Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.759903
Title: Masked conditional neural networks for sound recognition
Author: Medhat, Fady
ISNI:       0000 0004 7431 9217
Awarding Body: University of York
Current Institution: University of York
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Sound recognition has been studied for decades to grant machines the human hearing ability. The advances in this field help in a range of applications, from industrial ones such as fault detection in machines and noise monitoring to household applications such as surveillance and hearing aids. The problem of sound recognition like any pattern recognition task involves the reliability of the extracted features and the recognition model. The problem has been approached through decades of crafted features used collaboratively with models based on neural networks or statistical models such as Gaussian Mixtures and Hidden Markov models. Neural networks are currently being considered as a method to automate the feature extraction stage together with the already incorporated role of recognition. The performance of such models is approaching handcrafted features. Current neural network based models are not primarily designed for the nature of the sound signal, which may not optimally harness distinctive properties of the signal. This thesis proposes neural network models that exploit the nature of the time-frequency representation of the sound signal. We propose the ConditionaL Neural Network (CLNN) and the Masked ConditionaL Neural Network (MCLNN). The CLNN is designed to account for the temporal dimension of a signal and behaves as the framework for the MCLNN. The MCLNN allows a filterbank-like behaviour to be embedded within the network using a specially designed binary mask. The masking subdivides the frequency range of a signal into bands and allows concurrent consideration of different feature combinations analogous to the manual handcrafting of the optimum set of features for a recognition task. The proposed models have been evaluated through an extensive set of experiments using a range of publicly available datasets of music genres and environmental sounds, where they surpass state-of-the-art Convolutional Neural Networks and several hand-crafted attempts.
Supervisor: Chesmore, David ; Robinson, John Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.759903  DOI: Not available
Share: