Traditional speech recognition systems heavily rely on manually designed acoustic features and are limited by the inherent constraints of models that overlook continuity and dynamic characteristics. This leads to reduced recognition accuracy and high energy consumption. Recurrent neural networks (RNNs) have shown significant potential in temporal-data processing owing to their strong autonomous learning capabilities and numerous adjustable parameters. Here, we present the design of a high-performance Cu/CuBiSe2/SrNbO3/W memristor, which exhibits responses to temporal signals and achieves a 106 ON/OFF ratio, 47.5 pW switching power, 105 cycles duration, and switching time below 200 ns. The memristor simulates several key synaptic behaviors well, such as dual-pulse facilitation and long-term potentiation/inhibition. An efficient semi-hardware speech recognition system was developed based on the buildup of cross-bar array of the memristor. As a demonstration, the array is capable of accurately identifying and distinguishing four common directional sounds. The current works show that the high-performance memristor is promising in applications in brain-like computing, image recognition, and storage technologies.
Jin et al. (2026) studied this question.