-
7/27/2019 2003 Lu Ieee Iscas
1/4
A MICROPOW ER ANALOG VLSI PROCESSING CHANNEL FOR BIONIC EARS
ANDSPEECH-RECOGNITION FRONT ENDSTimoihyK.-T Lu, Michael Baker,
ChristopherD. althouse, Ji-Jon Sit, Serhii Zhak , and Rahul
Sarpeshkartimlu@,mit.edu an d rahuls@,mit.eduAnalog VLSI and B
iological Systems Group, Research Laboratory of
ElectronicsMassachusetts Institute of Technology
I1 Mass. Ave.Cambridge, MA 02139A B S T R A C T
Next-generation bionic ears or cochlear implan ts will be
fullyimplanted inside the body of the patient and consequently
havevery stringent requirements on the power consumption used
forsignal processing. We describe a low-power program mable
analogVLSI processing channel that implements bandpass
filtering,envelope detection, logarithmic mapping and
analog-to-digitalconversion. A bionic ear processor may be
implemented throughthe use of several such parallel channels. In a
proof-of-concept 1.5pm AMI M O S S implementation, the most
power-hungry channelof OUT system (7.5kHz center frequency)
consumed 7.8 pW ofpower, had an internal dynamic range (IDR) of
51dB. and provided64 discriminable levels of loudness per channel.
Such numbersalready satisfy the requirements of todays commercial
bionic earprocessors and can lower the power consum ption of even
advancedDSP processing schemes of the future by an order of
magnitude.Our processing channel is also well suited for use in low
powerspeech recognition front ends, which com monly require the sam
esequence of operations in cepshum-like front ends.
Futureimprovements in the interfaces between the various stages of
ourprocessing channel, which were nut optimized in
thisimplementation, promise a potential intemal dynamic range
ofmore than 60dB with little or no increase in power.
1. I N T R O D U C T I O NIn the past 25 years, the development
of bionic ears (BEs) has beensuccessful in restoring hearing tu the
profoundly deaf bystimulating the auditory nerve with implanted
electrodes to mimicthe natural response of the ear to sounds.Figure
I shows a common approach to signal processing in BEs.The .micr
ophon e transduces audio sign als to electrical signalswhich are
then fed into the audio front end (AFE). The AFE sensesand pre-am
plifies the microphone sign al, pre-emphasizes impo
rtantfrequencies in speech through filtering, and uses an automatic
gaincontrol (AGC) system tu compress the 8OdB input dynamic rangeof
sounds (30dB SPL-IlO dB SPL ) into a narrower internaldynamic range
(IDR) for subsequent processing. The IDR istypically 50dB. The
subsequent processing is composed of severalparallel channels each
of which extracts the envelope energy in adefined frequency band
and maps it via a logarithmic function intoa narrower dynamic range
of perceivable electrode stimulation.Todays processors typically
have 16 channels that together spanthe entire audio frequency
spectmm from 250Hz-IOkHz. Figure I
only shows two such channels. In our implementation, the
outputof the logarithmic map is digitized and sets the
programmableDAC bits of electrode-stimulation circuits. The
stimulationcurrents from the electrodes excite the auditory nerve
to evoke thesensation ofhearing [ I ] .
F i g u r e 1 . A C o m m o n A p p r o a c h to S i g n a l P r
o c e s s i n g i nBionicEarsIn this paper, we will not report on
implementations of the AFEand AGC circuits, since they are nut part
of the parallel channels ofprocessing but common to all of them,
Our own implementationsof these circuits, however, yield a combined
power consumptionfor the AFE and AG C which is about 125 pW .Analog
implementations afford power savings over thecombination of an
A-to-D converter and a DSP processor: Evenwith Moores law scaling,
the latter schem es are unlikely tu lowerthe power consumption of
32-channels of processing below a fewmilliwatts or 10 mW. The total
power consumption of a systemthat uses 32 of our processing
channels amounts to less than 0.4mW. Thus, our analog
implementation promises an order-of-magnitude improvement in power
consumption over that of evenadvanced DSP designs. Furthermore,
like digital implementations.our analog processing channel is
programmable.Subthreshold-MOS, silicon- cochle as, and analog
circuits forcochlear-implant processing have been previously
proposed [?-7]as means for implementing complex signal processing
with verylow power. This work proves the promise of such prior work
byachieving numbers that make an analog processing
systemcommercially feasible in the near term.
V-41-7803-776 -3/03/%17.002003 EE E
mailto:timlu@,mit.edumailto:rahuls@,mit.edumailto:rahuls@,mit.edumailto:timlu@,mit.edu
-
7/27/2019 2003 Lu Ieee Iscas
2/4
We implemented our system with the MOSlS AMI 1.5pmBiCMOS process
with 2.8V power supp lies. The chip contains ananalog front end,
two complete channels, and other'circuitry fordiagnostic purposes.
