A melody retrieval system on parallelized computers

This paper describes a method for a WWW-based melody-retrieval system, takes a melody sung by a user as a search clue and sent over the Internet and uses it to retrieve the song’s title from a music database of standard MIDI files(SMF). It was difficult to build a melody-retrieval service with a large database and with a lot of user accesses since it was quite difficult to build a system which could achieve both quick search and high matching accuracy. We propose a method of a scalable melody-retrieval system which achieves 70% matching accuracy against more than 20,000 pieces of music and its search time is within a few seconds. Introduction In building melody-retrieval services on the Internet, it is quite important to keep quick and accurate retrieval against a large database. In previous paper [7], we developed an effective indexing method for melody sequences. It was, however, difficult to keep matching accuracy when we limited the size of index data for reducing the search time. There was about 5 % matching accuracy lost as compared with the case where an index is not used. In addition, our previous system with a large database which contained 10,000 melodies could accept only one user access every second since the performance was limited by the ability of a computer. To solve these problems, we propose a method of parallel-ized melodyretrieval servers. The structure of server network is scalable. It can consist of only a server on a computer and it can also be extended to any number of server computers. Testing of our method with parallel-ized 5 PCs against a database containing 21,500 melody sequences showed that its retrieval time was 5.5 seconds on average. Its matching result was 70% for 1,200 inputs from 3 different people.