一种基于帧-音符方式的哼唱检索算法

A frame-to-note algorithm for query by humming

为了提高哼唱检索旋律匹配的速度和精度,提出了一种基于帧-音符方式的匹配算法。该算法针对旋律曲线的形状特点,采用基频序列表示哼唱片段,采用音符序列表示模板片段,根据累积权重估计基频跳变点位置,然后计算哼唱片段和模板片段之间的编辑距离。在MIREX08数据库上进行的实验结果表明:该算法检索时间为动态时间规整算法的0.013倍;与动态时间规整算法结果进行融合,最终平均排序倒数精度指标可以达到91.2%。

This paper presents a frame-to-note（FTN） algorithm to improve the speed and precision of the melody match for querying by humming（QBH）.According to the characteristics of tune curves,the humming phrase is denoted by the pitch sequence while the symbolizing template phrase is denoted by the note sequence.The pitch transition position is estimated based on the predefined weights,with the edit distance between the two phrases then calculated.Experimental results using the MIREX 2008 corpus show that the retrieval time of the FTN algorithm is 0.013 times that of the dynamic time warping（DTW） algorithm,and that the final fusion precision achieves a mean reciprocal rank of 91.2%.