Reservoir computing utilizing <scp> HfO ₂ </scp> ‐based ferroelectric neuromorphic devices with <scp> WO _<i>x</i> </scp> nano insertion layers for efficient speech recognition

Abstract Reservoir computing (RC) presents a computationally efficient alternative to conventional recurrent neural networks (RNNs) for temporal‐data processing. Traditional bio‐inspired auditory systems often face constraints due to limited computational power and high energy consumption, which impede speech‐recognition accuracy. In this work, we demonstrate high‐performance ferroelectric neuromorphic devices based on TiN/WO x /Hf 0.5 Zr 0.5 O 2 (HZO, 4 nm)/TiN heterostructures for constructing an artificial auditory nervous system for efficient voice recognition. The device exhibited a high remanent polarization ( P r ) of approximately 20.58 μC cm – 2 at 1.8 V and endurance exceeding 10 10 cycles. Density functional theory calculations and experiments confirm that the WO x interlayer regulates oxygen vacancy formation and migration within the HZO layer. By emulating essential biological synaptic plasticity functions, such as paired‐pulse facilitation and long‐term potentiation/inhibition, the ferroelectric tunnel junction‐based devices can perform signal processing and neural computation within the RC framework, achieving an accuracy beyond 99% across 12 categories of everyday vocabulary voice words. These findings provide a promising pathway for developing highly reliable and energy‐efficient neuromorphic artificial auditory systems. image