We built two complete channels to addressissues of matching that
are natural , to ask in analogimplemen tations. The layou t of the
system is shown in Figure 2 .T he two channels and the au dio front
end c an entirely fit in half ofthe 4.6 mm x 4 .7 mm chip used,
althou gh this was not done in thisimplementation.
Figure 2. VLSl Layout of Two Channels of Our System.This Chip is
4.6 mm x 4.7 m m on a 1.5 pm Process.The organization of this paper
is as follows. Section 2 describesthe programmable bandpass
filters. Section 3 presents theoperation of the envelope detectors.
Section 4 discusses theimplementation of the logarithmic map
circuit used to providedynamic range compression and AID
conversion. Section 5presents experimental data demonstrating the
operation of th eentire system. Section 6 describes how our system
could also beused as a front-end for speech recognition system. We
summ arizeand conclude our findings in Section 7 .
2. THE PROGRAMMABLE BANDPASS FILTERBandpass filters are a
crucial component in the signal-processingchain in BEs. The array
of bandpass filters mimics the frequency-to-place transformation of
the biological cochlea: High frequencysounds stimulate the basal
region of the auditory nerve while lowfrequency sounds stimulate
the apical region of the auditory nerve.In a BE, the electrodes
corresponding to high-frequency channelsprimarily stimulate basal
regions of the audi tory nerve, whi le theelectrodes corresponding
to low-frequency channels primarilystimulate apical regions of the
au ditory nerve.Subthreshold Gm-C filters are well suited for use
in BEs becausethey offer low power consumption and can be tuned
over a widefrequency range to cover the spectrum of hearing [8].
Thecapacitive-attenuation filter used in the cha nnel described
here hadfirst-order rolloffs and was single-ended like that
described in [8].Higher-order and fully differential versions of
such filters aredescribed in [ 9 ] .
The center frequency of the filter is programmed via 5 DAC
inputbits which set the bias current of the transconductor and
yield atotal of 32 possible different configurations to span the
frequencyrange of hcaring. Alteration of fhe DAC reference current
providesa further degree of freedom if needed. The programmability
of thefilter is important in acco mmo dating v ariations in
patients,
1D 10' t 07 roJ to4-yrr)
Figure 3. Programmability of the Bandpass Filter withDAC
Settings.Experimental results from two separate channels are shown
inFigure 3, which presents the programmability of the
capacitive-attenuation bandpass filters. As th e DA C bits are
adjusted, thecenter frequencies of the filters move from about 250
Hz to 10kHz. Furthermore, Figure 3 exhibits the close matching
betweenfilters from different channels on the same chip. Measurem
entsfrom t h i s chip demonstrate that the filters can swing 848
mVrmswith 5% T H D and have noise floors of 200 pVrms, yielding
aminimum dynamic range o f72 dB for this par t of the channel .The
power required for high-frequency filters is larger than
thatrequired for low-frequency filters. For a SkHz - IOkHz filter,
ourmeasured power dissipation was 2. 1 p W per channel;
powerconsumption dropped to 0.14 pW per channel for a 100 Hz ~
200Hz filter. Fully differential filters with second-order rolloff
slopeshave power consumptions of 0.23 pW and 6.36 pW for the
samefrequency ranges [9].
3. THE ENVELOPE DETECTOREnvelope detectors are important in the
design of BEs since theytransform the energy of bandpassed audio
signals to informationfor patients to process. Th e envelope
detection strategy used inthis paper is similar to that described
in (31 and consists of arectifying stage and a peak-detector stage.
Circuit innovations inthe rectifying stage allow us to achieve
superior dynamic range atthe same power consumption and are
described in some detail in acompanion paper at this conference
[IO]. We shall only brieflydescribe the ope ration of the envelope
d etector in this paper.
V-42
-
7/27/2019 2003 Lu Ieee Iscas
3/4
Th e envelope detector uses a wide-linear-range transconductor [
I I]to drive a class B mirror and thus perform rectification
andvoltage-to-current conversion. The rectified currents from
themirror are summed to produce a full-wave output and are fed into
acurrent-mode peak detector with asymmetric attack and
releasetimes. The peak detector is identical to that used in [3].
