AnEffective Design for Fast Query-by-Humming System with Melody Segmentation and Feature Extraction

A melody segmentation scheme based on pivotal points combined Deep Auto-Encoder (DAE) for Query-by-Humming (QBH) systems is investigated. In order to deal with personalization of human voice, especially the instability in rhythm, we adopt a joint melody segmentation and encoded features extraction method. Pitch sequence between pivotal points shows characteristic differences largely due to monolithic deviation on tones, dynamic diversification in vocal range and mutative scale of tempo. The pivotal points in melody involve extreme and some extreme midpoints which depict the global trends, simultaneously exhibit robust to humming melody. Based on feature extraction, a novel melody segmentation scheme is derived to solve the problem. Through the trained DAE from the deep learning, the deep level coding features in melody are acquired. The experiment results are carried out to demonstrate the validity of the proposed scheme.