AudioAlert: A Dataset for Audio-Based Danger and Distress Detection

Overview The AudioAlert dataset is a curated collection of audio recordings designed for the development and evaluation of automated danger detection systems, focused on women's safety. The dataset enables training models to distinguish between women's screams or distress sounds and normal, everyday environments. Dataset Characteristics Total Files: 389 audio samples Classes: 2 (Danger, Normal) Duration: 1 to 3 seconds per clip Sampling Rate: 16,000 Hz (16 kHz) Channels: 1 (Mono) Normalization: Amplitude-normalized Class Distribution and Provenance The dataset is categorized into two classes to support binary classification tasks: 1. Danger (177 samples) Description: Voices of women in distress, screaming or shouting for help. Sources: A combination of original live recordings from 54 consenting volunteers captured on consumer smartphones, and curated audio from social media, online sites, and cinematic sources. 2. Normal (212 samples) Description: Background sounds representing typical daily activities. This includes human conversation, traffic noise, vehicle honks (motorcycles, cars, trains), and general domestic or travel sounds. Sources: Original recordings captured using consumer smartphones in naturalistic environments by the authors. No volunteer actors were used for this class. Preprocessing Pipeline All 389 audio clips underwent a uniform preprocessing pipeline: Resampling: All audio was resampled to a uniform rate of 16 kHz. Channel Conversion: All audio was converted to mono. Amplitude Normalization: Waveforms were standardized by dividing by their maximum absolute amplitude to ensure consistent volume levels across all samples. Repository Structure The dataset is provided as an archive containing the raw audio files: README.md danger/ (177 distress audio clips) normal/ (212 daily activity audio clips) Ethical and Usage Considerations Informed consent was obtained from all volunteers who contributed distress recordings to the Danger class. The data is provided for research and development purposes in the domains of affective computing and safety-critical audio classification. License This dataset is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).