Audio Engineering Society Convention Paper · 2020-05-12 · Audio Engineering Society Convention Paper Presented at the 118th Convention 2005 May 28{31 Barcelona, Spain This convention

Audio Engineering Society

Convention PaperPresented at the 118th Convention2005 May 28–31 Barcelona, Spain

This convention paper has been reproduced from the author’s advance manuscript, without editing, corrections, orconsideration by the Review Board. The AES takes no responsibility for the contents. Additional papers may be

obtained by sending request and remittance to Audio Engineering Society, 60 East 42nd Street, New York, New York10165-2520, USA; also see www.aes.org. All rights reserved. Reproduction of this paper, or any portion thereof, isnot permitted without direct permission from the Journal of the Audio Engineering Society.

Modifications on NIST MarkIII array toimprove coherence properties among inputsignals

Luca Brayda1, Claudio Bertotti2, Luca Cristoforetti2, Maurizio Omologo2, Piergiorgio Svaizer2.

1Institut Eurecom, Sophia Antipolis, 2229 route des Cretes, 06904, France

2Istituto Trentino di Cultura (ITC)-irst, Povo, Via Sommarive 18, 38050, Trento, Italy

Correspondence should be addressed to Luca Brayda or Maurizio Omologo ([email protected],[email protected])

ABSTRACT

This work describes an activity that led to the realization of a modified NIST Microphone Array MarkIII.This system is able to acquire 64 synchronous audio signals at 44.1 kHz and is primarily conceived for far-fieldautomatic speech recognition, speaker localization and in general for hands-free voice message acquisitionand enhancement. Preliminary experiments conducted on the original array had showed that coherenceamong a generic pair of signals was affected by a bias due to common mode electrical noise, which turnedout to be detrimental for time delay estimation techniques applied to co-phase signals or to localize speakers.A hardware intervention was realized to remove each internal noise source from analog modules of the device.The modified array provides a quality of input signals that fits results expected by theory.

1. INTRODUCTIONNowadays, the microphone array technology has animportant impact for a variety of applications, in-cluding human-computer interaction and hands-freetelephony. In particular, research studies are beingconducted inside the European Project CHIL (see[1] and [2]), with the purpose of applying this tech-nology to acoustic scene analysis in meeting and lec-

ture scenarios and, more in general, in any smartroom equipped with multi-channel acoustic sensor-ing. In order to pick-up speech from a distance of 5-6meters as well as to apply effectively enhancementtechniques based on filter-and-add beamforming, theNIST MarkIII array [3] was considered as the bestdevice available for research purposes. This arrayconsists of eight microboards, each having eight mi-

Brayda et al. Modifications on NIST MarkIII array

crophone inputs and related amplification and A/Dconversion stages. The whole digital stream is even-tually made available to the user through a very ef-fective interface to Ethernet.

In general, using a 64-microphone array and an accu-rate time delay estimation technique, as that basedon Generalized Cross-Correlation (GCC) PHAseTransform (PHAT) described in [4] and [5], one cansolve the speaker localization problem and provideenhanced speech in a very effective way. However,system performance can highly depend on the qual-ity of the input signals.

One of the key points to derive excellent results fromthe above mentioned techniques is that input chan-nels be independent each other. For instance, if asynchronous common-mode noise occurs in two mi-crophones, a time delay estimation technique willreveal an artificial coherence at zero sample delay.The latter fact is equivalent to have an active noisesource in front of the array, which actually does notexist.

The present work1 was conducted starting from apreliminary observation that a 50 Hz interferencewas evident in all the input channels of the MarkIIIarray. Once eliminated that source of noise in theeasiest way, i.e. by replacing in-house alimenta-tion with rechargeable batteries, a consistent syn-chronous interference was still present in the inputsignals. Although this interference had a rathersmall dynamics, the coherence between two signalswas still biased at zero samples. To remove com-pletely or to deviate the given electrical interference,the hardware of the device was changed, based onsome substitutions of electrical components (e.g. po-larized capacitors, tension regulators, etc.) as well ason modifications of the power supply ground stagein order to feed each microboard and each micro-phone circuit with an independent power supply.The modification process was conducted in severalsteps, each revealing an objectively quantifiable im-provement with respect to the previous one.

