【作者】 许洁萍； 袁斌； 刘怡；

【Author】 XU JiePing, YUAN Bin, LIU Yi (Computer Department of School of Information, Renmin University, Beijing 100872, China)

【机构】 中国人民大学信息学院计算机系；

【摘要】 为了研究音乐检索系统中用户哼唱输入旋律轮廓上的错误,提升音乐检索系统的性能,本文首先设计开发了一个音乐样本的标注工具,实现了对用户哼唱输入的音符切分及基频提取,完成了用户哼唱输入的样本音乐到midi的转换。然后,对用户输入的检索序列,根据规则进行了插入、删除的对齐处理,并对处理后的数据进行了旋律轮廓错误的分析。最后,通过数据挖掘工具,研究了哼唱旋律数据流中,在强关联规则下音高差及音长比与旋律错误类型之间的关联关系。更多还原

【Abstract】 To study the error model for sung music queries, we compiled regular MIDI files, collected 193 hummed queries which are sung by 17 untrained subjects, belong to 50 different songs and include 3165 notes, and built an annotation tool which is used to segment music note and extract pitch of note. Thus, the sung music audio samples are transcribed as MIDI symbol queries matching to lyric and compared with regular MIDI queries that we have been written in midi database. We defined the error types including insertion error, deletion error and adding error, lessen error. The statistic results showed that insertion error is 14% and deletion error is 2% of all error, adding and lessen error can be ignored. So we defined some rules of processing insertion and deletion error, analyzed the hummed melody error distributing for insertion and deletion only. The results showed that most delta pitch concentrated on 0, ±1 and ±2, so we concluded that it is easier for people to hum the parts with smaller fluctuation for melody. The distributing results also indicated that 28% pitch melody contour is right-on, 39.2% is the same direction changing error, which is perceived acceptable and 32.93% is reverse pitch error which is perceived impossibly. In the end, the melody contours are analyzed and 2906 melody-error data streams are constructed. We used a data-mine tool of Classification Based on Associations which is built by Singapore National University, and studied the association rule among delta pitches, duration ratios and the error types of melody under the strong association rules which are min-confidence 90% , min-support 3% in right-on frequent items and min-confidence 100% and min-support 5% in completeness-error frequent items. The association results showed that the inter onset interval is more relative to user hummed error. If the delta pitch are between -3 to +3 or equal to 0, the hummed error should be corrected under the inter onset interval not changing. So the type error should be put right in an audio front end for query-by-humming systems in order to improve the performance for Music Information Retrieval. In future, we will enlarge melody database to study the results and build melody error model on finished research results in this paper.更多还原