Audio Digit Classification at the Edge Via PMUT Geometric Nonlinearity

We show that a piezoelectric micromachined ultrasonic transducer (PMUT), typically used as an acoustic transmitter and receiver, can also serve as an on-device audio processor for edge computing. Exploiting its intrinsic cubic nonlinearity modeled by the Duffing equation, we implement the PMUT as a physical reservoir computer. Simulations with experimentally calibrated parameters reproduce its nonlinear dynamics and closely match measurements. Using this platform, we perform 10-class audio digit classification (“0”–“9”) and improve recognition accuracy from 86.2% to 97.2% by strengthening the PMUT's nonlinear response. The results show the potential of building compact acoustic systems with the embedded edge processing capability for the time-series classification task.