Professional Documents
Culture Documents
INTRODUCTION
tain sequencesof observations
for the purposeof check-
ing the measurement procedures and the speakerand
ONSIDERABLE
variation
processes istobefound
inthe
of speechproductionbecauseof their listenerconsistency.The acoustic measurements were
complexityand becausethey depend upon the past made with the sound spectrograph;to minimize meas-
experienceof the individual. As in much of human urementerrors,a methodwasusedfor rapid calibration
behavior there is a self-correcting,or servomechanism of the recordingand analyzingapparatusby meansof
type of feedbackinvolvedas the speakerhearshisown a complextest tone.Statisticaltechniques wereapplied
voiceand adjustshis articulatorymechanisms. 1 to the resultsof measurements,both of the calibrating
In the elementarycaseof a word containinga conso- signalsand of the vowel sounds.
nant-vowel-consonant phoneme 2.3structure,a speaker's These methodsof measurementand analysishave
pronunciationof the vowel within the word will be beenfound to be preciseenoughto resolvethe effects
influencedby his particular dialectalbackground;and of different dialectal backgroundsand of the non-
his pronunciationof the vowel may differ both in random trends in speakers'utterances.Some aspects
phoneticqualityand in measurable characteristics
from of the vowel study will be presentedin the following
that producedin the word by speakerswith other paragraphsto illustratethe usefulness
of the methods
backgrounds.A listener,likewise,is influencedin his employed.
identificationof a soundby his past experience. EXPERIMENTAL PROCEDURES
Variations are observedwhen a given individual
makes repeated utterancesof the same phoneme.A The plan of the study is illustratedin Fig. 1. A list
very significantpropertyof thesevariationsis that they of words(List 1) waspresentedto the speakerand his
are not random in a statisticalsense,but show trends utterances of the words were recordedwith a mag-
and suddenbreaksor shiftsin level, and other typesof netictaperecorder.The list containedten monosyllabic
nonrandomfluctuations. 4 Variationslikewiseappearin wordseachbeginningwith I-hi and endingwith I-d]
the successive identificationsby a listenerof the same and differingonly in the vowel. The wordsusedwere
utterance. It is probable that the identification of heed,hid, head,had, hod,hawed,hood,who'd,hud, anal
repeatedsoundsis also nonrandombut there is little heard. The order of the words was randomized in each
direct evidencein this work to support such a con- list, and each speaker was asked to pronouncetwo
clusion. differentlists.The purposeof randomizingthe wordsin
A study of sustainedvowelswas undertakento in- the list was to avoid practiceeffectswhich would be
vestigate in a general way the relation between the associatedwith an unvarying order.
vowelphonemeintendedby a speakerand that identi- If a givenList 1, recordedby a speaker,wereplayed
fied by a listener,and to relate thesein turn to acous- back to a listener and the listener wero asked to write
tical measurements of the formant or energyconcentra- down what he heard on a secondlist (List 2), a com-
tion positionsin the speechwaves. parisonof List 1 and List 2 wouldrevealoccasional
In the plan of the study certainmethodsand tech-
niques were employedwhich aided greatly in the LEST
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
176 ß G. E. PETERSON AND H. L. BARNEY
FiG.2. Broadbandspectrograms
andamplitude
sections
of thewordlistby a femalespeaker.
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
METHODS USED IN A STUDY OF VOWELS 177
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
178 G. E. PETERSON AND H. L. BARNEY
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
METHODS USED IN A STUDY OF VOWELS 179
hearingof sounds.
Since the listening group was not given a seriesof
training sessionsfor thesetests,learningwould be ex- 400
[-
pected in the results of the tests.14Several piecesof
evidenceindicate a certain amount of practice effect, 60
I ,
but the data are not suchas to provide anything more (b) OBSERVER
• 50-
(C) OBSERVER013
ments, 109 had lessdisagreements, and 24 had the same
numberof disagreements as in the precedingtests.The
probability of getting this result had there been no
practice or other effect, but only a random variation
of observers'votes, would be about 0.01. When these
data are broken down into three groupsfor the men,
womenand childrenspeakers,the largestdifferencesin
numbersof disagreements for the originaland repeated Ii I I• •! a 3 u u., ^ •,
tests was on the childrens'words, indicating a larger
practice or learning effect on their sounds.The indi- Fro. 4. Observerdisagreements
in listeningtests.
