的步长条件使得匹配过程既能防止病态弯曲又能增加局部匹配分数。为了验证该算法在翻唱歌曲识别中的有效性,采用基于节拍同步的音级轮廓(PCP)特征作为测试对象,并利用最佳移位索引(OTI)实现基调不变性;根据所提取的特征构造交叉递归图(CRP),利用提出的局部匹配算法计算序列之间的相似度。实验结果表明,该方法获得了比传统匹配算法,如动态时间规整(DTW)、互相关和Qmax算法更高的识别准确率。

Cover Song Identification Based on Cross Recurrence Plot and Local Alignment

In this paper, a new local alignment algorithm based on Qmax is proposed to identify the cover versions. By changing the step size condition, the proposed algorithm can prevent the generating of pathological warping and improve the final score of local alignment. To verify the effectiveness of the proposed algorithm in cover song identification, the beat-synchronous pitch class profile (PCP) feature is taken as test object and the optimal transposition index (OTI) is used to achieve the key invariance. According to the extracted features, the cross recurrence plot (CRP) is constructed and the similarity is computed. It is shown from the experimental results that the proposed algorithm can achieve higher identification accuracy than the traditional alignment algorithms, e. g. , dynamic time warping (DTW), cross-correlation and Qmax.