In the remainder of this work, the basic theory ofmicrophone array processing will be introduced to-gether with a detailed description of the array de-

1This work was partially supported by the EU under theIntegrated Project CHIL and by the French MESR (Ministerede l’Enseignement Superieur et de la Recherche).

vice. In particular, the computation of the coher-ence between microphone signals will be describedwith a technical discussion and related figures. Sec-ondly, the basic hardware changes will be discussedwith suitable details for a possible intervention onthe circuitry in order to fix a similar platform. Moredetails about this activity can be found in [6].

2. MICROPHONE ARRAYS

2.1. Microphone arraysThe use of a microphone array for distant-talkinginteraction is based on the potentiality of obtaininga signal of improved quality, compared to the onerecorded by a single far microphone [7, 8, 9]. A mi-crophone array system allows the talker message tobe enhanced as well as noise and reverberation com-ponents to be mitigated, so that it can be used toachieve a hands-free human-machine voice interac-tion.

A microphone array consists of a set of acoustic sen-sors placed at different locations to spatially samplea sound pressure field. Using a microphone arrayit is possible to selectively pick-up a speech mes-sage, while avoiding the undesirable effects due todistance, background noise, room reverberation andcompetitive sound sources. This objective can beaccomplished by means of a spatiotemporal filteringapproach.

The directivity of a microphone array can be elec-tronically controlled, without changing the sensorpositions or requiring the talker to speak close to themicrophones. Moreover, detection, location, track-ing, and selective acquisition of an active talker canbe performed automatically to improve the intelli-gibility and quality of a selected speech message inapplications such as teleconferencing and hands-freecommunication (e.g. car telephony).

2.1.1. BeamformingA beamformer exploits the spatial distribution of

the elements of a microphone array to perform spa-tial filtering [10]. The microphone signals are appro-priately delayed, filtered and added to constructivelycombine the components arriving from a selected di-rection while attenuating those arriving from otherdirections. As a consequence, signals originatingfrom distinct spatial locations can be separated evenif they have overlapping bandwidths.

AES 118th Convention, Barcelona, Spain, 2005 May 28–31

Page 2 of 13


Σ1

s (t)0

s (t)

Delay

Delay

Delay

τ

τ

τ

1

0

w

w

w

0

1 h (t)z(t)

1

h (t)0

Filter

Filter

N-1 Filter h (t)N-1

N-1N-1s (t)

Fig. 1: Filter and sum beamformer.

Delay-and-sum beamforming is the simplest andmost straightforward array signal processing tech-nique. A proper delay τn is applied to each micro-phone signal sn(t), n = 1..N , in order to cophasethe desired component by compensating for the dif-ference in path lengths from the source to each mi-crophone. Each signal amplitude can be weighted bya coefficient wn in order to shape the overall polarpattern, and finally the N signals associated to N

microphones are summed together as follows:

z(t) =

N−1∑

n=0

wnsn(t − τn). (1)

By adjustment of the delays, the array can be elec-tronically steered towards different locations. De-sired main beam width and sidelobe level can beobtained with a proper choice of the coefficients wn.A delay, in the time domain, is equivalent to a linearphase shift in the frequency domain. Therefore thebeamformer can also be implemented by a properphase alignment in the frequency domain accordingto the relationship:

Z(f) =N−1∑

n=0

wnSn(f) e −j2πfτn . (2)

A more general beamforming structure is the filter-

and-sum beamformer described by the equation:

z(t) =

N−1∑

n=0

wn hn(t) ∗ sn(t − τn). (3)

In this case, an additional linear filtering with im-pulse response hn(t) is inserted in each channel priorto summation, according to the scheme of Figure 1,

θ

d

wavefront

0 1 2 N-1

Fig. 2: Uniform linear array in far-field conditions.

in order to perform more complex spatiotemporal fil-tering and directivity shaping, especially when deal-ing with broadband signals (e.g. speech). Moreover,the transfer functions of the filters can be chosenaccording to the statistical characteristics of the de-sired signal and interfering noise.