In ourimplementation, the release time is adjustable to suit
thepreferences of the patient. A slow feedback loop performs
offsetcorrection and ensures that offsets in the
wide-linear-rangetransconductor and in the two halves of th e
class-B mirror do notgreatly degrade the minimum detectable signal
of the envelopedetector.Figure 4 shows experimental results
obtained from the outputcurrent of the envelope detector circuit in
response to varying inputamplitudes at frequencies of IOOHz, kflz,
and I O kHz. Thelinear dynamic range of the envelope detector
circuit at 100 Hz sdemonstrably 60 dB if we only allow + IO%
deviations fromlinearity; at 1 WZ, he linear dynamic range is 59
dB. At 10 kHz,as described in [IO], high-frequency operation of the
envelopedetector results in residual dead-zone effects from the
rectifier inthe circuit, and the linear dynamic range of the
envelope detector isdegraded to 49dB. However, Figure 4 shows that
a monotonicresponse with good signal-to-noise is still obtained
over the entire60dB range of operation. We measu red a power
consumption of2.8 pW. Further unpublished optimizations in the
envelopedetector or an increase in power consumption to 5pW [ IO ]
canyield almost 60dB of linear dynamic range at all frequencies.
Inthis paper, we did not explore these issues further since the
overalldynamic range for the channel was lowered by interstage
couplinganyway.
F i g u r e 4. E n v e lo p e D et e c to r O u t p u t C u r r
e n t as aF u n c t i o n o f I n p u t A m p l i t u d e f o r D i
f f e r en t F r e q u e n c i es .D e v ia t io n s f r o m L i n
e a r i t y of i10% a re also s h o w n [IO].
4. LOGAFUTHMICMAPThe electrical dynamic range that is
psychophysically observed indeaf patients is usually between 3 dB
and 30 dB, with a typicalvalue being about 10 dB [IZ]. The
logarithmic-map stage maps the40dB-60dB internal dynamic range in
envelope-energies range intothis electrical dynamic range. It does
so by having the electrodestimulation currents be a linear function
of the logarithm of theenvelope energy. Independent of their
electrical dynamic range,deaf patients appear to perceive changes
in sound intensity ofabout I dB.Thus, a good patient with 30dB of
electrical dynamicrange may be able to resolve about 30
discriminable levels; toensure that such perception is possible,
the logarithmic map mustbe precise to at least 5 output bits. We
may achieve all of thesespecifications by building a low-power
current-input logarithmicA/D convener that is at least 5-bit
precise.It is generally accepted that the envelope in each band of
speechdoes not vary significantly faster than 100 Hz. Thus,
thelogarithmic AID converter need only sample at a rate greater
than200 Hz, the Nyquist frequency.The low power and relatively slow
bandwidth requirements weremet with a 6-bit diode-based logarithmic
dual-slope A D converter.Several circuit innovations to cancel
offset and temperaturedependence in the logarithmic map circuit
were employed but arebeyond the scope of this paper.
Iln (AIF i g u r e 5. Performance of the L o g a r i t h m i c M
ap Circuit.Figure 5 shows the overall performance of the
logarithmic mapcircuit. It is able to output a linear range in its
digital code as themeasured input current from the envelope
detector varies inconstant ratio increments over a 60 dB dynamic
range from 200 pAto 200 nA . Figure 5 was measured for a 7.5kHz
input. Wemeasured 1.68 pW in analog power and 1.26 pW in digital
powerfor this stage yielding a total of about 3 pW .
v-43
-
7/27/2019 2003 Lu Ieee Iscas
4/4
5. OVERALL SYSTEM PERFORMANCEThe overall dynamic range of a
single channel in this system wasmeasured for a 7.5kHz input to a
bandpass filter centered at7.5Khz. Figure 6demonstrates linear
operation from a 4 mVppinput to a 1.5 Vpp input, which is a dynamic
range of 5 1 dB. Ouroverall dynamic range is less than that of each
of our stagesprimarily because our proof-of-concept design did not
ensure thatthe minimum output noise floor stage of each stage of in
thecascade matched the minimum input noise floor of the next
stageof processing in the cascade. If these optimizations are made
infuture designs, we expect to achieve an overall dynamic range
ofoperation of 60dB with little or no increase in power.
.. VhM_Figure 6. Overall System Performance of the ChannelThe
total analog power measured for each channel is comprised of2.1 pW
n the bandpass filter, 2. 8 pW in the envelope detector, and3 p W n
the logarithmic map. Thus, each channel requires a totalof 7.9 p W
. For a 32-channel implant, the expected powerconsumption is thus
about 25 6 pW. f we add the 125 pW powerconsumption of the AFE, a
fully functional 32-channel bionic earprocessor would require less
than 4 0 0 pW of power. Using filterswith second-order rolloffs [9]
do not alter this conclusion if thepower scaling with center
frequency in the filters is included.Thus, our analog processing
channel offers an order-of-magnitudeimprovement over even advanced
A-D-and-DSP implementationsof the futurc.
