E1-ESR: Modelling intelligibility using discriminance scores derived from automatic speech recognition models

Automatic intelligibility assessment is important for accelerating the design and optimization of hearing devices by reducing reliance on time-consuming listening tests in the first design phases. Intelligibility models can be grouped into two broad classes: signal-based models use spectro-temporal measures such as the articulation index (AI), the speech transmission index (STI), or the speech intelligibility index (SII) as predictors fail to take contextual and phonetic information into account, whereas template-based models use slowly evolving cues to judge overall performance. In this project we aim to develop a novel, smoothly differentiable measure of speech intelligibility. By using a large-vocabulary continuous-speech recognition engine, our system will consider detailed phonetic and linguistic context information in the confusability assessment, and a bio-inspired feature extraction will allow us to model hearing impaired listening.

Host institution: Ruhr-Universität Bochum (RUB)
Supervisor names: Dorothea Kolossa and Torsten Dau (DTU)
Start date: 1.32013
PhD enrolment: Yes
Employment contract: Full-time early-stage research fellow, 36 months