The directional characteristics of the microphones,the locations of the array elements, and the overallarray geometry provide additional degrees of free-dom in designing the directivity pattern of a beam-former.

2.1.2. Uniform linear arrayThe uniform linear array is the most commonlyused sensor configuration in multichannel signal pro-cessing. It consists of N transducers located on astraight line and uniformly spaced by a distance d.

If a source is far from the array (in the so called“far field” region), then the arrival direction of thesound wavefronts is approximately equal for all sen-sors, and the propagating field can be considered toconsist of plane waves. If a wavefront reaches thesensors from a direction forming an angle θ with thenormal to the array, as depicted in figure 2, then therelative delay τa between wavefront arrivals at twoadjacent microphones is given by:

τa =d

csinθ. (4)

If the source is located close to the array (in the“near field” region), then the wavefronts of the prop-agating waves are perceivably curved with respect tothe dimensions of the array and the arrival directiondiffers from an element to another. In this case, the


Page 3 of 13


relative delays of wavefront arrival at successive sen-sors lie on a hyperbolic curve.

More details on this topic can be found in [8]. Inthe following of this work we refer to the case of amedium size uniform linear array consisting of 64sensors, with 2cm inter-microphone distance.

3. THE MARKIII MICROPHONE ARRAYThe NIST Microphone Array MarkIII [3] is an arrayof microphones composed by 64 elements, specifi-cally developed for voice recognition and audio pro-cessing. It records synchronous data at a samplerate of 44.1 kHz or 22.05 kHz with a precision of 24bits.

The particularities of this array are the modularity,the digitalization stage and the data transmissionvia an Ethernet channel using the TCP/IP protocol.

The array uses 64 electret microphones installed in amodular environment. Two main components con-stitute the system: a set of microboards for record-ing the signals and a single motherboard to transmitthe digital data over the network. There are eightmicroboards in the array, and every microboard isconnected to eight microphones. The first step doneby the microboard is the polarization of the micro-phones and the amplification of the signals. Electretmicrophones need a phantom power to work prop-erly and provide a low voltage signal. So the mi-croboard adapts the signals to be converted in thedigital format. The digitalization of the audio sig-nals is done on each microboard, using four dedi-cated stereo analog to digital converters. The choiceof putting the A/D converters as close as possible tothe microphones reduces the possibility of having theanalog signal disturbed by electrical interferences.

The task of the motherboard is to collect all thedigital signals from the single microboards, multi-plex them and pack all the data in a format suitablefor being sent over the network. The motherboarduses an Ethernet channel to transmit the digital sig-nals: it gets an IP address via a DHCP service andsends broadcast data on the network. If a PC needsaudio signals from the array, it has just to contactthe array using a certain protocol and read the datafrom the network card. Due to the huge amount ofmaterial (64 ch × 44100 samples/sec × 3 bytes =8.07 MB/sec), it has been chosen to use the UDP

protocol. This allows to transfer a big quantity ofdata, but lacks of integrity checks. If the receivingcomputer is momentarily not fast enough to read allthe packets, some packets are simply lost and therecorded signal will contain discontinuities. A soft-ware protocol to resend the lost packets has been im-plemented but is not encouraged for the high chancesto lose data again.

The weak part of the chain is the storing of the dataon the computer. In theory it could be possible toconnect the MarkIII array to a switch and then listento the data from a generic computer on the network.But since the transmission volume is very high, acomputer with a single network interface card is notable to get all the data and loses packets. This isa crucial aspect since missing samples in the signallead to worse performance of any of the above men-tioned technologies. The solution is to install a ded-icated network card on a PC and connect the arraydirectly to that machine. This leads to the loss offlexibility guaranteed by the Ethernet protocol, butat least allows to record seamless data. However,there is the necessity to tune the operating systemfor receiving a lot of UDP packets. This tuning couldnot be done for Microsoft Windows machines, forc-ing the array to be used only with UNIX/LINUXoperating systems. The machine connected to thearray has to be in any case pretty fast, able to storedata without losing incoming packets. This leads tothe necessity to have a dedicated machine only fordata recording, while real-time processing seems notfeasible at the moment.

