Deep audio-visual speech recognition

Abstract

Decades of research in acoustic speech recognition have led to systems that we use in our everyday life. However, even the most advanced speech recognition systems fail in the presence of noise. The degraded performance can be compensated by introducing visual speech information. However, Visual Speech Recognition (VSR) in naturalistic conditions is very challenging, in part due to the lack of architectures and annotations.

This thesis contributes towards the problem of Audio-Visual Speech Recognition (AVSR) from different aspects. Firstly, we develop AVSR models for isolated words. In contrast to previous state-of-the-art methods that consists of a two-step approach, feature extraction and recognition, we present an End-to-End (E2E) approach inside a deep neural network, and this has led to a significant improvement in audio-only, visual-only and audio-visual experiments. We further replace Bi-directional Gated Recurrent Unit (BGRU) with Temporal Convolutional Networks (TCN) to greatly simplify the training procedure.

Secondly, we extend our AVSR model for continuous speech by presenting a hybrid Connectionist Temporal Classification (CTC)/Attention model, that can be trained in an end-to-end manner. We then propose the addition of prediction-based auxiliary tasks to a VSR model and highlight the importance of hyper-parameter optimisation and appropriate data augmentations.

Next, we present a self-supervised framework, Learning visual speech Representations from Audio via self-supervision (LiRA). Specifically, we train a ResNet+Conformer model to predict acoustic features from unlabelled visual speech, and find that this pre-trained model can be leveraged towards word-level and sentence-level lip-reading.

We also investigate the Lombard effect influence in an end-to-end AVSR system, which is the first work using end-to-end deep architectures and presents results on unseen speakers. We show that even if a relatively small amount of Lombard speech is added to the training set then the performance in a real scenario, where noisy Lombard speech is present, can be significantly improved.

Lastly, we propose a detection method against adversarial examples in an AVSR system, where the strong correlation between audio and visual streams is leveraged. The synchronisation confidence score is leveraged as a proxy for audio-visual correlation and based on it, we can detect adversarial attacks. We apply recent adversarial attacks on two AVSR models and the experimental results demonstrate that the proposed approach is an effective way for detecting such attacks.