Perception of alarms in noise: development of methods for the evaluation of the perceived audibility

In most public and work areas, auditory alarms are used to alert people to dangerous situations that may require immediate action. The audibility of these alarms is essential to ensure the safety of individuals in the face of accident risk. In the interests of occupational safety and health, the protection of workers requires that alarm levels be set so as to guarantee their correct perception while limiting the risk of exposure to excessive levels. This issue is present in many noisy work environments, such as transport (land, rail and air), logistics, waste collection and treatment, and construction. This thesis focuses on the development and application of methods for assessing the audibility of alarms in noise. It aims to provide a better understanding of the notion of audibility in noise, by considering alarm and ambient noise levels commonly encountered in work environments. It also aims to provide methods for adjusting alarm levels, ensuring that they are audible enough to be perceived without effort by normal-hearing listeners, while not being excessively loud. We explore two complementary approaches. The first is based on the evaluation of detectability and audibility at supraliminal levels of alarms in noise using traditional psychoacoustical methods. After improving an existing model for predicting alarm detectability in noise, we develop an experimental method for subjective evaluation of alarm detectability and audibility. The second approach leverages machine learning methods to predict the audibility of alarms in noise at supraliminal levels. We collect and annotate a dataset through listening tests consisting of a subjective evaluation of the audibility of alarms presented in recordings of ambient noise. This dataset is used to train a convolutional neural network as part of a system for predicting alarm audibility. The results of the psychoacoustic approach reveal the main factors and mechanisms involved in the perception of alarm audibility at supraliminal levels. Based on these experimental results, we propose an audibility criterion that imposes alarm levels depending on the ambient noise level. This criterion ensures adequate audibility, while reducing the risk of overexposure that is potentially induced by the application of other existing criteria. Regarding the machine learning approach, our system is able to predict the audibility of alarms as assessed by normal-hearing human listeners, with a level of performance comparable to that of an average listener. Through a series of experiments, we study the impact of the dataset annotation strategy on the system performance. From this, we derive a set of methodological recommendations applicable to the collection of perceptually annotated datasets. Finally, we compare the system predictions with human assessments of audibility. We identify further analytical perspectives for a more exhaustive characterization of the system, which is necessary for any practical application.