Deep Learning (DL) has produced state-of-art results in many audio tasks. We are currently using DL approaches in neuromorphic auditory scene analysis. In this project, the student will implement a recurrent neural network model for estimating the number and identity of speakers in a conversation scenario. In particular, we focus on networks with a small memory footprint and low computational cost so that they can be implement on embedded systems. The use of the silicon DAS cochlea sensor and DVS vision sensor can also be incorporated as an additional sensor for this system. The project is suitable for a Semester or Master Thesis project.
Knowledge of Python, Matlab, some knowledge of deep networks, interest in hardware.
For more details, contact Shih-Chii Liu (firstname.lastname@example.org) and Enea Ceolini (email@example.com) and go to http://sensors.ini.uzh.ch for more information.