THE JOURNAL OF THE ACOUSTICAL SOCIETY OF AMERICA VOLUME 24, NUMBER 2 iMARCH, 1952
Control Methods Used in a Study of the Vowels
GORDON E. PETERSON AND HAROLD L. BARNEY
Bell Telephone Laboratories, Inc., Murray H•ll, New Jersey
(Received December 3, 1951)
ß
Relationships between a listener's identification of a spoken vowel and its properties as revealed from
acoustic measurement of its sound wave have been a subject of study by many investigators. Both the
utterance and the identification of a vowel depend upon the language and dialectal backgrounds and the
vocal and auditory characteristics of the individuals concerned. The purpose of this paper is to discuss
some of the control methods that have been used in the evaluation of these effects in a vowel study program
at Bell Telephone Laboratories. The plan of the study, calibration of recording and measuring equipment,
and methods for checking the performance of both speakers and listeners are described.-The methods are
illustrated from results of tests involving some 76 speakers and 70 listeners.
INTRODUCTION
ONSIDERABLE variation is to be found in the processes of speech production because of their
complexity and because they depend upon the past
experience of the individual. As in much of human
behavior there is a self-correcting, or servomechanism
type of feedback involved as the speaker hears his own
voice and adjusts his articulatory mechanisms. 1
In the elementary case of a word containing a conso-
nant-vowel-consonant phoneme 2.3 structure, a speaker's
pronunciation of the vowel within the word will be
influenced by his particular dialectal background; and
his pronunciation of the vowel may differ both in
phonetic quality and in measurable characteristics from
that produced in the word by speakers with other
backgrounds. A listener, likewise, is influenced in his
identification of a sound by his past experience.
Variations are observed when a given individual
makes repeated utterances of the same phoneme. A
very significant property of these variations is that they
are not random in a statistical sense, but show trends
and sudden breaks or shifts in level, and other types of
nonrandom fluctuations. 4 Variations likewise appear in
the successive identifications by a listener of the same
utterance. It is probable that the identification of
repeated sounds is also nonrandom but there is little
direct evidence in this work to support such a con-
clusion.
A study of sustained vowels was undertaken to in-
vestigate in a general way the relation between the
vowel phoneme intended by a speaker and that identi-
fied by a listener, and to relate these in turn to acous-
tical measurements of the formant or energy concentra-
tion positions in the speech waves.
In the plan of the study certain methods and tech-
niques were employed which aided greatly in the
collection of significant data. These methods included
randomization of test material and repetitions tb ob-
1 Bernard $. Lee, J. Acoust. Soc. Am. 22, 824 (1950).
• B. Bloch, Language 24, 3 (1948).
a B. Bloch, Language 26, 88 (1950).
s R. K. Potter and J. C. Steinberg, J. Acoust. Soc. Am. 26,
807 (1950).
175
tain sequences of observations for the purpose of check-
ing the measurement procedures and the speaker and
listener consistency. The acoustic measurements were
made with the sound spectrograph; to minimize meas-
urement errors, a method was used for rapid calibration
of the recording and analyzing apparatus by means of
a complex test tone. Statistical techniques were applied
to the results of measurements, both of the calibrating
signals and of the vowel sounds.
These methods of measurement and analysis have
been found to be precise enough to resolve the effects
of different dialectal backgrounds and of the non-
random trends in speakers' utterances. Some aspects
of the vowel study will be presented in the following
paragraphs to illustrate the usefulness of the methods
employed.
EXPERIMENTAL PROCEDURES
The plan of the study is illustrated in Fig. 1. A list
of words (List 1) was presented to the speaker and his
utterances of the words were recorded with a mag-
netic tape recorder. The list contained ten monosyllabic
words each beginning with I-hi and ending with I-d]
and differing only in the vowel. The words used were
heed, hid, head, had, hod, hawed, hood, who'd, hud, anal
heard. The order of the words was randomized in each
list, and each speaker was asked to pronounce two
different lists. The purpose of randomizing the words in
the list was to avoid practice effects which would be
associated with an unvarying order.
If a given List 1, recorded by a speaker, were played
back to a listener and the listener wero asked to write
down what he heard on a second list (List 2), a com-
parison of List 1 and List 2 would reveal occasional
LEST
I , • SPEAKER I'---I TAPE I •
RECORDER
II I
MEASURING •-I
Fro, •. Recording and measuring arrangements for vowel study.
