IEEE Circuits and Systems Magazine - Q2 2023 - 30

I. Introduction
E
dge audio devices are quickly gaining interest
in the Internet of Things (IoT) domain, with
particular focus on low-power devices that perform
smart pre-processing of the input before data
transmission to the cloud. Typical tasks performed
on these devices include voice activity detection
(VAD) and keyword spotting (KWS). As shown in Fig.
1, solutions for reported state-of-the-art edge audio
integrated circuits (ICs) come in two forms. The first
approach samples and quantizes the microphone
output signal at Nyquist or oversampling frequency
through an analog-to-digital converter (ADC). These
data samples are then further processed by a digital
signal processing block such as fast Fourier transform
(FFT), followed by triangular filtering, and logarithmic
compression.
The second approach is to replace the synchronous
ADC and the subsequent signal processing stages with
continuous-time (CT) analog circuits inspired by the biological
modeling of cochleas [1]. These designs implement
the frequency selective filtering properties of the
basilar membrane, rectifying properties of the biological
inner hair cells, and neuronal firing of the ganglion
cells [2], [3], [4], [5], [6], [7], [8], [9]. Since the FFT computation
circuit is typically the most power-hungry building
block of the entire audio feature extractor (FEx)
[10], the analog signal processing has been regarded
as a promising alternative in terms of better power efficiency
[11], [12] thereby it could be useful for tasks
implemented on low-power edge audio devices. The CT
analog filters on the state-of-art edge audio ICs for VAD
[13], [14], [15] and KWS [16], [17] adopt a set of secondorder
band-pass filters (BPFs).
Fig. 2 shows an example of a CT audio processing
stage. Here, a speech sample from the Google Speech
Command Dataset (GSCD) [18] is fed to a 16-channel second-order
Butterworth BPF bank [19]. It can be clearly
seen that each channel responds to different parts of
the speech sample depending on the instantaneous frequencies
in the speech. These filter responses can be
used for training a network on an audio task (see Fig. 1).
This article aims to provide an introductory survey
of voltage-domain CT analog filters leading to the
circuits that have been reported in recent edge audio
ICs. It will provide a unified analysis that covers gCm
and small-signal equivalent diagrams, based on a twointegrator-loop
biquad topology. To the best of our
knowledge, it is the first work to present the operating
principle of voltage-domain second-order filters
using an unified analysis that includes the operational
transconductance amplifier (OTA)-based, crosscoupled
source-follower (XSF), super source-follower (SSF), and
flipped voltage follower (FVF) biquad filters. Simulation
results are also provided to show support for the
proposed analysis. Note that the scope of this article is
geared to review transfer functions and thereby share
intuitive circuit insights, rather than discussing every
performance aspect of analog filter designs (e.g., noise,
distortion, or sensitivity).
The remainder of this article is organized as follows.
Section II introduces the basics of biquad filters and
discusses how a second-order BPF can be implemented
from the twointegrator-loop topology. Section III presents
the notation of a transconductor
which is used in the description of the
filters. Section IV and Section V present
the core analysis of the OTA-based
and source-follower (SF)-based filter
circuits. Section VI summarizes and discusses
the state-of-the-art approaches
for the design of edge audio ICs with
brief future research prospects. Section
VII concludes this article.
II. Review on Biquad Filter
Figure 1. Edge audio processing stages based on conventional and neuromorphic
approaches [20], [21].
The second-order filter is also called as
a biquadratic filter or a biquad filter. It is
because its transfer function is the ratio
of two quadratic equations.
The authors are with the Institute of Neuroinformatics, University of Zürich and ETH Zürich, 8057 Zürich, Switzerland (e-mail: kwantae@ini.uzh.ch;
shih@ini.uzh.ch).
30
IEEE CIRCUITS AND SYSTEMS MAGAZINE
SECOND QUARTER 2023
IEEE Circuits and Systems Magazine - Q2 2023

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q2 2023
