Abstract: Through analyzing the humming melodies pitch extraction and retrieval algorithm, a complete system framework of Query By Humming（QBH） is proposed. It includes the melody feature extraction and approximate melody matching in MIDI music database. The Mel Frequency Cepstral Coefficients（MFCC） is extracted. Through analyzing the theory of DTW algorithm, the cosine similarity of the delta duration sequence is added with the characteristics of the sound for performance improvement of the system. Experiments are conducted in a test set of 340 MIDI songs. The system gets a success rate of top-3 increased by 3.7% and a 16% time reduction.

Key words: query by humming, pitch track, melody match, dynamic time wrapping

摘要： 通过研究哼唱旋律基频提取和检索算法，给出了一个完整的基于哼唱的音乐检索系统框架。系统主要分析了旋律特征提取和近似旋律匹配部分。旋律特征提取部分采用基于差分Mel倒谱法求基频；旋律匹配部分对经典的动态时间弯折算法原理分析后，根据声音特征引入音长差序列的余弦相似度，提高了检索效率和精度。在340首MIDI歌曲的测试集上，前三位识别效率提高3.7%，用时降低16%，系统的性能有明显改善。

关键词: 哼唱检索, 基频提取, 旋律匹配, 动态时间弯折