A 28-nW Noise-Robust Voice Activity Detector with Background Aware Feature Extraction

In light of the increasing number of Internet of Things (loT) devices, such as intelligent vehicles and smart assistants, it has become imperative to develop low-power Voice Activity Detection (VAD) devices. The always-on VAD devices detect the voice to wake up the target system, thus dominating the standby power consumption of the loT devices. The typical VAD consists of a feature extractor and a neural-network-based classifier. The algorithm using the frequency-domain features which can be obtained by modulation frequency [1], fast Fourier transform (FFT) [2], and analog filter banks, can achieve high detection accuracy. However, the feature extractor induces high power consumption due to the complex operations. Alternatively, the time-domain analog VADs [5]-[6] achieve low power consumption, due to the lack of a frequency extractor, but also suffers from reduced accuracy that the audio amplitude is interfered with noise easily, especially in noisy environments with a signal-to-noise ratio (SNR) is lower than OdB. In summary, achieving high accuracy and low power consumption simultaneously in VAD devices is a critical challenge.