cated learningeffecton men'sand women'sspeechwas
nearly the same.When the data are classifiedaccording within an auditoriumupon intelligibilityhas been ob-
to the vowel sound,the learningeffectindicatedby the servedpreviouslyand is reportedin the literature.1•
repetitions
wasleaston ri-], I-x-I,and [u-I, and greatest
ACOUSTIC MEASUREMENTS
on
Another indication that there was a practice effect Calibrationsof Equipment
lies in the sequenceof total numbersof disagreements
by tests.From the secondto the seventhtest, the total A rapid calibrating technique was developedfor
number of disagreementsby all observersdiminished checkingthe over-allperformance of the recordingand
consistentlyfrom test to test, and the first test had con- analyzingsystems.This dependedon the use of a test
siderably more disagreementsthan the eighth, thus tone which had an envelopespectrumthat was essen-
stronglyindicatinga downwardtrend. With the speak- tially flat with frequencyover the voice band. The
ersrandomizedin their orderof appearancein the eight circuit used to generate this test tone is shown sche-
tests, each test would be expected to have approxi- maticallyin Fig. 5. It consists
essentiallyof an overload-
mately the same number of disagreements. The prob- ing amplifier and pulse sharpeningcircuit. The wave
ability of getting the sequenceof numbersof total dis- shapeswhich may be observedat several different
agreementswhich wasobtainedwouldbe somewhatless
pointsin the test tonegeneratorare indicatedin Fig. 5.
The test tone generatormay be driven by an input
than 0.05 if there were no learning trend or other non-
random effect.
sinewave signalof any frequencybetween50 and 2000
cycles.Figure6(a) showsa sectionof the test tonewith
It was alsofound that the listeningpositionhad an a 100 cycle repetition frequency,which had been re-
effect upon the scoresobtained. The observerswere cordedon magnetictape in place of the word lists by
arrangedin 9 rows in the auditorium, and the listeners the speaker, and then played back into the sourid
in the back 4 rowshad a significantlygreaternumberof spectrograph.The departure from uniform frequency
disagreements with the speakersthan did the listeners responseof the over-all systemsis indicated by the
in the first 5 rows. The effect of a listener'sposition shapeof the envelopeenclosingthe peaks of the 100
•4H. Fletcher and R. H. Galt, J. Acoust. Soc. Am. 22, 93 •* V. O. Knudsen and C. M. Harris, AcousticalDesigning in
0950). Architecture(JohnWiley and Sons,New York, 1950),pp. 180-181.
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
180 G. E. PETERSON AND H. L. BARNEY
w.E. 400 A
GERMANIUM
---- DIODE
250 V
cycleharmonics.With the 100 cyclesfrom the Labora- percent slower when playing back than it did on
toriesstandardfrequencyoscillatoras the drive signal, recording.
the frequencycalibrationof thesystems maybe checked The speedvariationson the soundspectrograph were
very readilyby comparison of the harmonicspacingon measuredwith the test tone applieddirectly, and the
the sectionwith the template scale.The amplitude maximum short time variations were found to be :t=0.3
scalein 6(a) is obtainedby insertinga puretoneat the percent.Suchdirect calibrationsof the frequencyscale
spectrograph in 5 db increments.The frequencyscale of the spectrograph, duringa periodof four weekswhen
for spectrograms may also be calibratedas shownin most of the spectrographic analysiswas done, showed
Fig. 6(b). The horizontallineshereare representationsmaximumdeviationsof +30 cyclesat the31stharmonic
of the harmonics of the test tone when the test tone of the 100 cycletest tone.During that perioda control
generatoris driven by a 500 cyclestandardfrequency. chart•6 of the measurements of the 3100 cyclecompo-
These lines further afford a means of checkingthe nent of the test tone showed a downward trend of about
amount of speedirregularity or wow in the over-all 10 cycles,which was attributed to changesin the elec-
mechanicalsystem.A calibrationof the time scalemay tonic circuit componentsof the spectrograph.As a
be obtainedby usingthe test tone generatorwith 100 result of thesecalibration tests, it was concludedthat
cycle drive and making a broad band spectrogramas the frequencyscaleof the soundspectrograph couldbe
shownin Fig. 6(c). The spacings betweenverticalstria- reliedupon as beingaccuratewithin :t=1percent.
