Highly Sensitive LM@MWCNT Flexible Piezoresistive Sensor for Signal Recovery and Speech Recognition Enabled by CNN

Abstract Flexible piezoresistive sensors often suffer from limited sensitivity, insufficient functional layer adhesion, and weak interference resistance, thereby restricting their applicability in physiological monitoring. A highly sensitive flexible sensor was developed by encapsulating Liquid Metal droplets (LM) within multi-walled carbon nanotube (MWCNT) to form LM@MWCNT composite droplets, which effectively overcame the aggregation issue of MWCNT and could be uniformly coated onto a polydimethylsiloxane (PDMS) substrate featuring an undulating microstructure. LM droplets were extruded via a mechanical sintering process to infiltrate the gaps, and the subsequent formation of an oxide layer upon oxygen exposure further strengthened the interfacial adhesion between the LM@MWCNT coating and the PDMS substrate. The assembled LM@MWCNT flexible piezoresistive sensor demonstrated a high sensitivity of 14.81 kPa⁻ ¹, exhibiting rapid response and recovery times of 43 ms and 46 ms, respectively, alongside excellent stability during 3000 cycles. By leveraging convolutional neural network (CNN), pulse signals were effectively processed to enhance their stability, while vibration signals generated during throat vocalization were accurately classified, enabling signal recovery after interference (MAE = 0.007) and precise speech recognition with an accuracy of 0.918. The proposed sensor offers significant potential for real-time physiological monitoring and voice-interactive communication.