Sound Event Classification with TinyML on ESP32-S3 for Risk Alerting of Hearing-Impaired People in Public Spaces

This paper presents a TinyML embedded system for environmental sound classification targeting hearing-impaired individuals in public spaces. A lightweight 1D Convolutional Neural Network (1D CNN) trained through the Edge Impulse platform is deployed on an ESP32-S3 microcontroller directly connected to a MAX9814 electret microphone, currently classifying two acoustic classes - vehicle horn and baby cry - entirely on-device without internet connectivity. Background silence is handled through a firmware-level amplitude threshold filter, bypassing unnecessary neural inference during acoustically inactive intervals and reducing processor workload. Experimental evaluation on physical hardware achieved 97.3% validation accuracy with 7 ms inference latency, confirming the feasibility of low-latency embedded audio classification for assistive safety applications. A BLE 5.0 haptic wearable is proposed as the next development stage.