Toward Interpretable and Real-time Speech Biomarkers of Stress Using Adaptive Sub-band Attention

Speech carries physiological indicators of stress through changes in respiration, vocal-fold tension, and articulatory stability. Conventional representations such as MFCCs and i-vectors smooth octave-level spectral variations and remain sensitive to speaker and noise variability. This study introduces an Adaptive Spectral-Vector (ASV) framework combining spectral contrast, Karhunen-Loève transformation, and i-vector modeling with a novel Adaptive Sub-Band Attention (ASBA) mechanism that adaptively enhances stress-relevant frequency regions while suppressing noise. Evaluations on SUSAS, Emo-DB, and RAVDESS using speaker-independent, cross-corpus, noise-added, and low-resource settings show strong performance: 91.7% intra-corpus accuracy, ≥82% cross-corpus accuracy, and 82.1% at 5 dB SNR. ASBA attention weights correlate significantly (p < 0.01) with physiological acoustic markers, supporting interpretability. With ~0.61M parameters and <10 ms latency on Raspberry Pi 4B and Jetson Nano, ASV enables real-time, on-device deployment. Although based on acted emotions, the framework establishes a reproducible and efficient foundation for future physiologically validated stress monitoring.