Robust and Efficient Pitch Tracking for Query-by-Humming

Several query-by-humming techniques have been proposed for content-based music retrieval. Reliable pitch extraction from humming is critical for such music retrieval systems to work well. Previous pitch detection methods are based on music note segmentation. Such methods require a user to separate each music note with silence or hum with specific syllables; otherwise the segmentation errors can seriously deteriorate pitch detection. This paper presents a pitch tracking method using a region growing approach, which does no rely on note segmentation. A novel harmonic analysis technique is proposed to efficiently group the partials of a harmonic in the power spectrum. The pitch tracking method operates by initially finding the frames with reliable pitch and subsequently finalizing the pitch detection by growing the pitch region. Experimental results show that the proposed method has high robustness and accuracy for continuous pitch contour extraction.