Class-specific classifiers in audio-visual speech recognition

In this paper, class-specific classifiers for audio, visual and audiovisual speech recognition systems are developed and compared with traditional Bayes classifiers. We use state-of-the-art feature extraction methods and develop traditional and class-specific classifiers for speech recognition, showing the benefits of a class-specific method on each modality for speaker dependent and independent set-ups. Experiments with a reference audio-visual database show a general increase in the systems performance by the introduction of class-specific techniques on both visual and audio-visual modalities.