- Position identifier: ESR10
- Host partner: UoC
Objectives
Much research effort has been invested in determining the source of the intelligibility gain of clear speech over conversational speech. Finding artefact-free modification algorithms to convert from conversational to clear speech remains a challenging problem. This project investigated parametric ways to describe, at a range of time resolutions (from glottal cycle to higher segmental levels), the role of temporal envelope and fine structure modulations in clear and conversational speech. A parametric representation facilitated then the design of mapping mechanisms, for instance using modern machine learning approaches, to associate parameters between the two styles. Applying the learned mapping functions, and using high-quality speech synthesis approaches (i.e., adaptive sinusoidal models) artefact-free modifications were obtained to transform conversational speech to clear speech effectively.
The ESR completed a project under the supervision of Professor Yannis Stylianou and Dr. George Kafentzis, with additional supervision from Prof. Valerie Hazan (UCL, London, UK), Dr. Olivier Rosec (Voxygen, France) and Dr. Chariton Papadakis (Chania General Hospital, Crete Greece).