Hardware-accelerated graph neural networks: an alternative approach for neuromorphic event-based audio classification and keyword spotting on SoC FPGA

A research team led by Tomasz Kryjak presents a hardware-accelerated graph neural network (GNN) for neuromorphic audio processing. Their system, implemented on a System-on-Chip FPGA, uses an artificial cochlea to convert audio into sparse event data. It achieves 92.3% accuracy on the SHD dataset (within 2.4% of SOTA) with 10x fewer parameters, and demonstrates keyword spotting with 95% word-end detection accuracy at just 10.53µs latency and 1.18W power.