research group:

email:

course:

start date:

publications:

Natural Language Processing Group

claireb@comp.leeds.ac.uk

PhD (part time)

February 2006

forthcoming for 2007

Within a phrase or sentence, or as a paragraph unfolds, certain words are made prominent by the speaker when that text is read aloud. How and why do we do this? What makes us pick one word rather than another? Is it all down to speaker choice and if not, what are the constraints?

People reading text silently to themselves can intuitively predict which words might be made prominent if that text were read aloud and there would be a fair amount of consensus on this. But can such predictions be automated so that output sentences from a speech synthesizer sound fairly natural to the human ear?

Large corpora exist of speech recordings and transcriptions of same from many different speakers and this information has been annotated at various different levels by experts. Layers of annotation might include syntactic, semantic or discourse-related information - and of course, prosodic information.

The success of automated predictions mentioned above will depend on which layers are used and how they are combined to model and test the unwritten rules people adhere to when they use their voice in a meaningful way to aid communication.