Bachelor Thesis: Rare Sound Detection using Deep Neural Networks on Low Energy Devices

Summary: The goal of this thesis is the development of a software prototype that can detect, recognize and localize dangerous sound events. The target platforms for the system are IoT/low energy devices. Therefore, a special emphasis is placed on energy and computational efficiency.  The thesis includes building a pipeline that can detect and localize special sound events in real time, developing a Deep Neural Networ (DNN) that can classify sound events as dangerous, optimize the system for usage on IoT devices, as well as deploying and testing the system on an IoT device. The software and thesis are developed in collaboration with MED-EL as a proof of concept for a possible future usage of such functionality in a cochlear implant.

Supervisor: Michael Felderer