Environmental sound monitoring using machine listening and spatial audio

This thesis investigates how the technologies of machine listening and spatial audio can be utilised and combined to develop new methods of environmental sound monitoring for the soundscape approach. The majority of prior work on the soundscape approach has necessitated time-consuming, costly, and non-repeatable subjective listening tests, and one of the aims of this work was to produce robust systems reducing this need.

The EigenScape database of Ambisonic acoustic scene recordings, containing eight classes encompassing a variety of urban and natural locations, is presented and used as a basis for this research. Using this data it was found that it is possible to classify acoustic scenes with a high level of accuracy based solely on features describing the spatial distribution of sounds within them. Further improvements were made when combining spatial and spectral features for a more complete characterisation of each scene class.

A system is also presented using spherical harmonic beamforming and unsupervised clustering to estimate the onsets, offsets, and direction-of-arrival of sounds in synthesised scenes with up to three overlapping sources. It is shown that performance is enhanced using higher-order Ambisonics, but whilst there is a large increase in performance between first and second-order, increases at subsequent orders are more modest.

Finally, a mobile application developed using the EigenScape data is presented, and is shown to produce plausible estimates for the relative prevalence of natural and mechanical sound in the various locations at which it was tested.