4. THE CROSS-POWER SPECTRUM PHASETECHNIQUEA microphone array performs a spatial sampling

of the acoustic wavefronts propagating inside an en-closure. It is often of interest the capability of com-paring the signals captured by different microphonesin order to calculate a degree of similarity betweenthem as a function of their mutual delay. Giventwo microphones and their related signals si and sk,it is possible to define a Coherence Measure (CM)function Cik(t, τ) that expresses for each delay τ ,the similarity between segments (centered at timeinstant t) extracted from the two signals. While thetwo microphones are receiving the wavefronts gen-erated by an active acoustic source, this function isexpected to have a prominent peak at the delay cor-


Page 4 of 13


responding to the direction of wavefront arrival (e.g.positive if the source is on the left and negative ifit is on the right). For each microphone pair a bidi-mensional representation of the CM function can beconceived. In this representation horizontal axis isreferred to time, vertical axis is referred to delay andthe coherence magnitude is represented by means ofa “heat” palette (see for example Fig. 6 and Fig. 8).A particularly convenient CM function can be ob-tained starting from a Crosspower Spectrum Phase(CSP) analysis [5, 11], also known as PHAT trans-form, a particular case of Generalized Cross Correla-tion [4]. The procedure for estimating a CSP-basedCoherence Measure (CSP-CM) starts from the com-putation of the spectra Si(t, f) and Sk(t, f) throughFourier transforms applied to windowed segments ofsignals si and sk centered around time instant t.Then these spectra are used to estimate the normal-ized Crosspower Spectrum:

Φik(t, f) =Si(t, f) · S∗

k(t, f)

‖Si(t, f)‖ · ‖Sk(t, f)‖(5)

that preserves only information about phase differ-ence between si and sk. Finally the inverse Fouriertransform of Φik(t, f) is computed:

Cik(t, τ) =

∫

∞

−∞

Φik(t, f)ej2πfτdf (6)

The resulting function (considered as dependent ofthe lag τ) is the transform of an all-pass functionand has a constant energy, mainly concentrated onthe mutual delays at which there is high correlationbetween the two channels. The CM representation isvery useful when it is necessary to locate the acous-tic source or to analyze the multipath propagationinside a room, as delays associated to direct wave-front and to principal reflection are easily detectable.The same representation turns out to be extremelyadvantageous also to analyze the mutual “indepen-dence” between the acquisition channels of an array.In the fact problems as cross-talk or common modenoise components generated within the acquisitiondevice are clearly put into evidence by the appear-ing of graphical patterns (i.e. lines) in the CM thatotherwise, in a quiet environment, should be ratheruniform along the τ coordinate.

5. THE MARKIII/IRST BASED ON BATTER-IESIn this section we describe the problems we encoun-tered with the array, originally designed at NIST[3], and how we solved them. It is worth noticingthat, for project constraints, the underlying purposeof this initial improvement was to obtain a perfor-mant device in a short time, no matter how complex,costly or reproducible the solution would be. Thisimprovement led us to obtain a first new prototypeof MarkIII, from now on called “MARKIII/IRST”.Each of the following subsections describes how thedisturbances were eliminated one by one. In somecases the final solution was obtained after many sub-sequent trials, fully described in [6]. Of all the prob-lems solved in [6], only a subset related to the qualityof the speech signals acquired is presented here.

5.1. Early saturation effect of microphonesIt was observed that when a speaker was near thearray the microphone signals immediately saturated.One could guess that the Panasonic microphoneswere too sensitive or the OPAMPs were pushed tothe limit. In any case, the device did not allowto control input levels. Moreover, it is worth not-ing that some microphones were more sensitive thanothers. The biggest ratio from the most sensitive (ch35 and ch 8, respectively, in the array available atITC-irst) was of 2:1, i.e. 6 dB in amplitude. Sinceno trimmer or other regulations of the input levelwere available, we eventually decided to physicallybypass the first amplification stage as described inthe following and shown in Figure 3. For compari-son purposes, one can find in [3] the original layoutof the NIST device.

