A Federated Approach for Adaptive Urban Sound Classification on TinyML Edge Devices

Cities exhibit sound patterns that vary across locations and time, while transmitting raw audio introduces communication and privacy concerns. We present a federated TinyML architecture for real-time urban sound classification on microcontroller-class edge devices. A compact audio embedding network is deployed as a frozen feature extractor, while a lightweight classifier head is trained on-device and shared via MQTT, enabling communication-efficient collaborative learning. The system is evaluated on ESP32 (Espressif Systems, Shanghai, China) hardware under cross-dataset transfer from UrbanSound8K to SONYC. Domain shift reduces baseline accuracy from 90.39% to 78.27%, while local adaptation and federated aggregation improve accuracy to approximately 85%, recovering most of the performance loss. Repeated aggregation further improves macro-F1 and class balance across heterogeneous data. Embedded measurements confirm real-time inference (~250 ms per window) with negligible overhead, while each update exchanges only a compact classifier head (~1.2 kB). These results demonstrate that adaptive classification can be achieved on resource-constrained nodes in distributed smart-city networks.