Toward Zero-Power Feature Extraction For Speech Wakeup Using Helmholtz Resonator Array

This work demonstrates for the first time a temporal feature extraction method based on a 3D-printed Helmholtz resonator array, which selectively amplifies specific frequency bands for temporal feature extraction, for ultra-low-power speech wakeup applications. In particular, putting the mechanical acoustic filtering ahead of microphones successfully performs classifications of 10 speech commands with performance comparable to traditional Mel-spectrogram, which relies on power-hungry short-time Fourier transform (STFT) or digital/analog filtering. This approach reduces energy consumption for data processing, enabling near-zero-power speech wakeup through ultra-low-power feature extraction, making it an ideal solution for energy-constrained keyword recognition systems.