Fredrik Bajers Vej 5
P.O. Box 159 DK-9100 Aalborg
Phone: +45 9940 9940
22.11.2017 kl. 13.00 - 22.11.2017 kl. 16.00
Speech Intelligibility Prediction for Hearing Aid Systems
Hearing loss is a serious medical condition, which, primarily due to the aging of the population, affects increasingly many individuals. The typical approach to treatment involves the daily use of hearing aids, which amplify and enhance incoming sounds. Immense research efforts continuously strive to improve the performance of such hearing aids. While the performance of hearing aids can be measured in many ways, speech intelligibility is possibly the most important one. Designing hearing aids, which are good in this respect, is made difficult by the fact that speech intelligibility is cumbersome to measure. Typically, speech intelligibility has to be measured using time consuming and expensive listening experiments involving many listeners.
In this thesis we study an emerging alternative to listening experiments: Speech Intelligibility Prediction (SIP) algorithms. Such algorithms estimate the outcome of listening experiments, using audio recordings from the studied acoustical conditions. We propose and investigate several SIP algorithms extending the popular Short-Time Objective Intelligibility (STOI) measure. While these algorithms are applicable for a wide range of use-cases, we have investigated particular applications in the development of hearing aid systems.
Specifically, we propose 1) multiple binaural SIP algorithms, which can account for the advantage obtained in binaural listening environments, and 2) two non-intrusive SIP algorithms, which can predict intelligibility without requiring a clean speech reference signal. The proposed measures are shown to yield predictions which are accurate in comparison with competing algorithms, across a broad range of acoustical conditions including different configurations of noise types, non-linear processing, reverberation, and spatial source positions. In particular, using one of the proposed algorithms, it was possible to predict the result of a listening experiment, which had previously been carried out as part of product development within Oticon A/S. Thus, in principle, a subset of such experiments could be replaced by cheaper and less time consuming predictions. This underlines the potential of SIP as a highly valuable tool within the development of hearing aids.
Associate Professor Flemming Christensen, Aalborg University, Denmark (Chairman)
Dr. rer, nat.Thomas Brand, Oldenburg University, Germany
Reader Mike Brookes, Electrical and Electronic Engineering, Imperial College London
Professor Zheng-Hua Tan, Aalborg University, Denmark
Professor Jesper Jensen, Aalborg University, Denmark
Jan Mark de Haan, Oticon A/S, Denmark
Professor Søren Holdt Jensen, Aalborg University, Denmark
After the defence there will be a reception at Fredrik Bajers Vej 7, B4-211
Free of charge
Signal and Information Processing, Department of Electronic Systems
Aalborg University, Fredrik Bajers Vej 7 A4-106