Multimodal feature extraction and fusion for audio-visual speech recognition

Multimodal signal processing analyzes a physical phenomenon through several types of measures, or modalities. This leads to the extraction of higher-quality and more reliable information than that obtained from single-modality signals. The advantage is two-fold. First, as the modalities are usually complementary, the end-result of multimodal processing is more informative than for each of the modalities individually, which represents the first advantage. This is true in all application domains: human-machine interaction, multimodal identification or multimodal image processing. The second advantage is that, as modalities are not always reliable, it is possible, when one modality becomes corrupted, to extract the missing information from the other one. There are two essential challenges in multimodal signal processing. First, the features used from each modality need to be as relevant and as few as possible. The fact that multimodal systems have to process more than just one modality means that they can run into errors caused by the curse of dimensionality much more easily than mono-modal ones. The curse of dimensionality is a term used essentially to say that the number of equally-distributed samples required to cover a region of space grows exponentially with the dimensionality of the space. This has important implications in the classification domain, since accurate models can only be obtained if an adequate number of samples is available, and obviously this required number of samples grows with the dimensionality of the features. Dimensionality reduction is thus a necessary step in any application dealing with complex signals, and this is achieved through selection, transforms or the combination of the two. The second essential challenge is multimodal integration. Since the signals involved do not necessarily have the same data rate, range or even dimensionality, combining information coming from such different sources is not straightforward. This can be done at different levels, starting from the basic signal level by combining the signals themselves, if they are compatible, up to the highest decision level, where only the individual decisions taken based on the signals are combined. Ideally, the fusion method should allow temporal variations in the relative importance of the two streams, to account for possible changes in their quality. However, this can only be done with methods operating at a high decision level. The aim of this thesis is to offer solutions to both these challenges, in the context of audio-visual speech recognition and speaker localization. Both these applications are from the field of human-machine interaction. Audio-visual speech recognition aims to improve the accuracy of speech recognizers by augmenting the audio with information extracted from the video, more particularly, the movement of the speaker's lips. This works well especially when the audio is corrupted, leading in this case to significant gains in accuracy. Speaker localization means detecting who is the active speaker in a audio-video sequence containing several persons, something that is useful for videoconferencing and the automated annotation of meetings. These two applications are the context in which we present our solutions to both feature selection and multimodal integration. First, we show how informative features can be extracted from the visual modality, using an information-theoretic framework which gives us a quantitative measure of the relevance of individual features. We also prove that reducing redundancy between these features is important for avoiding the curse of dimensionality and improving recognition results. The methods that we present are novel in the field of audio-visual speech recognition and we found that their use leads to significant improvements compared to the state of the art. Second, we present a method of multimodal fusion at the level of intermediate decisions using a weight for each of the streams. The weights are adaptive, changing according to the estimated reliability of each stream. This makes the system tolerant to changes in the quality of either stream, and even to the temporary interruption of one of the streams. The reliability estimate is based on the entropy of the posterior probability distributions of each stream at the intermediate decision level. Our results are superior to those obtained with a state of the art method based on maximizing the same posteriors. Moreover, we analyze the effect of a constraint typically imposed on stream weights in the literature, the constraint that they should sum to one. Our results show that removing this constraint can lead to improvements in recognition accuracy. Finally, we develop a method for audio-visual speaker localization, based on the correlation between audio energy and the movement of the speaker's lips. Our method is based on a joint probability model of the audio and video which is used to build a likelihood map showing the likely positions of the speaker's mouth. We show that our novel method performs better than a similar method from the literature. In conclusion, we analyze two different challenges of multimodal signal processing for two audio-visual problems, and offer innovative approaches for solving them.