I created the Psychoacoustic‑Stress‑Detector repository on my GitHub in February 2026. It is a Python project (3 ★) that combines deep‑learning sound classification with psychoacoustic metrics to estimate human‑perceived stress. The system runs via a Gradio UI, supports single‑file, batch, and live microphone inputs, and can generate per‑sample PDF reports with geospatial stress mapping using Folium. github profile
The repository language is Python.
The repository has 3 stars on GitHub.
The repository was created in Feb 2026 and last updated in Feb 2026.
I selected ResNet‑18 as the final classification model because it achieved the highest average accuracy of 89.53 % and best F1 of 0.9003 (Fold 5) on UrbanSound8K.
The dataset used is UrbanSound8K with 8,732 audio clips across 10 classes and 10 strict folds.
The stress score is computed as 0.5 × Roughness + 0.3 × Tonality + 0.2 × Sharpness; Roughness is weighted highest due to its strong correlation with perceived disturbance.
Uncertainty is estimated via prediction entropy and calibrated with temperature scaling to avoid overconfident outputs.