UFRB-ESD (UFRB Emergency Sound Dataset)

This is the UFRB-ESD (UFRB Emergency Sound Dataset), an open-access audio resource developed as part of an undergraduate thesis (Final Project) in Computer Engineering at the Universidade Federal do Recôncavo da Bahia (UFRB), Brazil. The dataset was created to address the gap in datasets focused on short-duration emergency sounds, aiming at the development of assistive technologies for the hearing impaired (e.g., low-latency wearable devices). Dataset Content: 2,849 audio clips in .wav format (16kHz, mono). Standard Duration: 1 second per clip. 9 Classes: buzina (horn), tiro (gunshot), explosão (explosion), sirene (siren), grito (scream), colisão (crash/collision), freio (brakes), cachorro (dog bark), and outros (other/non-emergency sounds). All clips were collected from public sources and underwent a manual curation and validation process. The dataset was validated with four deep learning architectures (MLP, 2D CNN, ResNet-18, ResNet-50), achieving a maximum accuracy of 86.67% (ResNet-18), which confirms its viability and challenging nature for classification tasks.