Mark Lister Undergraduate Dissertation 2016/17

Acoustic Scene Classification

Supervised by J.Barker

Abstract

As humans we use information we extract from sound all the time to give us clues about the world around us. There is a wealth of information in these sound signals which computers can use to also learn information about the environment they are in, extracting this and classifying the resulting data is known as acoustic scene classification. This report investigates the background literature to this problem before two novel feature extraction techniques, sound texture and i-vectors are chosen, and used to improve an MFCC baseline system. These techniques both showed specific perfor- mance improvements on certain scenes. Early and late fusion techniques were applied to combine the three features, leading to a final system combining all three features together. This final system was used on the DCASE challenge data to reduce the error of the provided baseline system by 13.6%.