EECS Main
>
Events
Event Details
|
Graduate Research Seminar: David Little12 noon - 1:30
April 24, 2008
Ford ITW Auditorium
David Little
"Learning musical instruments from polyphonic audio with weak labels" | Abstract: We are interested in developing a system that can learn to recognize individual sound objects in an auditory scene where multiple sound sources may be occurring simultaneously. Here, we focus on sound object recognition as applied to music audio mixtures. To accomplish the task of learning instruments from polyphonic audio previous systems have required fairly stringent conditions on the labeling of the data. Although all of the pieces to accomplish this task using much more weakly labeled data already exist, we are not aware of anyone who has yet put these pieces together, and tried to train a system using weakly labeled data. Specifically our system learns to identify a single target instrument class, given positive examples that include that instrument in the mixture of sounds, and negative examples that do not include the target instrument class. We show in our experiments that it not only possible to learn from polyphonic audio, but that it works significantly better (78% accuracy compared to 55%) than learning from isolated examples when identifying instruments in novel mixtures.
Calendar of all GRS events. |
|
