Modern sound recognition techniques are based on machine learning, which requires large quantities of annotated data. This limits the sound recognition to specific tasks where sufficient data are available. Project ZERO-ACOUSTIC will develop methods that allow recognizing sounds using their textual descriptions, attributes such as their material, or images. Furthermore, it will develop active learning methods that continuously improve the recognition by requesting information from humans in those cases where the recognition is uncertain. We will also model the subjectivity of human input, so that learning is possible from information provided by non-experts

2021

2025