Programmable Speech Recognition Based on Cu/CuBiSe₂/SrNbO₃/W Memristors

Area of Science:

• Materials Science
• Neuroscience
• Computer Science

Background:

• Traditional speech recognition systems struggle with accuracy and energy efficiency due to manual feature design and model limitations.
• Recurrent neural networks (RNNs) show promise for temporal data but require efficient hardware implementations.
• Memristors offer potential for mimicking synaptic behavior and enabling neuromorphic computing.

Purpose of the Study:

• To design and characterize a high-performance memristor for temporal signal processing.
• To develop an efficient semi-hardware speech recognition system utilizing memristor cross-bar arrays.
• To demonstrate the memristor's capability in recognizing directional sounds.

Main Methods:

• Fabrication of a Cu/CuBiSe2/SrNbO3/W memristor device.
• Characterization of memristor performance, including ON/OFF ratio, switching power, endurance, and switching time.
• Simulation of synaptic behaviors like dual-pulse facilitation and long-term potentiation/inhibition.
• Development of a speech recognition system using a memristor cross-bar array.

Main Results:

• The memristor achieved a 10^6 ON/OFF ratio, 47.5 pW switching power, 10^5 cycles endurance, and sub-200 ns switching time.
• The device successfully emulated key synaptic functions.
• The semi-hardware speech recognition system accurately identified four common directional sounds.

Conclusions:

• The high-performance memristor is a promising candidate for energy-efficient, accurate speech recognition.
• The developed system demonstrates the potential of memristors in brain-like computing, image recognition, and storage.
• Further research into memristor-based neuromorphic systems is warranted.