Séminar from Francesco Foscarin, CEDRIC computer science lab
Localization: IC2, boardroom
Speaker: Francesco Foscarin
The musical score: a challenging goal for Automatic Music Transcription
The musical score is a complex semiotic object that excels at conveying musical information in a human readable format. A lot of music nowadays is available exclusively as recorded performances, so systems able to automatically transform those performances into musical scores would be extremely beneficial for performers and musicologists. This task, called automatic music transcription (AMT), comprises a large number of subtasks that generically involve the sequential transformation of the performance higher level representations. Despite the score being the ultimate goal, most of the approaches in the literature stop short of such an output, only considering lower level formats such as MIDI files. Indeed, the handling of musical scores constitutes a great challenge for automatic approaches.
We believe that a clear model of the information contained in a musical score would help the development and the evaluation of such “full” AMT systems. In particular we advocate for a clear separation between the music that the score is encoding (i.e. the musical content) and the set of notation symbols that is employed to represent it (i.e. the score engraving). This presentation is structured in four parts: the first details a dedicated model for AMT and related tasks. The second describes a system for monophonic MIDI performance parsing, that jointly performs note quantization and metrical alignment to build the musical content level of the score. The third part focuses on the graphical symbols in the score, with a tool that computes and displays the differences at graphical level between two musical scores. Finally we present a new dataset of annotated classical music performances aligned with musical scores; we employ a semi-automatic workflow that relies on the information in the musical content of the score, and on MIDI alignment tools, to speed up the tedious task of manual annotation.