The problem of feature selection has been thoroughly analyzed in the context of pattern classification, with the purpose of avoiding the curse of dimensionality. However, in the context of multimodal signal processing, this problem has been studied less. Our approach to feature extraction is based on information theory, with an application on multimodal classification, in particular audio-visual speech recognition. Contrary to previous work in information theoretic feature selection applied to multimodal signals, our proposed methods penalize features for their redundancy, achieving more compact feature sets and better performance. We propose two greedy selection algorithms, one that penalizes a proportion of feature redundancy, while the other uses conditional mutual information as an evaluation measure, for the selection of visual features for audio-visual speech recognition. Our features perform better than linear discriminant analysis, the most usual transform for dimensionality reduction in the field, across a wide range of dimensionality values and combined with audio at different quality levels.