LISTEN: Lightweight Transformer for real-time industrial sound monitoring on edge devices

Deep learning for industrial sound analysis has been held back by two opposing problems: general-purpose sound models are too computationally heavy for on-site use, while task-specific models require large annotated datasets that are infeasible to collect for each new machine or process. Researchers from Korea and the US have solved this by creating LISTEN, a lightweight foundation model built specifically for industrial acoustics. They use knowledge distillation from a large teacher model called IMPACT, compressing it into a Student model optimized for edge devices.

LISTEN's architecture freezes the backbone and trains only a shallow head on minimal target-process data, avoiding full fine-tuning. In tests across diverse manufacturing environments, it matches IMPACT's performance while running on low-cost IIoT-connected hardware. The team validated end-to-end real-time monitoring on a live CNC machine, demonstrating how the system can detect faults like tool wear or chatter from sound alone. This opens the door to affordable, scalable predictive maintenance on factory floors without cloud dependency.