Answers are grounded in this paper and cite their sources.
The scarcity of datasets designed for emergency sound classification poses a challenge to the development of assistive technologies for people with hearing impairments. This work presents the construction and evaluation of UFRB-ESD, a public dataset containing 2,849 one-second clips distributed across nine classes: horn, dog, collision, explosion, brake, scream, siren, gunshot, and others. Its composition includes a majority class of everyday non-emergency sounds, representing situations in which alerts must be distinguished from other environmental sounds. To characterize the dataset, an MLP, a 2D CNN, a ResNet-18, and a ResNet-50 were evaluated using stratified group ten-fold cross validation, ensuring that clips from the same source recording remained in a single fold. The 2D CNN achieved the highest mean macro AUC (94.90%) and weighted F1-score (76.30%), whereas the MLP achieved the highest mean accuracy (77.32%). Although numerical differences were observed among the models, simultaneous intervals obtained by grouped resampling did not indicate consistent superiority of any architecture. The results establish reproducible references, reveal the classification challenges present in the dataset, and reinforce the potential of UFRB-ESD as a resource for research on multiclass classification and assistive alert systems.
No takes yet. Share an insight, caveat, or question.
Gabriel Sampaio de Oliveira (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: