Towards Data-Driven Models of Rhythm
Our goal is to learn audio-driven models of rhythm to be used in music similarity tasks and creative music generation. In music information retrieval, much rhythm-related research has focused on the transcription of rhythm from audio: onset, beat, downbeat, meter, and tempo estimation. With accurate transcriptions, rhythmic similarity can be computed with several proposed methods. However, algorithmic solutions to rhythm transcription are still error-prone and often break down for non-western musics, which are structured differently than western music. Some rhythmic similarity measures have been proposed that do not rely on accurate transcriptions, but these methods neglect key aspects of rhythm or have made invalid assumptions in their evaluations. In this talk, I will discuss our efforts thus far to learn models of rhythm from audio that overcome these issues.