tions in this case correspondto one-hundredthof a
second intervals. Formant Measurements
In the processof recordingsomeof the word lists, Measurementsof both the frequencyand the ampli-
it was arrangedto substitutethe calibratingtest tone tude of the formants were made for the 20 words re-
circuit for the microphonecircuit, and record a few cordedby eachof the 76 speakers. The frequencyposi-
seconds of test tone between the lists of words. When the tion of each formant was obtained by estimating a
word listswereanalyzedwith the spectrograph, the ac- weighted average of the frequenciesof the principal
companyingtest tone sectionsprovided a means of components in the formant. (Seereference4 for a dis-
checkingthe over-allfrequencyresponse of the recorder cussionof this procedure.) When the principalcompo-
and analyzer,and the frequencyscaleof the sectioner. nents in the formant were symmetricallydistributed
The effectof speedvariationsin eitherthe recorderor about a dominant component, such as the second
the soundspectrograph is to changethe frequencyscale. formantof ['A• hudin Fig. 2, thereis little ambiguity
A seriesof measurements with the 100 cycletest tone •6"A.S.T.M. manual on presentationof data," Am. Soc.Test-
showedthat the tape recorderran approximatelyone ing Materials (Philadelphia,1945), AppendixB.
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
METHODS USED IN A STUDY OF VOWELS 181
//
in choosing the formantfrequency. Whenthe distribu-
tionis asymmetrical, however, asin thefirstformantof 400
[-zr']heardin Fig. 2, the differencebetweenestimated 380
formantfrequency andthat assigned by the earmaybe
appreciable. 36O // //// / ß ß
Oneof the greatestdifficulties
in estimatingformant o-•
.J //// ß //
• _/ ß /// R OFDIFFE'RENCES=I?.2
,,- 240 - '/ // ESTIMATED½ OF
-:..:•-<f::--
..-• :::--., ../-- -, , ;•,;,.,"•'•y'*;•, •';,. ,. ,s2; '., :, . } *,.': ;' %.;½.::-.
LL220•/// // DIFFERENCES:
15.3
220 240 260 280 300 320 340 360 380 400 420
..:.: -.½ .., .--. ..... ., .½,• , ,v. ,,' ,'....½1;.:::';•:'½;.; F1 OF FIRST CALLING IN CYCLES PER SECOND
betweenvowelclassification
by ear and by measured
valuesof formant frequencies.
-" 4000 .... ';..... ; .*.;
Amplitudes wereobtainedby assigning a valuein
decibels
to the formantpeak.In the caseof the ampli-
tude measurements it was then necessaryto apply a
0 ' :- -"' '::;L '.... ,•:.:-'-":-::-:'"
-..:-,,'......
'-:':--'-
....
t-t-l:- : ................ correctionfor the over-allfrequencyresponse of the
ß
• . . ... :...... . ..................
system.
.• ......... ..::--.:-:-}
_.;•.:;::::•:**;
-
The procedure of makingduplicaterecordings
and
.• ....
•. . .....
..... .:. . analysesof the ten wordsfor eachof the speakers
providedthebasisforessential
checksonthereliability
of the data.
...........
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
182 G. E. PETERSON AND H. L. BARNEY
Fro. 8. Frequency of secondformant versusfrequency of first The variation of the measureddata for a group of
formantfor ten vowelsby 76 speakers. speakersis muchlargerthan the variationencountered
in repetitionswith the samespeaker,however,as may
steadystateperiodof the vowel.When corrected,these be shownby the data for F1 and F• for the 76 speakers.
88 pointswerewithin the =1= 3 • limits. Of the remaining In. Fig. 8 are plotted the pointsfor the secondcallingby
30 pointswhichwerestill outsidethe limits, 20 werethe eachspeaker, withthepointsidentified accordingto the
result of the individuals' having produced pairs of speaker'sword list. The closedloops for each vowel
soundswhich were unlike phonetically,as shown by have been drawn arbitrarily to enclosemost of the
the resultsof the listeningtests. points; the more extremeand isolatedpointswere dis-
The duplicate measurementsmay also be used to regardedso that in generaltheseloopsincludeabout
show that the difference between successive utterances 90 percent of the values.The frequencyscaleson this
of the samesoundby the sameindividual is much less and Fig. 9 are spacedaccordingto the approximation
significant statistically than the difference between to an aural scaledescribedby Koenig, which is linear to
utterancesof the same soundby different individuals. 1000cpsand logarithmicabove?
