A Noise Robust Content-Based Music Retrieval System for Mobile Devices

This paper proposes a noise robust content-based music retrieval system for mobile devices. It takes the user's humming/singing audio input and queries the desired songs from music database. Since the system is deliberately designed for mobile devices, noise disturbance are inevitable in practical application. In order to improve the noise robustness of the retrieval system, we propose a new humming/singing audio feature extraction algorithm. A frame-to-note matching engine is employed to compute the similarity distance. The experimental results show that the proposed algorithm is efficient and robust under various noisy environments and SNR levels. For 91.46% queries, the correct songs can be retrieved among the top-10 matches in clean condition. About 85% average success rate of top-10 returns can be obtained in most noisy conditions. Even in low SNR conditions, the proposed algorithm can still achieve acceptable performance.