研究了基于哼唱的歌曲检索算法,提出了一个完整的哼唱检索框架,由特征提取模块、歌曲模板库模块、旋律匹配模块组成。为改善系统性能,在旋律提取部分采用基于小波变换的基音提取方式。旋律匹配模块在对传统的动态时间弯折进行分析后,对之进行了改进。在548首歌曲的测试集上,该系统的识别效果达到89.1%。
The methods of QBSH(Query by Singing/Humming) is focused on, and a frame of a complete system, which consists of feature extraction module, MIDI music database, and melody match module, is proposed. In order to improve the system performance, an improved wavelet-based method for pitch track in melody match module and an improved DTW(Dynamic Time Warping) method in melody match module are proposed. The experiment of 548 MIDI songs show that the system performance has a satisfying result.