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. Broad band spectrograms and amplitude sections of the word list by a female speaker.
differences, or disagreements, between speaker and
listener. Instead of being played back to a listener,
List ! might be played into an acoustic measuring
device and the outputs classified according to the
measured properties of the sounds into a List 3. The
three lists will differ in some words depending upon the
characteristics of the speaker, the listener, and the
measuring device.
A total of 76 speakers, including 33 men, 28 women
and 15 children, each recorded two lists of 10 words,
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
making a total of 1520 recorded words. Two of the
speakers were born outside the United States and a few
others spoke a foreign language before learning English.
Most of the women and children grew up in the Middle
Atlantic speech area. 5 The male speakers represented a
much broader regional sampling of the United States;
the majority of them spoke General American. 5
The words were randomized and were presented to a
group of 70 listeners in a series of eight sessions. The
listening group contained only men and women, and
represented much the same dialectal distribution as
did the group of speakers, with the exception that a
few observers were included who had spoken a foreign
language throughout heir youth. Thirty-two of the 76
speakers were also among the 70 observers.
The 1520 words were also analyzed by means of the
sound spectrograph. 6,7
Representative spectrograms and sections of these
words by a male speaker are shown in Fig. 3 of the
paper by R. K. Potter and J. C. Steinberg; 4 a similar
list by a female speaker is shown here as Fig. 2. 8 In the
spectrograms, we see the initial [h• followed by the
vowel, and then by the final ['d•. There is generally a
part of the vowel following the influence of the [h• and
preceding the influence of the [d• during which a
practically steady state is reached. In this interval, a
section is made, as shown to the right of the spectro-
grams. The sections, portraying frequency on a hori-
zontal sca;le, and amplitude of the voiced harmonics on
the vertical side, have been measured with calibrated
Plexiglass templates to provide data about the funda-
mental and formant frequencies and relative formant
amplitudes of each of the 1520 recorded sounds.
LISTENING TESTS
The 1520 recorded words were presented to the group
of 70 adult observers over a high quality loud speaker
system in Arnold Auditorium at the Murray Hill
Laboratories. The general purpose of these tests was to
obtain an aural classification of each vowel to supple-
ment the speaker's classification. In presenting the
words to the observers, the procedure was to reproduce
at each of seven sessions, 200 words recorded by 10
speakers. At the eighth session, there remained five
men's and one child's recordings to be presented; to
these were added three women's and one child's record-
ings which had been given in previous essions, making
again a total of 200 words. The sound level at the ob-
servers' positions was approximately 70 db re 0.0002
dyne/cm 2, and varied over a range of about 3 db at the
different positions.
In selecting the speakers for each of the first seven
5 C. K. Thomas, Phonetics of American English, The Ronald
Press Company (New York, 1947).
6 Koenig, Dunn, and Lacy, J. Acoust. Soc. Am. 17, 19 (1946).
? L. G. Kersta, J. Acoust. Soc. Am. 20, 796 (1948).
s Key words for the vowel symbols are as follows: I-i'] heed,
l-z-1 hid, l-e-1 head, Ire-1 had, [-a• father, [o-1 ball, [rr-1 hood, [u-I
who'd, I%-1 hud, [3'] heard.
FREQUENCY OF SECOND FORMANT IN CYCLES PER SECOND
2500 2000 154)0 1000 500 100 20
2o •z
,
I'k
IIIIIIIIll I I
lOO •
•o•o
õ00 • •n
1000"
Fro. 3. Vowel loop with numbers of sounds unanimously classified
by listeners; each sound was presented 152 times.
sessions, 4 men, 4 women, and 2 children were chosen
at random from the respective groups of 33, 28, and 15.
The order of occurrence of the 200 words spoken by the
10 speakers for each session was randomized for pre-
sentation to the observers.
Each observer was given a pad containing 200 lines
having the 10 words on each line. He was asked to
draw a line through the one word in each line that he
heard. The observers' seating positions in the audi-
torium were chosen by a randomizing procedure, and
each observer took the same position for each of the
eight ses.sions, which were given on eight different days.
