DISOPT Seminar at EPFL, Bernoulli Centre, with members of the Chair of Discrete Optimization and DCML,
abstract: Music transcription is the problem of converting a music performance into music notation. We consider in this work the case where the input performance is given by a sequence of timestamped events, typically a MIDI file, whereas the output is a structured music score (e.g. in XML format). In order to tackle the problem of extracting a structure from linear input, we rely on parsing techniques, using an apriori hierarchical model of music notation given by generative tree grammars.
In this presentation we shall see how this approach helps to make relevant and inter-related decisions, and compute transcription solutions optimal with respect to both the fitness of the output to the input, and a measure of readability of music notation. Some transcription case studies for monophonic and drum datasets will also be presented.