哼唱检索中改进的动态时间规整算法

Improved Dynamic Time Warping Algorithm in Query by Humming

针对传统动态时间规整算法只考虑音高特征而不考虑音长特征的缺点,提出改进的算法,采用音高差和音长差共同构成算法中的代价函数,在此基础上实现了一个哼唱检索系统的原型。在数据库容量为115首乐曲和118个哼唱片段的测试中,该算法的前10位命中率为81.0%,前3位命中率为72.4%,性能优于4种同类算法。

Aiming at the drawback that traditional Dynamic Time Waxping（DTW） algorithm only considers pitch and hardly regards duration, this paper presents an improved algorithm with the help of note duration. Both delta pitch and delta duration are used to construct the cost function in the algorithm. A demo system of Query By Humming（QBH） is implemented. Experiments are conducted to test the algorithm with 118 humming clips and 115 songs in music database. The success rate of top-10 and top-3 of the improved algorithm are 81.0% and 72.4% respectively, exceeding the other four algorithms.