Abstract: A two-layer neural network for humming feature extraction is proposed for note segmentation and note recognition，when users sing consecutively with arbitrary words.The first-layer BP network is used to divide the humming data into the independent notes.The second-layer RBF network is used to receives the MIDI pitch of the independent note.The experimental results show that this method can extract the humming features with high accuracy，and it’s suitable to the QBH（Query By Humming） system.

摘要： 针对用户以任意字词连续哼唱的情况下，哼唱特征提取中音符分割、音符识别难度大的问题，提出了一种基于两级神经网络的哼唱特征提取方法。第一级采用BP神经网络实现哼唱音符分割，得到独立音符；第二级采用RBF神经网络识别分割出的各个音符，获得音符的MIDI音高值。实验结果表明，该方法能较好地完成哼唱特征的提取，适合于实际哼唱检索系统对连续哼唱的要求。