VoAD: A Sub-μW Multiscene Voice Activity Detector Deploying Analog-Frontend Digital-Backend Circuits

This brief proposes a multi-scene low power voice activity detector deploying both analog and digital circuit, named VoAD. For the speech feature extractor as frontend, the analog circuit is integrated with an extra channel for noise prediction under a wide SNR range. For the voice classifier as backend, a hybrid neural network (HNN) accelerator with precision-aware approximate computing and reconfigurable work modes is proposed to achieve high accuracy and low power. The approximate computing techniques are used to accelerate the neural network layers, the batch-normalization and activation layers. Fabricated with an industrial 22-nm CMOS process, experimental results show that the HNN inspired VoAD can achieve accuracy upto 97.2%, with 508nW power consumption in clean scenes, and over 91.4% accuracy with 578nW in noisy scenes with −5dB SNR, respectively. Compared to the state-of-the-art silicon-based works, the proposed VoAD realizes sub- $\mu \text{W}$ power consumption and enhanced recognition accuracy under multi-scenes in a wide SNR range.