The two capacitors C1 and C6, placed at the verybeginning and at the very end of the amplificationstage, were 1 µF polarized capacitors in the originaldesign. They were substituted with two polyester0.47 µF capacitors, which generate much less noise.The first amplification stage was then bypassed viaa 0.47 µF capacitor, keeping the second stage polar-ization to the phantom GND with a 100 kΩ resistor.As a result, the original gain of 68 was reduced to6.8, which is suitable to avoid any signal clipping.

5.2. 50 Hz disturbanceIn our preliminary recordings (done in an insulatedroom) we observed the presence of a perceivable 50Hz interference. We realized the disturbance was


Page 5 of 13


Fig. 3: Modifications of the amplification stage in the first prototype (MarkIII/IRST).

due to the power supply: this problem was solved bysubstituting the 220V-AC to 9V-DC power adaptor,provided with the array, with a Pb rechargeable bat-tery. This was not the final solution, as in a secondstep we solved the device noise problem (see Sec-tion 5.3) by switching to a battery power supply forthe whole analog part. It is worth noting that, evenwith the best battery-based power supply available,still a light 50Hz disturbance persisted: it was muchlower than the one coming from the AC current andit was totally due to environmental electromagneticfields. By consequence it was definitely eliminatedby surrounding the MarkIII with a Faraday cage.

5.3. Device noiseThe device noise represents the major obstacle to

the use of the MarkIII for speaker localization andbeamforming purposes. It is also subtle to detect,as this problem is neither perceivable in normallyreverberant rooms nor evident through waveform orspectral analysis of a single channel.

The device noise problem was evident once elimi-nated the 50 Hz interference (see Section 5.2). In

other words, the following experiments regard theuse of the MarkIII array powered by a rechargeablebattery and installed in a very quiet insulated room.The room is characterized by less than 30 dBA back-ground noise level (that is very close to the acousticsof an anechoic chamber) and a reverberation timelower than 100 ms. Recordings were done at 44.1kHz. As discussed below, the electrical problem canbe revealed both at single channel level (perceptu-ally evident through listening tests) and at inter-channel correlation level (through inter-channel co-herence measurements) analysis.

5.3.1. Single channel analysisThe device noise can be perceptually detected onlyin recordings taken in a very silent room, becausein this condition it can be distinguished from realbackground noise. Alternatively it can be detected,without the need of an anechoic chamber, by man-ually detaching the microphones from the boards:the signals acquired from the array is then only purenoise coming from the devices. The effect of the de-vice noise can be observed in Figure 4, where two


Page 6 of 13


Fig. 4: Spectra corresponding to a 600 ms of back-ground noise. The red, upper line hints at the signalquality of the original MarkIII, while the black, lowerone hints at the signal quality of the MarkIII/IRST.A reduction of 20 dB is evident at most of the fre-quencies.

average spectra of 600 ms of silence sequence areprovided. The red, upper line is relative to a singlechannel of the original MarkIII array. The black,lower line is relative to a silence sequence of the samelength recorded with the MarkIII/IRST. The envi-ronmental conditions were approximately the same,but clearly the device noise affects the whole spec-trum. According to the given figures, more than20 dB noise reduction was obtained at almost allthe frequencies. Another very detailed analysis wasdone by shortcutting each microphone input in or-der to measure only the board circuitry noise and,also in this case, a noise reduction of about 15-20dB was observed. To better understand the entityof the noise, Figure 5 is related to some silence col-lected in the ITC-irst insulated room. From Figure5, one can observe that the noise dynamics (between-300 and +300) involves about 9 bits. It was clearthat losing 9 bits out of the first 16 most significantones was a heavy limitation to the potential of thisarray.

5.3.2. Cross-channel analysisAn analysis of the CSP (described in Section 4) be-tween pairs of channels put into evidence other prob-lems related to the so called “device noise”. Thisnoise component, which can be observed in all thechannels, is neither acoustic noise nor transductionnoise of the microphones. It dominates over acousticbackground noise of a relatively quiet environment.

Fig. 5: Analysis of a background noise sequence of32ms length. The lower left part of the figure reportsthe spectrogram. The log power spectrum is given inthe right part. The device noise is here more evidentboth in its dynamics and in its spectral characteris-tics. Note that the slope of the signal is due to a 2.5Hz interference characterizing the given recordings.