The randomizing of the speakers in the listening
sessions was designed to facilitate checks of learning
effects from one session to another. The randomizing
of words in each group of 200 was designed to minimize
successful guessing and the learning of a particular
speaker's dialect. The seating positions of the listeners
were randomized so that it would be possible to de-
termine whether position in the auditorium had an
effect on the identification of the sounds.
DISCUSSION OF LISTENING TEST RESULTS
The total of 1520 sounds heard by the observers con-
sisted of the 10 vowels, each presented 152 times. The
ease with which the observers classified the various
vowels varied greatly. Of the 152 ['i-] sounds, for in-
stance, 143 were unanimously classified by all observers
as Ei}. of the 152 sounds which the speakers intended
for ['(•}, on the other' hand, only 9 were unanimously
classified as Ea} by the whole jury.
These data are summarized in Fig. 3. This figure
shows the positions of the 10 vowels in a vowel loop in
which the frequency of the first formant is plotted
against the frequency of the second forma. nt 9 on mel
scales; •ø in this plot the origin is at the upper right.
The numbers beside each of the phonetic symbols are
the numbers of sounds, out of 152, which were unani-
mously classified as that particular vowel by the jury.
It is of interest in passing that in no case did the jury
agree unanimously that a sound was something other
than what the speaker intended. Figure 3 shows that
o R. K. Potter and G. E. Peterson, J. Acoust. Soc. Am. 20, 528
(1948).
10 S.S. Stevens and J. Volkman, Am. J. Psychol. 329 (July,
•94o).
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
[i], ['•r], [a•-], and ['u] are generally.quite well under-
stood.
To obtain the locations of the small areas shown in
Fig. 3, the vowels were repeated by a single speaker on
twelve different days. A line enclosing all twelve points
was drawn for each vowel; the differences in the shapes
of these areas probably have little significance.
When the vowels are plotted in the manner shown in
Fig. 3, they appear in essentially the same positions as
those shown in the tongue hump position diagrams
which phoneticians have employed for many years. n
The terms "high, front, low back" refer to the tongue
positions in the mouth. The ['i], for instance, is made
with the tongue hump high and forward, the [u-] with
the hump high and back, and the [a] and ['a•] with the
tongue hump low.
It is of interest that when observers disagreed with
speakers on the classification of a vowel, the two
classifications were nearly always in adjacent positions
of the vowel loop of Fig. 3. This is illustrated by the
data shown on Table I. This table shows how the ob-
servers classified the vowels, as compared with the
vowels intended by the speakers. For instance, on all
the 152 sounds intended as [i-] by the speakers, there
were 10,267 total votes by all observers that [hey were
['i'], 4 votes for ['f], 6 votes for ['e-], and 3 votes for ['o.'].
Of the 152 [-a'] sounds, there was a large fraction of the
sounds on which some of the observers voted for ['o-].
[•'] was taken for ['e-] a sizable percentage of the time,
and ['e-] was called either If] or ['a•-] (adjacent sounds
on the vowel loop shown in the preceding Fig. 3) quite
a large number of times. ['a-] and [o-], and [',t-] and
were also confused to a certain extent. Here again, as
in Fig. 2, the [i-], [•r-], ['•e-], and ['u-] show high intel-
ligibility scores.
It is of considerable interest that the substitutions
shown conform to present dialectal trends in American
speech rather well, • and in part, to the prevailing vowel
shifts observable over long periods of time in most
languages. •a The .common tendency is continually to
shift toward higher vowels in speech, which correspond
to smaller mouth openings.
The listener, on the other hand, wotfld tend to make
the opposite substitution. This effect is most simply
described in terms of the front vowels. If a speaker
produces ['f] for [e-I, for example [m•n-] for [men-] as
currently heard in some American dialects; then such
an individual when serving as a listener will be inclined
to write men when he hears ['m•n-]. Thus it is that in the
substitutions hown in Table I, [•-] most frequently
became ['e'], and [e-] most frequently became ['a•-]. The
explanation of the high intelligibility of ['a•-] is probably
based on this same pattern. It will be noted along the
n D. Jones, An Outline of English Phonetics (W. Heifer and Sons,
Ltd., Cambridge, England, 1947).
•2 G. W. Gray and C. M. Wise, The Bases of Speech (Harper
Brothers, New York, 1946), pp. 217-302.
