Use this URL to cite or link to this record in EThOS:
Title: Models for learning reverberant environments
Author: Papayiannis, Constantinos
ISNI:       0000 0004 7659 0692
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Reverberation is present in all real life enclosures. From our workplaces to our homes and even in places designed as auditoria, such as concert halls and theatres. We have learned to understand speech in the presence of reverberation and also to use it for aesthetics in music. This thesis investigates novel ways enabling machines to learn the properties of reverberant acoustic environments. Training machines to classify rooms based on the effect of reverberation requires the use of data recorded in the room. The typical data for such measurements is the Acoustic Impulse Response (AIR) between the speaker and the receiver as a Finite Impulse Response (FIR) filter. Its representation however is high-dimensional and the measurements are small in number, which limits the design and performance of deep learning algorithms. Understanding properties of the rooms relies on the analysis of reflections that compose the AIRs and the decay and absorption of the sound energy in the room. This thesis proposes novel methods for representing the early reflections, which are strong and sparse in nature and depend on the position of the source and the receiver. The resulting representation significantly reduces the coefficients needed to represent the AIR and can be combined with a stochastic model from the literature to also represent the late reflections. The use of Finite Impulse Response (FIR) for the task of classifying rooms is investigated, which provides novel results in this field. The aforementioned issues related to AIRs are highlighted through the analysis. This leads to the proposal of a data augmentation method for the training of the classifiers based on Generative Adversarial Networks (GANs), which uses existing data to create artificial AIRs, as if they were measured in real rooms. The networks learn properties of the room in the space defined by the parameters of the low-dimensional representation that is proposed in this thesis.
Supervisor: Naylor, Patrick A. ; Evers, Christine Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral