Introduction: This study investigates real-world voice communication between pilots and air traffic controllers (ATCs) during emergencies to identify acoustic indicators of situational stress.
Method: We analyzed 75 recordings from LiveATC.net, manually annotated for speaker (pilot vs ATC) and flight phase (routine vs emergency). A total of 27 acoustic features were extracted using Parselmouth. Feature selection involved ANOVAs with FDR correction and collinearity checks. Speaker-specific Linear Discriminant Analyses (LDA) and Leave-One-Event-Out cross-validations were applied.
Results: Twenty-two features for ATCs and twenty for pilots significantly differed between the two flight phases. LDA models achieved cross-validated accuracy of 73.6% (ATCs) and 78.7% (pilots). Shared markers of stress included increase in F0 mean, F0 min, intensity max, utterance duration, jitter and shimmer, as well as decrease in HNR and CPP, indicating increased vocal effort and reduced voice quality during emergencies. Differences by role emerged: ATCs showed a substantial increase in RMS energy, number of pauses, and spectral tilt, while pilots exhibited higher articulation rate under stress.
Conclusions: This study highlights robust, role-specific acoustic markers of acute stress in real-world aviation emergencies. These findings may inform real-time voice-based stress monitoring tools to enhance aviation safety.