Ultra-Low-Power Voice Activity Detection System Using Level-Crossing Sampling

This paper presents an ultra-low-power voice activity detection (VAD) system to discriminate speech from non-speech parts of audio signals. The proposed VAD uses level-crossing sampling for detection. useless samples in the signal are eliminated due activity-dependent nature this scheme. A 40 ms moving window with a 30 overlap is exploited as feature extraction block, within which output analog-to-digital converter (LC-ADC) counted feature. only variable used distinguish and segments input number LC-ADC time window. achieves average 91.02% hit rate 82.64% over 12 noise types at ?5, 0, 5, 10 dB signal-to-noise ratios (SNR) TIMIT database. including LC-ADC, extraction, classification circuits was designed 0.18 µm CMOS technology. Post-layout simulation results show power consumption 394.6 nW silicon area 0.044 mm2, makes it suitable always-on device automatic recognition system.