CCStereo: Audio-Visual Contextual and Contrastive Learning for Binaural Audio Generation

TL;DR

This paper introduces CCStereo, a novel audio-visual model for binaural audio generation that uses contrastive learning and dynamic normalization to improve spatial accuracy and reduce overfitting to room environments.

Contribution

We propose a new model with an audio-visual conditional normalization layer and contrastive learning to enhance spatial sensitivity in binaural audio generation.

Findings

Achieves state-of-the-art accuracy on FAIR-Play and MUSIC-Stereo benchmarks.

Effectively utilizes test-time augmentation for improved performance.

Enhances spatial detail preservation in binaural audio synthesis.

Abstract

Binaural audio generation (BAG) aims to convert monaural audio to stereo audio using visual prompts, requiring a deep understanding of spatial and semantic information. However, current models risk overfitting to room environments and lose fine-grained spatial details. In this paper, we propose a new audio-visual binaural generation model incorporating an audio-visual conditional normalisation layer that dynamically aligns the mean and variance of the target difference audio features using visual context, along with a new contrastive learning method to enhance spatial sensitivity by mining negative samples from shuffled visual features. We also introduce a cost-efficient way to utilise test-time augmentation in video data to enhance performance. Our approach achieves state-of-the-art generation accuracy on the FAIR-Play and MUSIC-Stereo benchmarks.