Speedup of a time-sequence retrieval method for musical audio signals by integrating similar melodic intervals

摘要

We have developed a music retrieval system that takes a humming query and finds similar audio intervals (segments) in a music audio database. Because our previous time-sequence matching method was not efficient for a databese containing several similar intervals we organize the database into an efficient network by integrating similar intervals. For this purpose we propose a model-driven path RIFCDP (mpRIFCDP), which is an improved version of the conventional RIFCDP for finding similar intervals considering temporal warps. While the conventional RIFCDP was not able to deal with a transposed melody, our mpRIFCDP can appropriately find intervals similar to such a melody by considering multiple possibilities of transpositions. Test results show the effectiveness of the proposed method.