An analysisof varianceof the data in Fig. 7 showsthat Considerableoverlappingof areasis indicated,par-
the differencesbetween callings of pairs are not sig- ticularlybetweenE•r-]and Ee-],E•r-]and Ev-],Ev-]and
nificant. However, the value for the varianceratio when Eu-],and Ea-]and Eo-].In the caseof the E•r-]sound,it
comparingspeakersis much larger than that corre- may be easily distinguishedfrom all the othersif the
spondingto a 0.1 percentprobability. In other words, third formant frequencyis used,as the positionof the
if the measurements shownin Fig. 7 for all callingsby third formant is very closein frequencyto.that of the
all speakerswere assumedto constitute a body of second.
statisticallyrandomdata, the probabilityof having a The data of Fig. 8 show that the distribution of
variance ratio as high as that found when comparing pointsin the F1--F•plane is continuousin goingfrom
speakerswould be lessthan one in a thousand.There- soundto sound;thesedistributionsdoubtlessrepresent
TABLEI. Classifications
of vowelsby speakersand by listeners.Vowelsas classifiedby listeners.
i I 8 • o o u u A •
10267 4 6 ... 3 ...........
. 6 954• 694 "5 1 1 .... 56
257 9014 949 1 3 ... i.i "5 51
ß 1 300 9919 2 2 15 39
Vowelsintendedby speakers ß 1 19 8936 1013 '• "' 228 7
... 1 2 590 9534 71 5 62 14
... 1 1 16 51 9924 96 171 19
1 2 78 10196 2
'"1 1 "• 540 '1•'7 103 ... •7• 21
... 23 6 2 3 ...... 2 10243
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
METHODS USED IN A STUDY OF VOWELS 183
i i Ig It• ct o TJ u A 3'
Fundamental
frequencies
M
W 136 232
235 135 130
223 127 124
210 212 129 232
216 137 231
141 221
130 218
133
(cps) Ch 272 269 260 251 256 263 276 274 261 261
Formant frequencies(cps)
M 270 390 530 660 730 570 440 300 640 490
Ft W 310 430 610 860 850 590 470 370 760 500
Ch 370 530 690 1010 1030 680 560 430 850 560
M 2290 1990 1840 1720 1090 840 1020 870 1190 1350
F2 W 2790 2480 2330 2050 1220 920 1160 950 1400 1640
Ch 3200 2730 2610 2320 1370 1060 1410 1170 1590 1820
M 3010 2550 2480 2410 2440 2410 2240 2240 2390 1690
F3 W 3310 3070 2990 2850 2810 2710 2680 2670 2780 1960
Ch 3730 3600 3570 3320 3170 3180 3310 3260 3360 2160
large differencesin the way individuals speak the The plot has alsobeen simplifiedby the omissionof
sounds.The values for F3 and the relative amplitudes [3.•. The [3.• producesextensiveoverlapin the [u•
of the formantsalso have correspondingly large varia- regionin a graph involvingonly the first two formants.
tions between individuals. Part of the variations are As explainedpreviously,however,the [3.-] may be
becauseof the differencesbetween classesof speakers, isolatedfrom the othervowelsreadily by meansof the
that is, men, womenand children.In general,the chil- third formant.
dren'sformantsare highestin frequency,the women's When only vowelswhichreceived100percentrecog-
intermediate, and the men's formants are lowest in nition are plotted, the scatter and overlap are some-
frequency. what reducedover that for all callings.The scatter is
These differencesmay be observedin the averaged greater,however,than might be expected.
formant frequenciesgiven on Table II. The first for- If the first and secondformantparametersmeasured
mants for the children are seen to be about half an from thesewordswell definedtheir phoneticvalues;
octave higher than those of the men, and the second and if the listeningtestswere an exact meansof classi-
and third formants are also appreciablyhigher. The fying the words, then the points for each vowel of
measurementsof amplitudesof the formants did not
show decideddifferencesbetween classesof speakers,
and so have been averagedall together.The formant
amplitudesare all referredto the amplitudeof the first
formantin [a•, whenthe total phoneticpowersof the
vowelsare correctedsoas to be relatedto eachotherby
the ratiosof powersgivenby Fletcher.•a
Various methods of correlating the results of the
listening tests with the formant measurementshave
been studied. In terms of the first two formants the
nature of the relationshipis illustratedin Fig. 9. In this
figuremeasurements for all vowelsof both callingsare
plotted in which all membersof the listening group
agreedwith the speaker.Sincethe valuesfor the men
and the childrengenerallylie at the two endsof the dis-
tributions for each vowel, the confusionbetweenvowels
is well illustratedby their data; thus the measurements
for the womenspeakershave been omitted.
