Low-power multichannel spectro-temporal feature extraction circuit for audio pattern wake-up

In many distributed sensing applications, continuous sensor monitoring requires processing with a significant energy footprint, which hinders autonomous operation and battery lifetime of sensor nodes. In our research we explore the power savings gained by splitting the hardware architecture for continuous monitoring into two stages: an always-on ultra-low-power mixed-signal wake-up circuit placed near the sensor, performing coarse recognition (e.g. wake-up circuit) and waking up the main digital processing unit only on event detection. This enables for activation of energy-hungry digital processing only at the rate of event occurrence without penalising responsiveness and monitoring continuity. We focus on the wake-up circuit performing recognition of spectro-temporal audio patterns, consisting of spectro-temporal feature extraction, and the classification sub-circuits. We propose a novel design of the feature extraction circuit. It consists of a spectral decomposition multi-channel analog band-pass filter bank, implemented in generalized impedance converter topology (GIC), and the bank of passive channel detectors for measuring the intervals of in-band signals. Experimental filter characterization demonstrated the benefits of proposed filtering topology for low-power applications in the audio frequency range even with operational amplifiers of very limited bandwidth. Detector's response was verified in multi-channel environment. Preliminary analysis showed power consumption ranging from 10.5 to 13.5 μW per channel using off-the-shelf components.