Teaching Machines to Listen

Sound event detection from audio has many applications, including assistive technologies for the elderly and hearing impaired, monitoring of industrial activity, noise, or biodiversity. State-of-the-art methods rely on annotated data for learning sound classes, but often there is a gap between how humans are recognizing and annotating sounds and how automatic methods need and are able to provide this information. To date, there is no standard method for collecting and annotating data. Moreover, there is no relevant research on directly utilizing the noisy data produced through manual annotation. The goal of this project is to develop methods for using subjective, noisy data produced by multiple, non-expert annotators, such that the system learns from human opinion, with typical variability of labels and temporal consistency. In addition, the project will develop subjective, user-friendly measures, that reflect human tolerance for certain type of confusions and misalignments.