Automatic Identification of Visemes

We are investigating methods for determining automatically what are the visemes of a language. Most previous approaches to answering this question have used some form of subjective assessment, whereby the visual confusion of phonemes (usually consonants) in various (VCV) contexts are determined by presenting to viewers silent visual stimuli, then asking the viewers to identify the consonant. The phonemes that are (visually) confused often are grouped together and said to form a viseme (visual phoneme).

There has been much work investigating this and many phoneme-to-viseme mappings have been proposed. In this project we are applying computer vision and pattern recognition techniques to determine such a mapping without subjective assessment, and we are keen compare the results with viseme groups that were derived using subjective evaluation.

Contact: Sarah Hilder, Barry-John Theobald and Richard Harvey

Last Updated on March 2010

A range of common lip shapes