Biomimetic-Voltage Nitride-Enhanced Memristor Arrays for Femtojoule In-Memory Biosignal Processing

Abstract Edge audio recognition systems demand extreme energy efficiency and noise resilience for deployable intelligence. However, conventional approaches typically rely on separate denoising and recognition modules, incurring significant overhead and latency that hinder deployment on severely resource-constrained edge devices. Inspired by the biological auditory system, we propose an ultralow voltage memristor-based neuromorphic system that monolithically integrates on-chip audio denoising and recognition. Our Pt/IGZO/SiNx/Ta memristor achieves a record-low biomimetic-voltage of 19 mV, near-ideal switching steepness (32 µV/decade), and high endurance (106 cycles). Fabricated into a 1-kb crossbar array with high uniformity, the system implements partitioned processing: a denoising region utilizing volatility for real-time noise suppression, which activate based on input signal strength and operate efficiently without requiring additional erasure steps, and a recognition region based on non-volatility for executing high-precision classification. This end-to-end solution attains 100% accuracy for 10-class audio signals (500 samples) after 15 training epochs with post-denoising, consuming merely 1.44 fJ per operation for denoising, outperforming non-denoised approaches in convergence speed and robustness. Confusion matrices confirm > 90% class-specific accuracy across all classes under noise, establishing a pathway for miniaturized, energy-scalable edge hardware.