# CLEAR2 dataset

This dataset was presented in the article "NAAQA: A Neural Architecture for Acoustic Question answering" submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence in 2021.
https://arxiv.org/abs/2106.06147

The code to generate this dataset is available at : https://github.com/J3rome/CLEAR-AQA-Dataset-Generator

## Structure
- scenes/ : 1 json file per set (Train/val/test)
    - Specify the order and the timings of each sounds in a scene
- questions/ : 1 json files per set (Train/val/test).
    - Specify the questions and answers for each scenes.
    - The functional program of the question is also provided
- audio/ : Acoustic scenes recordings (FLAC)
    - train/
    - val/
    - test/
- attributes.json : List all possible answers (Split by question categories)