The lines on Fig. 9 are the sameas the boundaries..
drawn in Fig. 8. As indicatedpreviously,somevowels
received100 percentagreementmuch more frequently o •oo ,•oo eoo eoo •ooo •oo •,•oo
•nE•UE•½¾ oF F• iN CYCLESPEn SECOND
than others.
Fro. 9. Frequencyof secondformant versusfrequencyof first
•aH. Fletcher,SpeechandHearing(D. Van NostrandCompany, formant for vowels spoken by men and children, which were
Inc., New York, 1929), p. 74. classifiedunanimouslyby all listeners.
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
184 G. E. PETERSON AND H. L. BARNEY
Fig. 9 shouldbe well separated.Words judged inter- urements on vowel soundsmay be summarizedas
mediatein phoneticpositionshouldfall at intermediate follows.
positionsin sucha plot. In other words,the distribu- 1. Calibration and measurement techniques have been de-
tions of measuredformant valuesin theseplots do not velopedwith the soundspectrographwhich make possibleits
correspondclosely to the distributionsof phonetic usein a detailedstudy of the variationsthat appearin a broad
values. sampleof speech.
It is the presentbelief that the complexacoustical 2. Repeated utterances, repeated measurementsat various
stagesin the vowel study, and randomizationin test procedures
patterns representedby the wordsare not adequately have madepossiblethe applicationof powerfulstatisticalmethods
representedby a single section,but require a more in the analysisof the data.
complexportrayal. The initial and final influencesoften 3. The data, whensoanalyzed,revealthat both the production
shownin the bar movementsof the spectrograms are of and the identificationof vowel soundsby an individualdepend
on his previouslanguageexperience.
importancehere.•ø The evaluationof these changing 4. It is alsofound that the productionof vowel soundsby an
bar patterns of normal conversationalspeechis, of individual is not a randomprocess,i.e., the valuesof the acoustic
course,a problemof major importancein the study of measurements of the sounds are not distributed in random order.
the fundamentalinformationbearingelementsof speech. This is probably true of many other processesinvolving indi-
A further studyof the vowelformantsis nownearing viduals' subjectiveresponses.
5. Finally, the data showthat certain of the vowelsare gener-
completion. This study employs sustainedvowels, ally better understoodthan others,possiblybecausethey repre-
without influences,obtained and measuredunder con- sent "limit" positionsof the articulatorymechanisms.
trolled conditions.The general objectivesare to de-
termine further the most fundamental means of evalu- ACKNOWLEDGMENTS
ating the formants,and to obtain the relationsamong The work which we have discussed has involved the
the variousformantsfor eachof the vowelsasproduced contributionsof a number of people. We shouldlike
by differencespeakers.When this informationhasbeen to acknowledgethe guidanceof Mr. R. K. Potter and
obtainedit is anticipatedthat it will serveasa basisfor Mr. J. C. Steinbergin the plan of the experiment,and
determiningmethods of evaluating and relating the the contribution of Dr. W. A. Shewhart who has assisted
changingformantswithin wordsasproducedby various in the design and interpretation of the study with
speakers. respectto the applicationof statisticalmethods.We
SUMMARY are indebted to Miss M. C. Packer for assistance in
statisticalanalysesof the data. We wishto acknowledge
The results of our work to date on the develop- also the assistancegiven by Mr. Anthony Prestigia-
ment of methodsfor making acousticand aural meas- como,Mr. GeorgeBlake, and Miss E. T. Leddy in the
•9Potter, Kopp, and Green, VisibleSpeech(D. Van Nostrand recordingand analysisof the sounds and in the prepara-
Company,Inc., New York, 1947). tion of the data.
Downloaded 25 Sep 2013 to 165.123.225.35. Redistribution subject to ASA license or copyright; see http://asadl.org/terms