Search for: All records

Generating realistic audio for human actions is critical for applications such as film sound effects and virtual reality games. Existing methods assume complete correspondence between video and audio during training, but in real-world settings, many sounds occur off-screen or weakly correspond to visuals, leading to uncontrolled ambient sounds or hallucinations at test time. This paper introduces AV-LDM, a novel ambient-aware audio generation model that disentangles foreground action sounds from ambient background noise in in-the-wild training videos. The approach leverages a retrieval-augmented generation framework to synthesize audio that aligns both semantically and temporally with the visual input. Trained and evaluated on Ego4D and EPIC-KITCHENS datasets, along with the newly introduced Ego4D-Sounds dataset (1.2M curated clips with action-audio correspondence), the model outperforms prior methods, enables controllable ambient sound generation, and shows promise for generalization to synthetic video game clips. This work is the first to emphasize faithful video-to-audio generation focused on observed visual content despite noisy, uncurated training data. 

We propose a new unsupervised model for mapping a variable-duration speech segment to a fixed-dimensional representation. The resulting acoustic word embeddings can form the basis of search, discovery, and indexing systems for low- and zero-resource languages. Our model, which we refer to as a maximal sampling correspondence variational autoencoder (MCVAE), is a recurrent neural network (RNN) trained with a novel self-supervised correspondence loss that encourages consistency between embeddings of different instances of the same word. Our training scheme improves on previous correspondence training approaches through the use and comparison of multiple samples from the approximate posterior distribution. In the zero-resource setting, the MCVAE can be trained in an unsupervised way, without any ground-truth word pairs, by using the word-like segments discovered via an unsupervised term discovery system. In both this setting and a semi-supervised low-resource setting (with a limited set of ground-truth word pairs), the MCVAE outperforms previous state-of-the-art models, such as Siamese-, CAE- and VAE-based RNNs.