Audio-visual speech recognition by speechreading

Speechreading increases intelligibility in human speech perception. This suggests that conventional acoustic-based speech processing can benefit from the addition of visual information. This paper exploits speechreading for joint audio-visual speech recognition. We first present a color-based feature extraction algorithm that is able to extract salient visual speech features reliably from a frontal view of the talker in a video sequence. Then, a new fusion strategy using a coupled hidden Markov model (CHMM) is proposed to incorporate visual modality into the acoustic subsystem. By maintaining temporal coupling across the two modalities at the feature level and allowing asynchrony in the state at the same time, a CHMM provides a better model for capturing temporal correlations between the two streams of information. The experimental results demonstrate that the combined audio-visual system outperforms the acoustic-only recognizer over a wide range of noise levels.