Leaky Integrate‐and‐Fire Dynamics of HZO/ZnO Neuromorphic Devices for Edge Audio Recognition Applications

Abstract

Always‐on acoustic recognition is becoming essential for intelligent vehicles, robotics, and embedded sensor nodes, yet conventional audio pipelines remain energy‐intensive because they rely on continuous sampling, digital feature extraction, and computation of heavy inference. Materials that encode temporal sound features directly in hardware can enable privacy‐preserving on‐device intelligence with reduced complexity. Here we demonstrate device‐level temporal encoding for edge sound recognition using a ferroelectric oxide heterostructure that merges leaky integration, facilitation, and short‐term relaxation within a single two‐terminal device. The W/HZO/ZnO/W device exhibits polarization‐controlled transport coupled to oxygen vacancy dynamics, yielding a pronounced polarity‐dependent response with memory‐like behavior in one bias regime and resistive operation in the opposite regime. Under pulse stimulation, the device shows progressive current accumulation followed by gradual decay, enabling paired pulse facilitation and short‐term memory without an external reset. Integrated with a microcontroller that converts audio waveforms into pulse trains, distinct spoken commands generate reproducible transient current fingerprints that serve as separable physical features for recognition. Complementing the measured device encoding, network‐level simulations using a compact multi‐branch spiking architecture indicate the classification achievable when this event‐coded representation feeds a trained spiking network, pointing toward event‐driven inference with lower computational load than dense digital baselines.