Noise-Robust Speech-Based Severity Assessment for Emergency Calls

This study aims to automatically classify emergency calls into serious (life-threatening) and minor (non- life-threatening) cases using acoustic features and machine learning models, thereby contributing to automated triage support in emergency response systems. Two enhancement strategies—noise reduction and data augmentation—are investigated to improve robustness in real-world call environments. Accurate triage during emergency calls is critical for optimizing resource allocation and ensuring timely medical response. Building on our previous exploratory analysis of acoustic features, this study advances toward practical deployment by addressing two key challenges: noisy real-world conditions and limited training data. To mitigate background noise and enhance feature stability, Wiener filtering was integrated into the preprocessing pipeline. Data scarcity was addressed through augmentation strategies, including moderate pitch shifting (±2 semitones) as well as comprehensive augmentation with pitch, volume, and noise perturbations. Acoustic features—including fundamental frequency statistics, Mel-frequency cepstral coefficients, and spectral descriptors— were extracted from call recordings provided by the Tokyo Fire Department. Three classifiers (Logistic Regression, Support Vector Machine, and Random Forest) were trained and evaluated using stratified cross-validation. Performance was primarily assessed by area under the ROC curve (AUC) and recall, given the critical importance of minimizing false negatives in emergency triage. Results showed that noise reduction improved robustness, while full augmentation yielded the greatest gains in predictive accuracy, with Random Forest achieving an AUC of 0.93. These findings demonstrate the feasibility of acoustic-based severity classification in emergency calls and highlight the potential of recall-oriented decision- support systems for emergency dispatchers. Future work will focus on real-time implementation and integration into dispatch operations.