In noisy occupational settings, the audibility of acoustic alarms is crucial for alerting workers to potential dangers. To ensure the audibility of warning signals, the international standard ISO 7731 requires that alarms be clearly audible and provides the level to play the auditory alarms relative to ambient noise. However, the expression clearly audible is not defined and the level specifications lead to excessive alarm levels for high noise levels, thereby exposing workers to dangerous high sound levels. In order to ensure that alarms are audible without being excessively loud, the goal of our work is to propose an automatic approach to assess their audibility at supraliminal levels, thus avoiding having to test them experimentally. Our contribution is: (1) the development of an experimental method to assess the audibility of the alarms according to ISO 7731; (2) using this method, the collection of extensive perceptual data on a minimum of 12 normal-hearing listeners (minimum 2000 training samples involving 70 alarms and 50 noises); (3) the development of a convolutional neural network that is learned and evaluated from the perceptual data, showing strong generalization; (4) a detailed analysis of both the evaluation data and the model performance.
