Sound Source Localization and Virtual Testing of Binaural Audio
Citation:
O'Dwyer, Hugh, Sound Source Localization and Virtual Testing of Binaural Audio, Novel Methods for the Analysis and Assessment of Spatial Audio, Trinity College Dublin.School of Engineering, 2021Download Item:
Abstract:
This thesis is concerned with aspects of our perception of spatial audio and with novel technology used to record, analyse and present it. The field of spatial audio has become increasingly prevalent with mediums such as virtual reality and augmented reality becoming readily available to anyone with a smart phone. The sense of immersion when experiencing a virtual environment is not only reliant on visuals but also on audio with an accurate portrayal of an auditory scene crucial to creating a sense of realism. The work presented in this thesis focuses on the analysis of spatial audio including ambisonic and binaural signals. Fundamentally, this research is focused on localization with machine learning tools used to evaluate localization cues in these binaural signals.
Preliminary chapters on the perception and reproduction of spatial audio provide a relevant background to the research presented in this thesis.
The major contributions of this thesis can be described as follows:
A subjective study comparing the quality of spatial audio over different headphones is presented. Using virtual headphone test methodologies, a series of MuSHRA tests were conducted which examined both the perceived spatial quality and general audio quality preference of a variety of stimuli. The results of the study show that there is a strong correlation between the perception of general audio quality and spatial audio quality presented via headphones. This is true however only when using a transfer function based virtual headphone test methodology is used. The study shows that results found using a binaural recording based approach to virtual headphone testing are inconsistent.
This investigation into the capture and presentation of binaural audio is followed by studies in sound source localization in binaural signals using machine learning. This work explores the relationship between spectral and interaural cues with source direction.
These studies focus entirely on binaural audio which has been synthesised using convolution with Head Related Impulse Responses (HRIRs). Novel methods are presented which use a combination of feature extraction with logistic classification to develop a proficient algorithm to localise sound sources in terms of both horizontal and vertical position. While the frequency dependent cues which allow us to detect sound source elevation are difficult to model using traditional methods, the results of these studies show they can be modelled accurately using deep learning and gammatone cepstral coefficients.
To improve the performance of this algorithm in the mismatched Head-Related Transfer Function (HRTF) condition, a clustering algorithm is introduced and used to cluster similar HRTF sets together providing a more robust performance. This clustering approach is evaluated in both anechoic and adverse environments and is used to identify anthropomorphic features which are important for determining our sense of vertical localization.
Sponsor
Grant Number
Science Foundation Ireland (SFI)
Description:
APPROVED
Author: O'Dwyer, Hugh
Other Titles:
Novel Methods for the Analysis and Assessment of Spatial AudioAdvisor:
Boland, FrankPublisher:
Trinity College Dublin. School of Engineering. Discipline of Electronic & Elect. EngineeringType of material:
ThesisCollections
Availability:
Full text availableMetadata
Show full item recordLicences: