For the music retrieval using query-by-humming, the correct segmentation of the humming note is an important guarantee to extract the effective music melody feature information. In this paper, a two-level note segmentation method based on double-threshold endpoint detection and fundamental frequency curve fluctuation measure is proposed, then melody is extracted after segmentation, and the humming music retrieval system with the improved dynamic time warping algorithm is constructed. Experiment results show the segmentation correct ratio reaches 90%, the retrieval correct rate is improved by 5% and the retrieval time is reduced by 12.6% compared with the existing traditional humming music retrieval system.