A Robust Music Retrieval Method for Query- by-Humming

In this paper, we present a novel melody representation and matching method, which is both robust against pitch errors and invariant to liner or non-linear tempo variation. The melody of a music item or a query is represented by a point sequence, which is derived from the pitch contour of the melody. This point sequence is invariant to the time or speed in the original melody contour. Important anchor points in the sequence, called melody skeleton, are used for melody searching and alignment. A specialized dynamic programming algorithm is proposed for robust melody skeleton matching and alignment. The melody similarity measurement of the whole point sequences is then computed based on the precise alignment. We achieve (1) tempo invariance by using the novel melody representation, (2) robustness to pitch error by using melody skeleton matching, and (3) high retrieval accuracy by using melody alignment. Furthermore, the melody skeleton matching can serve as a filter that rejects the large portion of the wrong candidate, which promote the retrieval efficiency. Abstract The increasing availability of digital music has created a need for effective music retrieval methods. In this paper, we present a novel content-based music retrieval method that is robust against pitch errors and tempo variations in the queries, which is especially advantageous for query-by-humming. The melody of a music item and that of a hummed query are represented by point sequences, and the skeletons of the melody are used to do melody searching and alignment. A novel point skipping dynamic programming algorithm is proposed for robust and efficient melody skeleton matching. The melody similarity measure is then computed based on the alignment of the point sequences. Our experiment, including a comparison to our previous methods, has demonstrated the performance of the method.