It exhibits a “common mode” within the 8 channelsof each array microboard. Different modules (e.g.from channel 1 to channel 8, and from channel 9to channel 16) have different and uncorrelated noisecomponents. This is evident on the basis of a CSPanalysis.

Figures 6 and 7 show the noise coherence betweenchannels 1 and 8, which was derived from the anal-ysis of a chirp-like signal reproduced through a hi-filoudspeaker placed at the left side of the array: inthis case, a strong coherence is evident between the(mainly electrical) noise sequences. A strong coher-ence at 0 samples is equivalent, for any localizationalgorithm, to determine a direction of arrival froman acoustic source right frontal to the array. In prac-tice, the device noise takes all the energy of the CSPand concentrates it where no sources actually exist.Figure 7 specifically shows how the artificial peak,at 0 samples, dominates the secondary, true, peaklocated at +5 sample delay.

On the other hand, the same analysis repeatedon channels 1 and 9, which are on two differentmicroboards and therefore have no common modenoise, demonstrates the absence of any coherence at


Page 7 of 13


Fig. 6: Chirp signals acquired in an insulated room before the intervention on the device. As the two channelsbelonged to the same microboard, there is a high peak of CSP function at 0 samples inter-microphone delay,which masks the true peak: this means a strong coherence between the device noise sequences.

any particular delay.

5.3.3. Device noise removalThe single and cross-channel analysis clearly showthe effect of the device noise. We describe in the fol-lowing how we detected its origin and how we elimi-nated it. It is worth noting that a better solution wasfound with the next prototype, the MarkIII/IRST-Light. The device noise was caused from the tensionregulator LM2940 (see technical documentation ofMark III in [3]). There is one such a regulator foreach of the 8 microboards. This tension regulatorprovides the operation voltage to 8 Panasonic mi-crophones, to 4 A/D converters and to 8 OPAMPs.As mentioned in Section 5.3.2, the device noise hasa common mode within the 8 channels of each arraymicroboard.

In order to keep the original device layout, the prob-lem was solved by physically removing such regula-tors and feeding the analogue part of every boarddirectly with a circuit of battery designed ad hoc,while the digital part remained fed with a new trans-former stabilized and filtered ad hoc. It is worthnoting that part of the device-noise is caused bythe LM2940 and part by the surface-mounted polar-

Fig. 7: A slice of the CSP-gram in a fixed instantshows the artificial peak of the CSP, which masksthe true one, located at a 5 samples delay.

ized capacitors, which should theoretically removethe regulator noise. These capacitors have an in-ner leakage current which creates the necessary oxidebetween the armors, thus generating a disturbance.Hence, they were substituted with polyester capaci-tors, which are bigger but generate much less noise.An effective solution was to feed the analogue partof each microboard with 4 × 1.2V, 5Ah batteries (atotal of 32 batteries), so to guarantee the galvanic


Page 8 of 13


Fig. 8: Signals extracted from Channel 1 and Channel 8 after our intervention. The peak of the CSP functionreported in the lower part of the figure shows a strong coherence only when the chirp is played.

de-coupling of each power supply source. Neverthe-less the best solution (see Section 6) turned out to berising the power supply ground of each microphonewith respect to the real ground. The use of externalbatteries required an analysis of power consumptionsprior to any decision about the components to buy.This analysis, together with a history of our sev-eral trials and the corrected layouts of the circuitryaround the removed tension regulators, is detailedin [6].

After our intervention the device noise disappeared:Figures 8 and 9 have to be compared with Figures6 and 7 respectively. The delay in samples at whichthe chirps arrive at the two microphones is clearlydetected. In fact, by comparing the CSP coherencemeasures of Figures 6 and 8, it is evident that theconstant yellow stripe at zero samples, caused bythe device noise, has disappeared completely. Withthe new device, the coherence representation is nowhighlighting the true interchannel delay (i.e. +5samples). For a single frame, this fact is evidentin the main peak depicted in Figure 9.

