Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.626574
Title: Development and evaluation of objective measures to predict the speech intelligibility of unilateral and bilateral cochlear implant users in noise and reverberation
Author: Cosentino, S.
ISNI:       0000 0004 5362 3957
Awarding Body: University College London (University of London)
Current Institution: University College London (University of London)
Date of Award: 2014
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Please try the link below.
Access through Institution:
Abstract:
Algorithms to predict speech intelligibility are useful tools for reliable, low-cost, and repeatable assessment of speech processing technologies, thus replacing costly and time-consuming subjective tests. Of particular practical interest is their application for hearing-impaired listeners, such as cochlear implant (CI) users, where speech enhancement is a possible way to reduce the gap to the higher scores achieved by normal hearing listeners. An additional feature that such measures may possess is the non-intrusiveness, which refers to the capacity of providing a speech intelligibility prediction without the need of a priori information, such as clean reference signals. This thesis includes behavioural and numerical experiments designed to investigate how speech intelligibility in unilateral and bilateral CI users can be predicted via objective measures. It is hypothesised that accurate predictions of speech intelligibility scores can be obtained by incorporation of CI-specific features into the proposed objective measures. In Chap. II, several intrusive measures originally developed for normal hearing listeners are employed to predict the speech intelligibility of unilateral CI users. A non-intrusive measure is also proposed and evaluated against other non-intrusive and intrusive measures, and shown to achieve good performance in noisy and reverberant environments. This measure relies on temporal envelopes and uses filterbanks that were derived for CI users of the Nucleus Freedom. In Chap. III, a non-intrusive measure of the bilateral advantage to speech intelligibility in CI is proposed and evaluated against simulated and publicly available data. As a first step, a non-intrusive model is developed that can predict the binaural advantage to speech intelligibility in normal hearing listeners. Behavioural data was collected in order to validate the accuracy of the model’s predictions. Both the normal hearing and the CI versions of the model measured the contribution of two psychoacoustic phenomena: the better ear effect, and the binaural unmasking. For the CI version of the model, the computation of these two contributions were modified to incorporate known limitations of the electric hearing, such as reduced time and frequency resolution, and the poor sensitivity to interaural time differences. With these modifications, also the CI version of the proposed model yielded accurate predictions of speech intelligibility scores obtained from simulated and subjective data. In Chap. IV, one of the unilateral intrusive measures described in Chap. II is used to optimise the parameters of a filterbank with respect to speech intelligibility in a simulated CI hearing environment. Across the unilateral and bilateral/binaural investigations, it can be concluded that temporal envelope modulations, amongst other factors, play a major role in predicting speech intelligibility for CI users. This is also the case for predictions for normal hearing subjects, and a few models already exist that use alterations in the temporal envelopes domain to predict the effect on speech intelligibility. However, this study shows that not all information employed by current models is necessary for such predictions. Non-intrusive models, such as those described in Chap. I and II, can be as accurate as models that require complete information of target and interferers signals. This means that that these measures could be applied to speech enhancement, providing a quick and inexpensive way to improve current settings in modern CI.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.626574  DOI: Not available
Share: