Confusion modelling for automated lip-reading using weighted finite-state transducers.

摘要: Automated lip-reading involves recognising speech from only the visual signal. The accuracy of current state-ofthe-art systems is significantly lower than that obtained by acoustic recognisers. These poor results are most likely due to lack information about production available in signal: for example, it impossible discriminate voiced and unvoiced sounds, or many places articulation, signals. Our approach this problem regard signal as having been produced a speaker who has reduced phonemic repertoire attempt compensate this. In respect, similar dysarthric speech, which control over their articulators, leading them speak with distorted set phonemes. previous work, we found use weighted finite-state transducers improved recognition performance on considerably. paper, report applying technique lip-reading. works, but our initial not good those using conventional approach, discuss why might be so what prospects future investigation are.