Phoneme-to-viseme Mapping for Visual Speech Recognition

Phonemes are the standard modelling unit in HMM-based continuous speech recognition systems. Visemes are the equivalent unit in the visual domain, but there is less agreement on precisely what visemes are, or how many to model on the visual side in audio-visual speech recognition systems. This paper compares the use of 5 viseme maps in a continuous speech recognition task. The focus of the study is visual-only recognition to examine the choice of viseme map. All the maps are based on the phoneme-to-viseme approach, created either using a linguistic method or a data driven method. DCT, PCA and optical flow are used as the visual features. The best visual-only recognition on the VidTIMIT database is achieved using a linguistically motivated viseme set. These initial experiments demonstrate that the choice of visual unit requires more careful attention in audio-visual speech recognition system development.