5.4. 8 kHz and 16 kHz common ground noiseA further problem was observed by analyzing thespectrogram of some utterances. This problem be-came evident once both the 50 Hz and the device

Fig. 9: A slice from the CSP-gram in a fixed instantreveals now the true peak at a 5 samples delay. Thedevice noise is totally absent.

noise problems were solved. Two disturbances atabout 8 kHz and 16 kHz appeared in the spectro-gram, as shown by Figure 10: two relatively strongstripes appear in red and violet in the spectrogramon the left part of the picture, which correspond tothe two peaks evident in the right part.

Though the disturbance was present at frequenciesnot closely related to the speech signal, it was veri-fied that it did not come from the environment and itwas then worth to investigate, as it represented an-other common mode noise component across differ-


Page 9 of 13


Fig. 10: The 8 kHz and 16 kHz disturbance peaksare evident in the right part of the picture, wherethe spectrum of a silence segment is taken after theutterance depicted in the left part. Notice the ab-sence of the device noise, removed as described inSection 5.3

ent channels preventing a clean data collection. Wediscovered it was due to the coupling between thedigital and the analog ground. This coupling wasmade around the A/D converter PCM1802: the de-vice was originally provided with two separate pinsfor the two grounds. In the original project of theMarkIII the two pins were connected via a shortcircuit. This makes the analog ground, which theaudio signal relies upon, coincident with the dig-ital ground, which collects the noise coming fromthe various integrated devices, such as the A/D con-verter and the two tension regulators. The final so-lution consists in avoiding the common ground byfeeding each microboard separately with an indepen-dent group of batteries, thus obtaining 8 groups of 4x 1.2V, 5Ah batteries. Figure 11 shows the batterybox entirely built at ITC-irst. More pictures anddetails are available in [6].

6. THE MARKIII/IRST-LIGHTThis section reports on further improvements of theMarkIII/IRST. For the purpose of making the mod-ifications we did in the MarkIII/IRST easily repro-ducible by an expert in electronics in every labo-ratory of the CHIL project [1], we were motivatedto find another solution. The new prototype, from

Fig. 11: Inside of the power supply box: from the8 groups of 4 batteries the power supply passesthrough the red and violet cables, placed on pur-pose in those positions. The transformer, which pro-vides power supply for the digital part (in acquisi-tion state) and recharges the batteries (in rechargingstate), appears in the center of the box.

now on called “MARKIII/IRST-Light”, solves thesame problems reported in section 5 in a very effi-cient, cheap and replicable way. It even performsbetter than the MARKIII/IRST in terms of SNRand coherence measures: see details in [12]. Themultichannel corpus being collected at University ofKarlsruhe [1], is based on the use of this improvedrelease of the device.

6.1. Manual gain correctionIn order to better exploit the acquired signal dy-namic range, in the new layout (Figure 12) we choseto keep both the amplifiers while reducing the to-tal gain and making it tunable: the potentiometerR11 allows the total gain to be in the range 12÷16.7(R11’s nominal value gives a total gain of 15), whichis both a compromise between good amplificationand clipping avoidance, and a way to cope with thedifferent sensitivity of the electret Panasonic micro-phones. Notice that R11 must be of high quality(possibly of plastic-film type).

6.2. High impedance microphone power supplyThe main purpose of the MarkIII/IRST-Light is toreduce complexity and cost while keeping, and pos-sibly improving, performance. It was realized that


Page 10 of 13


Fig. 12: Modifications of the amplification stage in the MarkIII/IRST-Light. Notice the high impedancepower supply stage, which connects each group of 8 microphones on the same microboard to a dual positive-negative power supply.

Fig. 13: Battery saver microboard layout.


Page 11 of 13


performance could even be improved with a differentapproach, taking into account that noises, circulat-ing both on the analog and on the digital ground,could be deviated instead of suppressed. In order tofeed microphones with a very clean supply, a highimpedance path was designed for the DC comingfrom the batteries and each microphone power sup-ply ground level was rised with respect to the realground. A typical π RC cell scheme was built via R1,R2, R3 and C1. This is feasible because the electretmicrophones power consumption is very low. Noticethat:

• C1 and CB are preferably of Tantalium type;

• C2 and C7 are preferably of Polyester type;

• there is one CB every 8 microphones, i.e. oneper microboard.

