Occupational noise exposure is a widespread concern, impacting millions of workers. The present research focuses on the audibility of acoustic alarms to ensure worker safety while minimizing exposure to unnecessarily high alarm levels. It introduces a laboratory experiment carried on normal-hearing participants to assess the perceived audibility of acoustic alarms in various workplace noise conditions. The experiment aimed to enhance comprehension of the audibility of acoustic alarms at supra-threshold levels, sought to facilitate the formulation of improved guidelines for alarm design. The results reveal the inappropriateness of the most commonly employed alarm level setting criterion of the ISO 7731 international standard, leading to excessive alarm levels in highly noisy work environments. Based on our data, we propose a revised value for this criterion. In addition, an acoustical analysis of the sounds used in the experiment shows that alarms that are more salient are perceived as more audible, thereby providing leads for alarm design. The study also introduces an innovative technique using a convolutional neural network model to predict the audibility of alarms in noise. Moving beyond generic arbitrary criteria, this data-driven approach leverages knowledge from perceptually annotated examples sourced from our contributed dataset. Evaluation on the experimental data and further analysis of the model outputs demonstrate solid alignment of the model predictions with human perception.