Acoustic scene classification using spatial spectrum estimation

Abstract

Analysis of audio from our surroundings gives us important cues about the acoustic scene, with automatic analysis usually done by sound event detection or analysing the audio scene as a whole. On the other hand, inspecting the auditory space characteristics, or openness of the space, is a much less studied aspect. This thesis aims to study the classification of audio scenes based on the aforementioned auditory space characteristics with the use of different audio features. In this work, log-mel band energies and spatial spectrums for the audio recordings are calculated and used in the classification. The results revealed that best performance is obtained when using the combination of mel features and spatial spectrum, instead of either one of them. It was also observed how any differences inside a class can affect the results.