Key information:
| Student | Hsuan-Yang Wang |
| Academic Supervisor | Christine Evers, Philip Nelson |
| Cohort | 1 |
| Pure Link | Active Project |
It has been well known for many years that the signals entering the two ears are used in combination by the human auditory system to enable the signal from a sound source to be isolated from those from many other sources present. This has long been known as the “cocktail party effect” where a listener can focus on a single conversation with another speaker, despite the surrounding background noise generated by multiple other conversations. This phenomenon has been widely studied and the extent to which the human hearing system can isolate, for example, a single speech sound in the presence of background noise, is reasonably well established. However, this remarkable property of the human hearing system is far from completely understood.
The objective of this project is to emulate for machine listening the human ability to detect, localize, and focus on multiple, simultaneously active sound sources in the cocktail party scenario. The project will investigate computational models of the human auditory system, including, and may consider extensions based on neuronal dynamics. Novel algorithms for deep learning will incorporate the proposed models for tasks, such as sound event detection & classification, sound source localization & tracking, or blind source separation. Computational aspects will be considered in order to benchmark the practical feasibility of the developed algorithms on embedded systems.