Mathieu Giraud

Can a Computer Understand Musical Forms?

Musical forms structure the musical discourse through repetitions and contrasts. The forms of the Western common practice era (binary, ternary, rondo, sonata form, fugue, variations...) were widely studied by music theorists, and often formalized and codified centuries later after their emergence. The musical forms are used for pedagogical purposes, in music analysis as in composition, and some of these forms (like variations or fugues) are also a principle of composition found inside large-scale works. As computer scientists working in music information retrieval (MIR) on symbolic scores, computational analysis of musical forms is challenging.  Very constrained forms (like a ‘school fugue’) can have rather consensual analysis and be good benchmarks for MIR algorithms. Can algorithms really make advances in musical form analysis?  On elaborated structures, MIR analysis possibly helps systematic musicology studies, but actual musicological or aesthetic considerations are still to be done by human expertise. More generally, symbolic MIR methods on musical forms are in their infancy and are very far from the analysis done by musicians: thinking on codification of musical forms will help the development of better MIR algorithms. I will present our results in computational music analysis on two cases (fugues and sonata forms) using both discrete tools (based on string comparison) and statistical approaches. I will present our benchmark data, and will also discuss the computational gap between the analysis of short simple forms and the analysis of works that deviate from the standard schemas.