Towards Digital Synthesis of Variable Q-Factor Direct-Conversion for Low-Power Edge Sensing
Parthojit Chakraborty, Kazuki Maari, Jim Bartels, Alexandre Varieras, Aravind Tharayil Narayan, Ludovico Minati, Shiro Dosho, Hiroyuki Ito · 2023
Feature extraction through modulation of input signal is promising for reducing complexity of digital neural network architectures. This prompts research into design paradigms such as direct-conversion that have not been explored yet for low-power edge sensing. This paper reports direct-digital conversion receiver of discrete-time input voice signal into pulse stream or spikes towards speech recognition. Feedforward neural network trained with receiver output operating 11 frequencies yields 96% accuracy. Output spikes of receiver circuit implemented in HDL is a function of input frequency - 80 % spike rate decrease from intended value ascribed to low-resolution input signal is reported.