6. USE FOR SPEECH-RECOGNITION FRONTENDSThe processing channel
that we have described may be easilyprogrammed to create an array
of filters that form a M e1 filter bank[13 ] . The logarithm of the
envelope energies of these filters thenyields a very cepstral-like
representation except that the final stepof the computation, the
discrete cosine transform is omitted [14].The latter transform, or
alternate transforms, may easily beperformed on the digital numbers
output by OUT channel in arelatively cheap fashion by a subsequent
DSP. The relatively
cxpensive filtering and log operations are performed by our
analogchannel. saving power.7. CONCLUSIONS
Experimental data from our chip proves that analog
VLSlimplementations o f processing channels for bionic ears or
low-power speech-recognition front ends ca n yield
order-ofmagnitudepower savings over even advanced DSP
implcmentations of thefuture. Such im plementations are therefore,
likely to be very usefulin fully implanted bionic ear systems or in
portable speechrecognition system s of the future , especially if
they areprogrammable like OU T implementation.
REFERENCES[I][2 ]
P. C. Loizou, "Mimicking lhe human car," IEEE Signol process
in^Mng. , Vol. 5,pp. 101-130, Sept. 1998.R. J . W . Wang, R.
Sarpcshkar, M. Jahri, and C . Mcad. "A low-poweranalog front-end
module for cochlcar implants." prcscntcd at the XV IWorld Congress
on Otorhinolatyngalogy, Sydney March 1997.R. Sarpeshkar, R. F-Lyon,
and C. A . Mcad, "A low-pawcr wide-dynamic-rangc analog VLSl
cochlea." Analog I n r e gmed Circsirs andSignol Pruce.wing,Vol.
16, pp . 245-274, 1998.W . Gemanov i x and C. Taurmarou, "Design of
a micropowci wrrcnt~modc log-domain analog cochlear implant.'' IEEE
Trans. o n CircuitsondSystemsI1: A n a i u g o n d DS P, Val. 47,
pp. 1023-1046, Oct. 2000.T. S.Landc, J. T. Marienborg. and Y .
Bcrg, "Ncuromorphic cochleaimplants." In Proc . o / r he IEEE
Inrernationol Symposium "11 Circurtson d Sy1enr.r ISCAS ZOUO), Vol.
4, pp. 40 1.404, 200 0.R. F. Lyon and C . Mead, "A n analog
electronic cochlea." IEEE Trons.on Acur r s l i cs . Speech, and
Signal Processsing.. Vol. 36 , pp. I 119-1134.Jul. 1988.
171 W. Liu, A.G. Andrcou, and M.H. Goldstcin,
"Voiced-speechrcpresentation by an analog silicon model of thc
auditory pcriphcly."IEEE T r a m on Neural Ne w o rk r , Vol. 3,
No. 3, pp. 477-487. M ay1992.C. D. Salthouse and R. Sarpeshkar. "A
micrapowcr band-pass filterfor us c in bionic cars,'' In Proc.
"/,hi- I E E E Inreroocronal Sympo.siunlon Circuilj andSwrenrx
(ISCAS 2002). Vol. 5 . pp. 189-192.2002.[91 C. D . Salthouse and R.
Sarpcshkar. "A practical micropowerprogrammable handpass filter for
usc in bionic cam," IEEE Journal on[ l o ] M . Bakcr and R.
Sarpeshkm, "A micropower envclope detcctor foraudio applications",
submitted for revicw to rh e IEEE lnlernalivnol
Symposium on Csircuits andSysrems. 2003.[Ill R. Sarpeshkar, R.
F. Lyon, and C. A . Mcad, "A low-power wide-l inear~mnge
ransconductancc amplifier," Analog Inlegra,rd Circails
a n d S i ~ ~ a l P r o c e s s i n g ,ol. 13,No. 112. MaylJune
1997.[ I 2 1 J.-J. Sit, "A low-power analog logarithmic map circuit
with offset andtcmpcraturc compcnsation for use in bionic cars."
M.S. thesis,Massachusctts Institute ofTcchnolo gy, Cam bridge, MA,
2002.[ 13 1 B. C . J . Moore, h lmducc ion Io the
PJychologvofHearing, Academic
Press, London, 1997.[ I 41 L. Rabiner and 8:H. Juang,
Fundomenlab qf Speech Recognition.Englcwood Cliffs, NJ . Prenticc
Hall, 1993.
[3 ]
[4 ]
[5 ]
[6 ]
[ S I
Sol idsrale c i r c u i r s , in prcss, 2003.
V-44
