Section: New Results

Weakly-Supervised Discriminative Model for Audio-to-Score Alignment

We consider a new discriminative approach to the problems of segmentation and of audio-to-score alignment. For each musical event, templates have to be built or learnt before performing any alignment. Because annotating a large database music files would be a tedious task, we develop an original approach to learn templates without annotations, but only the knowledge of the music scores associated to music files. We consider the two distinct informations provided by the music scores: (i) an exact ordered list of musical events and (ii) an approximate prior information about relative duration of events. We extend the celebrated Dynamic Time Warping algorithm (DTW) to a convex problem that learns optimal classifiers for all events while jointly aligning files, using this weak supervision only. We show that the relative duration between events can be easily used as a penalization of our cost function and allows us to drastically improve performances of our approach. We describe in details our approach and preliminary results obtained on a large-scale database in [18] .

This work was done in collaboration with the SIERRA project-team at Inria Paris.