Federated Data Analytics Framework for Distributed Sound Source Localization Using Wireless Microphone Networks and TinyViT Learning

Abstract

Wireless microphone node arrays offer a flexible sensing layer for acoustic monitoring, emergency perception, conference-room interaction, and mobile robot audition, but their localization accuracy is often affected by non-identical node placement, packet-level instability, and acoustic-domain non-IID data. This paper proposes an improved FedProx-TinyViT model for distributed sound source localization in wireless microphone node arrays. Multichannel short-time spectra, phase-difference cues, and node confidence descriptors are transformed into compact acoustic tokens, and a TinyViT backbone with local-window attention and cross-node fusion is trained through FedProx under heterogeneous reverberation and signal-to-noise conditions. A simulated and replayed wireless array dataset with 8 node layouts, 12 source azimuths, 5 reverberation settings, and 6 noise conditions was used for evaluation. Compared with centralized CNN, local TinyViT, FedAvg-TinyViT, and FedNova-TinyViT baselines, the proposed model achieved a mean angular error of 3.84 degrees, a within-5-degree localization rate of 91.6%, a 38.7% reduction in uplink payload, and 18.4 ms edge inference latency on an ARM-class node. Under 20% packet loss, the method obtained a mean error of 5.21 degrees, whereas FedAvg-TinyViT reached 7.46 degrees. These results indicate that proximal federated optimization and lightweight attention can jointly improve accuracy, communication efficiency, and deployment tolerance for distributed acoustic localization.