6.3. Battery saver microboardWe built and inserted into the Faraday cage a furthermicroboard (Figure 14) to power the microphonesonly when the MarkIII is acquiring signals: this issimply done by letting this microboard be driven bythe Capture Led ([3], page 40). The purpose is to letbatteries last as long as possible: we placed a seriesof 4 Alcaline batteries. The new microboard amplifi-cation stage layout is depicted in Figure 13. We esti-mated the MarkIII/IRST-Light can continuously ac-quire for 150 hours with this configuration, but onecould freely make the series voltage be in the range[4,5 - 9V] or different combinations series-parallel toincrease the duration. The battery saver microboardneeds three signals from the motherboard: “Point 1”is the signal coming from the Capture Led, “Point2” is the motherboard power supply for the relais,“Point3” is the motherboard GND. A small batterytester was added to check the batteries state.

7. CONCLUSIONS

This work reported on a recent activity conductedat ITC-irst laboratories which allowed us to realize anew release of the NIST MarkIII microphone array.

The current prototype is able to provide clean signalsthat are suitable for speech enhancement as well asfor automatic speaker localization purposes, thanksto improved characteristics in terms of coherenceamong different channel signals. The MarkIII/IRST

Fig. 14: Battery saver microboard, inserted in theFaraday cage of the array.

Light prototype is presently used at the Universityof Karlsruhe to record a large corpus of seminarsand meetings for benchmarking of various speechand acoustic related technologies under study insidethe Integrated European CHIL project [1].

Moreover, the new hardware layout is being usedat NIST to produce a new generation of MarkIIIarrays, on the basis of the above described inter-ventions. The resulting analogue circuitry, togetherwith a very effective digital section formerly designedand realized by NIST, makes the new device a veryuseful tool for future research and prototyping in thefield of microphone arrays and distant-talking inter-action.

8. REFERENCES

[1] URL: http://chil.server.de

[2] D. Macho et al. “First experiments of automaticspeech activity detection, source localizationand speech recognition in the CHIL project”,Hands-Free Speech Communication and Micro-

phone Arrays Workshop, CAIP, Piscataway,2005.

[3] Cedrick Rochet,URL: http://www.nist.gov/smartspace/toolChest/cmaiii/userg/Microphone Array Mark III.pdf


Page 12 of 13


[4] C.H. Knapp and G.C. Carter, “The generalizedcorrelation method for estimation of time de-lay”, IEEE Trans. on Acoustics, Speech andSignal Processing, vol. 24, n. 4, pp. 320–327,1976.

[5] M. Omologo, P. Svaizer “Acoustic event local-ization using a Cross-power Spectrum Phasebased technique”, in Proc. IEEE Int. Conf.Acoust., Speech, Signal Processing, Adelaide,1994.

[6] C. Bertotti et al.,URL: http://www.eurecom.fr/∼brayda/MarkIII-IRST.pdf

[7] J. Flanagan, D. Berkley, G. Elko, J. West andM. Sondhi, “Autodirective Microphone Sys-tems”, Acustica, vol. 75, pp. 58–71, 1991.

[8] M. Omologo, P. Svaizer, R. De Mori, ”AcousticTransduction”, in R. De Mori (ed.) “Spoken Di-alogues with Computers”, pp. 23-67, AcademicPress, London, UK, 1998.

[9] M. Brandstein and D. Ward (eds.) “MicrophoneArrays Signal Processing Techniques and Appli-cations”, Springer-Verlag, 2001.

[10] B.D. Van Venn and K.M. Buckley, “Beamform-ing: a versatile approach to spatial filtering”,IEEE ASSP Magazine, April 1988.

[11] M. Omologo, P. Svaizer, “Use of the cross-power-spectrum phase in acoustic event loca-tion”. IEEE Trans. on Speech and Audio Pro-cessing, vol. 5, n. 3, pp. 288–292, 1997.

[12] C. Bertotti et al.,URL: http://www.eurecom.fr/∼brayda/MarkIII-IRST-Light.pdf


Page 13 of 13

Audio Engineering Society Convention Paper · 2020-05-12 · Audio Engineering Society Convention Paper Presented at the 118th Convention 2005 May 28{31 Barcelona, Spain This convention

Documents