J3romee/CLEAR

# 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)

原始信息汇总

CLEAR2 数据集概述

数据集来源

• 该数据集在2021年提交至IEEE Transactions on Pattern Analysis and Machine Intelligence的文章"NAAQA: A Neural Architecture for Acoustic Question answering"中被提出。

数据集结构

• scenes/: 每个集合（训练/验证/测试）包含一个JSON文件，指定场景中每个声音的顺序和时间。
• questions/: 每个集合（训练/验证/测试）包含一个JSON文件，指定每个场景的问题和答案，以及问题的功能程序。
• audio/: 包含声学场景的录音（FLAC格式），分为训练集、验证集和测试集。
• attributes.json: 列出所有可能的答案，按问题类别划分。

数据集生成代码

• 数据集生成代码可在GitHub上获取：CLEAR-AQA-Dataset-Generator