Time-Encoded Mostly Digital Feature Extraction for Voice Activity Detection Tasks

In this manuscript, we propose a time-domain, mostly digital band-pass filter architecture for audio signal feature extraction. The filter is composed of a single voltagecontrolled-oscillator (VCO) as signal encoder and monostables generating digital pulses at the edges of the pulse frequency modulated VCO output. The band-pass filtering phenomena are observed when we combine these digital pulses with different lengths, and its functionality is proved by the behavioral simulations performed with a circuit proposal built in MATLAB/Simulink. Compared to other state-of-the-art solutions, our filter architecture can take the benefit from advancement of digital technology nodes and design tools thanks to its mostly digital implementation. Moreover, the band-pass filter possesses friendly programmability, whose center frequency can be easily tuned by modifying the digital pulse length, and the single VCO encoder promises an easy calibration circuit. Finally, a recurrent neural network is trained and tested based on the features extracted from this architecture, the excellent speech/non-speech hitting rate shows the feasibility of our band-pass filters for voice activity tasks.