We are developing a method for the analysis of non-Eurogenetic musical contexts that use a combination of machine learning methods and music-theoretic analysis and ethnomusicological inquiry. Our machine learning methods do not rely on explicit assumptions of structure, do not require annotated training data, and critically, are the first known methods which allow the practitioner to combine arbitrarily many input representations, each of which may encode different aspects of music (pitch, rhythm, harmony, instrumentation, etc.) that may or may not become informative for structure. Moreover, this approach provides simultaneous analyses at multiple levels of granularity, allowing the viewer to dynamically adapt the resolution of the analysis to suit their interest. Our music- theoretic and ethnomusicological inquiries provide the necessary cultural bias (aka “domain knowledge”) to improve the algorithms in their understanding of these musics.
Projects: