The rising popularity of neural networks, combined with the recent proliferation of online audio-visual media, has led to a revolution in the way machines encode, recognize, and generate acoustic and visual speech. Despite the ubiquity of naturally paired audio-visual data, only a limited number of works have applied recent advances in deep learning to leverage the duality between audio and video within this domain. This thesis considers the use of neural networks to learn from large unlabelled datasets of audio-visual speech to enable new practical applications. We begin by training a visual speech encoder that predicts latent features extracted from the corresponding audio on a large unlabelled audio-visual corpus. We apply the trained visual encoder to improve performance on lip reading in real-world scenarios. Following this, we extend the idea of video learning from audio by training a model to synthesize raw speech directly from raw video, without the need for text transcriptions. Remarkably, we find that this framework is capable of reconstructing intelligible audio from videos of new, previously unseen speakers. We also experiment with a separate speech reconstruction framework, which leverages recent advances in sequence modeling and spectrogram inversion to improve the realism of the generated speech. We then apply our research in video-to-speech synthesis to advance the state-of-the-art in audio-visual speech enhancement, by proposing a new vocoder-based model that performs particularly well under extremely noisy scenarios. Lastly, we aim to fully realize the potential of paired audio-visual data by proposing two novel frameworks that leverage acoustic and visual speech to train two encoders that learn from each other simultaneously. We leverage these pre-trained encoders for deepfake detection, speech recognition, and lip reading, and find that they consistently yield improvements over training from scratch.