Automatic Contextual Audio Denoising dataset

The Automatic Contextual Audio Denoising (ACAD) dataset is used in the publication Automatic Contextual Audio Denoising. In ACAD task, target and noise distributions are defined under the learned audio context, and different audio contexts can give different definitions of target and noise. In this dataset specifically, noise refers to outlier sound events that do not belong to the current scene’s sound event distribution.

The audios are mono, 10-second long, sampled at 44100Hz. The provided audio comes in pairs of noisy and clean (target) audio signals across 6 acoustic scenes

- Kitchen

- Park

- Restaurant

- Restroom

- Street

- Subway

The acoustic scene audio from CochlScene dataset and the noise components from FSD50K are used to create the dataset. Details on the dataset construction is specified in the publication, available at https://arxiv.org/abs/2605.22262. For details in the dataset structure, please check the README.md file.