Search

This paper presents an ultra-low-power voice activity detector (VAD). It uses analog signal processing for acoustic feature extraction (AFE) directly on the microphone output, approximate event-driven analog-to-digital conversion (ED-ADC), and digital deep neural network (DNN) for speech/non-speech classification. New circuits, including the low-noise amplifier, bandpass filter, and full-wave rectifier contribute to the more than 9 $\times $ normalized power/channel reduction in the feature extraction front-end compared to the best prior art. The digital DNN is a three-hidden-layer binarized multilayer perceptron (MLP) with a 2-neuron output layer and a 48-neuron input layer that receives parallel event streams from the ED-ADCs. To obtain the DNN weights via off-line training, a customized front-end model written in python is constructed to accelerate feature generation in software emulation, and the model parameters are extracted from Spectre simulations. The chip, fabricated in 0.18- $\mu \text{m}$ CMOS, has a core area of 1.66 $\times $ 1.52 mm2 and consumes 1 $\mu \text{W}$. The classification measurements using the 1-hour 10-dB signal-to-noise ratio audio with restaurant background noise show a mean speech/non-speech hit rate of 84.4%/85.4% with a 1.88%/4.65% 1- $\sigma $ variation across ten dies that are all loaded with the same weights. [ABSTRACT FROM AUTHOR]

In this paper, we investigate neural circuit architectures encoding natural visual scenes with neuron models consisting of dendritic stimulus processors (DSPs) in cascade with biophysical spike generators (BSGs). DSPs serve as functional models of processing of stimuli up to and including the neuron's active dendritic tree. BSGs model spike generation at the axon hillock level where neurons respond to aggregated synaptic currents. The highly nonlinear behavior of BSGs calls for novel methods of input/output (I/O) analysis of neural encoding circuits and novel decoding algorithms for signal recovery. On the encoding side we characterize the BSG I/O with a phase response curve (PRC) manifold and interpret neural encoding as generalized sampling. We provide a decoding algorithm that recovers visual stimuli encoded by a neural circuit with intrinsic noise sources. In the absence of noise, we give conditions on perfect reconstruction of natural visual scenes. We extend the architecture to encompass neuron models with on–off BSGs with self- and cross-feedback. With the help of the PRC manifold, decoding is shown to be tractable even for a wide signal dynamic range. Consequently, bias currents that were essential in the encoding process can largely be reduced or eliminated. Finally, we present examples of massively parallel encoding and decoding of natural visual scenes on a cluster of graphical processing units (GPUs). We evaluate the signal reconstruction under different noise conditions and investigate the performance of signal recovery in the Nyquist region and for different temporal bandwidths. [ABSTRACT FROM PUBLISHER]