This thesis examines the application of a particular neuromorphic computational model, the Boltzmann Machine, to the evaluation of laryngeal behaviour using parameters derived from the acoustic analysis of irregularities in the periodic structures of speech signals. Over the last twenty five years, researchers in various fields such as speech science, laryngology, speech pathology and phonetics have demonstrated a growing interest in the acoustic characterisation of healthy and pathological voices. This research activity has been in response to the need for non-invasive and quantitative techniques for the assessment of layrngeal function. Over the past five years neuromorphic computation has undergone a dramatic transformation with the development of powerful learning algorithms and the promise of highly parallel implementations taking advantage of developments in high density integrated circuit technology. These neuromorphic systems are machines that behave in brain-like ways and compute by absorbing experience. The Boltzmann Machine learning algorithm provides a formally guaranteed procedure for performing gradient descent in a global error measure. This thesis presents, for the first time, results which demonstrate the potential of the Boltzmann Machine approach to the detection of laryngeal pathologies. A simulation environment for Boltzmann Machines was successfully developed which provided acceptable speeds of operation for the sizes of network investigated, and the quantity of training data used, provided approximations to the <i>theoretical</i> Boltzmann Machine were made. Chapter 4 presents details of this implementation. Experiments using various topologies of Boltzmann Machine made use of ten intonation and perturbation parameters, derived from the analysis of waveform perturbations of fundamental frequency and amplitude evidenced in samples of connected speech from groups of healthy and pathological male speakers. A series of experiments are presented in Chapter 6 which evaluate the performance of various Boltzmann Machine topologies and data representation formats to the differentiation between groups of healthy speakers and speakers with known pathological conditions of the larynx. From these experiments it was concluded that the intonation and perturbation parameters could be processed using a Boltzmann Machine to provide useful differentiation between groups of healthy speakers and speakers with known pathological conditions. Chapter 7 presents a series of experiments using various topologies of Boltzmann Machine to differentiate between broad classes of pathologies using ten intonation and perturbation parameters. The experiments showed that it was possible for various pathology groups to be differentiated in a training group. The successful development of a neuromorphic system for determination of laryngeal function associated with healthy and pathological phonation has a number of potential applications, including screening, differential diagnosis and tracking changes in the conditions of laryngeal pathology.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:663032 |
| Date | January 1990 |
| Creators | Trehern, Jocelyn Frank |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/14573 |
Page generated in 0.0024 seconds