The Sound of Pixels

TL;DR

This paper presents PixelPlayer, a system that learns to associate sounds with specific image regions and separates audio sources using unlabeled videos, leveraging visual-audio synchronization without manual labels.

Contribution

It introduces the Mix-and-Separate framework that jointly learns sound localization and separation from unlabeled videos, advancing audio-visual source separation techniques.

Findings

Outperforms baselines on the MUSIC dataset

Learns to ground sounds in visual regions

Enables independent volume adjustment of sound sources

Abstract

We introduce PixelPlayer, a system that, by leveraging large amounts of unlabeled videos, learns to locate image regions which produce sounds and separate the input sounds into a set of components that represents the sound from each pixel. Our approach capitalizes on the natural synchronization of the visual and audio modalities to learn models that jointly parse sounds and images, without requiring additional manual supervision. Experimental results on a newly collected MUSIC dataset show that our proposed Mix-and-Separate framework outperforms several baselines on source separation. Qualitative results suggest our model learns to ground sounds in vision, enabling applications such as independently adjusting the volume of sound sources.