A Three-Layer Approach for Music Retrieval in Large Databases

An effective music information retrieval (MIR) system should provide fast queries to music databases taking into account musical features relevant to the task, such as transposition invariance, polyphony of music and the fact that there might be some ‘extra intervening musical elements’ (such as grace notes) within the database occurrence of the query pattern. The importance of efficiency is due to the need to find musical documents in possibly enormous databases. In this paper, we introduce an approach using three different matching layers each of which is possible to find transposition invariant occurrences of given musical query patterns in polyphonic music databases. The advantage of the layers is in sorting them in an order of decreasing efficiency as regards to the speed, but in an increasing order of carefulness put in the searching process. Thus, a repeatedly occurring musical pattern used as a query pattern should be found very efficiently, while searching for a query pattern corresponding to a rare database occurrence with arbitrarily many extra intervening elements is allowed to take more time.