Abstract: Query by humming is a user-friendly content-based music retrieval interface. To develop a practical music retrieval system, heuristic rules are concluded based on the analysis of music database to navigate note segmentation. Both rhythm and pitch information are extracted from acoustic input, rhythm is used to locate music database region, and pitch contour is used to match melody in this region; Recurrent neural network is employed to remember pitch contour, and weight metrics of recurrent neural networks are used as indices of music database, which guarantees the retrieval time to be linear to database size. Experiment results show the effectiveness of the presented approach.