Harry Howarth Undergraduate Dissertation 2017/18

Sound Event Detection for Smart Cars

Supervised by J.Barker

Abstract

Sound events are everywhere, the popularity of including the detection of them in context aware systems such as smart cars and smart cities is rapidly increasing. This project produced and experimented with a system for classifying sound events using a Convolutional Recurrent Neural Network. The system was evaluated using the same metrics used in the 2017 DCASE Challenge so that they could be compared to the results of the systems entered into their competition.

Sound Event Detection for Smart CarsThe experimentation in this project was using different training techniques, such as data augmentation, to try and improve on the results of a sound event detection system with a state of the art architecture.

The data augmentation techniques leveraged managed to achieve results that improved on this projects baseline system. For audio tagging this project's best F-Score improvement over the baseline was 15.3\%. For sound event detection Segment Based Error Rate was 25.4\% lower in this project's system, that leveraged data augmentation, compared to the system that won the 2017 DCASE Challenge, which did not use data augmentation.