Description

The SAR-Dataset is a synthetic audio dataset designed for training and evaluating acoustic classifiers to detect human-made distress signals in post-disaster environments. This dataset supports research in drone-based search and rescue operations where acoustic detection can complement visual sensing systems.

Dataset Generation

The dataset is generated from isolated sound samples in the PostDisasterDataset [1] by combining human-made distress signals with realistic background noises typical of post-disaster environments (thunder, rain, engines, tools). Each human distress sound is overlaid with randomly selected background noise attenuated by 6 dB, resulting in mean SNR values between 10-20 dB across classes.

Dataset Details

Total samples: 15,000 ten-second audio recordings

Classes (5): crying, coughing, screaming, knocking, none (3,000 samples each)

Split: 80% training (12,000) / 10% validation (1,500) / 10% test (1,500)

Audio format: 10-second mono recordings at 32 kHz

SNR range: 10-20 dB mean

File Naming Convention

[UniqueID]_[HumanClassID]_[BackgroundClassID].wav

Human Class ID

• 0 – coughing

• 1 – crying

• 2  – knocking

• 3 – screaming

Background Class ID

• 0 – engine

• 1 – heavy equipment

• 2 – rain

• 3 – thunder

• 4 – tools

• 5 – water

• 6 – wind

[1] Tuan-Anh Pham, Seong-Hun Park, Jong-Hoon Lee, Sang-Jin Han, Jungyu Choi, and Sungbin Im. An audio dataset and hierarchical taxonomy for post-disaster scenes. In Proceedings of Acoustics 2024. Australian Acoustical Society, November 2024.