Music Retrieval as Text Retrieval: Simple Yet Effective

This poster reports on the latest findings of the author concerning the development of a simple approach to the storage and retrieval of music information. Using the McNab et al. collection of 9354 folksongs [1], we have developed and tested a series of test databases where monophonic melodies are represented as a collection of interval-only n-grams (i.e., length-n substrings of the signed differences between pitches). These melodic n-grams of length-4, 5 and 6 can be treated like artificial "words". By treating n-grams as "words" we have been able to apply traditional text retrieval methods to the music information retrieval (MIR) problem. The traditional text retrieval system, SMART, was used to test the hypothesis that music information can indeed be treated as text. Randomly selected extracts from the databases were used to simulate potential queries. The results were evaluated using the standard text retrieval normalized precision and recall measures. Several test databases performed very well, confirming the notion that a simple, text-styled, approach to MIR is indeed feasible