Air Traffic Controller-Pilot Speech Analysis: A Bayesian Statistical Framework for Future Aviation Systems Safety

This study introduces a logistic regression-based stress level metric to (1) assess stress levels of Air Traffic Controllers (ATC) and pilots, and (2) detect off-nominal events using audio-derived features. With over 16,264 labeled ATC-pilot utterances and scalar and multi-dimensional speech attributes, it is demonstrated that a developed Stress Score metric can capture meaningful distinctions between the nominal and off-nominal utterances. The Beta distribution fit further captures the skewed nature of Stress Scores, indicating that the model follows a non-linear but quantifiable pattern. The Bayesian inference of scalar features using Markov Chain Monte Carlo sampling and posterior distributions provides further insight into the significance of specific speech attributes generated during the process of Deep Speech Pattern (DSP) analysis. DSP involves the generation of speech attributes, or data points, representing the tonal qualities of an audio segment, such as Loudness and Tempo. The study found that speech attributes, such as Loudness, Articulation, Noise Level, and Pitch Mean, have positive coefficients, indicating that their high values increase the likelihood of the off-nominal classification. On the other hand, speech attributes Spectral Centroid, Bandwidth, and Rolloff show negative coefficients, indicating that their flatter or degraded frequency profiles may be associated with higher Stress Scores. The study demonstrates the feasibility of applying an audio-based speech attribute analysis using machine learning methods, logistic regression classifiers, and Bayesian statistics, to act as a first-layer shield in aviation safety modeling scenarios. The model demonstrates that the DSP method can be used to enhance aviation safety.