Field Study Evaluation of Cepstrum Coefficient Speech Analysis for Fatigue in Aviation Cabin Crew

Impaired neurobehavioral performance induced by fatigue may compromise safety in 24-hour operational environments such as aviation. As such, non-invasive, reliable, and valid methods of objectively detecting compromised performance capacity in operational settings could be valuable as a means of identifying, preventing, and mitigating fatigue-induced safety risks. One approach that has attracted attention in recent years is quantitative speech analysis, but the extent of its operational feasibility, validity of the metrics, and sensitivity to operationally-relevant factors in aviation remains unknown. To this end, the present report offers an initial proof-of-concept evaluation of a speech analysis method based on Cepstrum Coefficient modeling, using voice files from a broad sample of 195 cabin crew personnel collected during the 2009-2010 United States Civil Aerospace Medical Institute-sponsored Flight Attendant Field Study. Using a personal digital assistant device, participants recited five standardized phrases in random order before and after each workday and sleep episode throughout their respective 3-4 week study periods. Operational acceptability of the procedure was high, as indicated by high protocol compliance and, despite the inherent variability of the timing and environments in which the test sessions occurred, the 13,975 files from 2,795 valid sessions were of sufficient quality for formal analysis. Individualized “baseline” speech models were built from the files collected during test sessions coinciding with optimal neurobehavioral performance, as determined by 5-min Psychomotor Vigilance Test (PVT) reaction times(RT), then speech deviation scores relative to individual baseline models were calculated for the test sessions that preceded and concluded each “trip” of multiple consecutive work days. Regarding validity, speech scores correlated significantly with PVT RTs and Lapses (RTs > 500 msec), with a stronger relationship to Lapses, but high variability at the low range of both performance variables suggests the influence of other factors. Regarding sensitivity to operational factors, average Pre-Trip vs. Post-Trip speech scores differed significantly, although scores unexpectedly decreased from Pre to Post, an artifact attributable to the composition of the baseline session pool. Nonetheless, the pattern of speech data echoed performance data from a previous report in which scores were most affected in crew of Regional carriers, with Junior seniority, and in Domestic operations. These initial results reveal promising validity and sensitivity of Cepstrum Coefficient modeling for speech signal analysis of fatigue in dynamic operational environments. Remaining questions underscore the need to further explore the dataset to determine the precise relationship between speech production and neurobehavioral performance capacity, the parameters for constructing individualized models, and standardized quantitative speech-based definitions of fatigue.