•a L. Bloomfield, Language (Henry Holt and Company, New
York, 1933), pp. 369-391.
vowel loop that a wide gap appears between [a• and
[a-]. The [a'] of the Romance languages appears in this
region. Since that vowel was present in neither the lists
nor the dialects of most of the speakers and observers
the [•e-] was usually correctly identified.
The [i'] and the [u-] are the terminal or end positions
in the mouth and on the vowel loop toward which the
vowels are normally directed in the prevailing process
of pronunci.ation change. In the formation of ['i'] the
tongue is humped higher and farther forward than for
any other vowel; in [u-] the tongue hump takes the
highest posterior position in the mouth and the lips are
more rounded than for any other vowel. The vowels ['u-]
and [i-] are thus much more difficult o displace, and a
greater stability in the organic formation of these sounds
would probably be expected, which in turn should mean
that these sounds are recognized more consistently by a
listener.
The high intelligibility of ['•r-] probably results from
the retroflexion which is present to a marked degree
only in the formation of this vowel; that is, in addition
to the regular humping of the tongue, the edges of the
tongue are turned up against the gum ridge or the hard
palate. In the acoustical pattern the third formant is
markedly lower than for any other vowel. Thus in both
physiological nd acoustical phonetics the ['•r'] occupies
a singular position among the American vowels.
The very low scores on [•-] and [o-] in Fig. 3 un-
doubtedly result primarily from the fact that some
members of the speaking group and many members of
the listening group speak one of the forms of American
dialects in which [a'] and ['o-] are not differentiated.
When the individuals' votes on the sounds are an-
alyzed, marked differences are seen in the way they
classified the sounds. Not only did the total numbers of
agreements with the speakers vary, but the proportions
of agreements for the various vowels was significantly
different. Figure 4 will be used to illustrate this point.
If we plot total numbers of disagreements for all tests,
rather than agreements, the result is shown by the
upper chart. This shows that [•-], [e-I, [a-], [o-], and
had the most disagreements. An "average" observer
would be expected to have a distribution of disagree-
ments similar in proportions to this graph. The middle
graph illustrates the distribution of disagreements given
by observer number 06. His chief difficulty was in dis-
tinguishing between [a-] and ['o-]. This type of distribu-
tion is characteristic of several observers. Observer 013,
whose distribution of disagreements is plotted on the
bottom graph, shows a tendency to confuse [¾] and
more than the average.
The distributions of disagreements of all 70 observers
differ from each other, depending on their language
experience, but the differences are generally less ex-
treme than the two examples shown on Fig. 4. Thirty-
two of the 70 observers were also speakers. In cases
where an observer such as 06 was also a speaker, the
remainder of the jury generally had more disagreements
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
with his [o-] and [o• sounds than with the other sounds
he spoke. Thus it appears that if a speaker does not
differentiate clearly between a pair of sounds in speak-
ing them, he is unlikely to classify them properly when
he hears others speak them. His language experience, as
would be expected, influences both his speaking and his
hearing of sounds.
Since the listening group was not given a series of
training sessions for these tests, learning would be ex-
pected in the results of the tests. 14 Several pieces of
evidence indicate a certain amount of practice effect,
but the data are not such as to provide anything more
than a very approximate measure of its magnitude.
For one check on practice effect, a ninth test • was
given the jury, in which all the words having more than
10 disagreements in any of the preceding eight tests
were repeated. There was a total of about 175 such
words; to these were added 25 words which had no
disagreements, picked at random from the first eight
tests. On the ninth test, 67 words had more disagree-
ments, 109 had less disagreements, and 24 had the same
number of disagreements as in the preceding tests. 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 groups for the men,
women and children speakers, the largest differences in
numbers of disagreements for the original and repeated
tests was on the childrens' words, indicating a larger
practice or learning effect on their sounds. The indi-
cated learning effect on men's and women's speech was
nearly the same. When the data are classified according
to the vowel sound, the learning effect indicated by the
repetitions was least on ri-], I-x-I, and [u-I, and greatest
on
Another indication that there was a practice effect
lies in the sequence of total numbers of disagreements
by tests. From the second to the seventh test, the total
number of disagreements by all observers diminished
consistently from test to test, and the first test had con-
siderably more disagreements than the eighth, thus
strongly indicating a downward trend. With the speak-
ers randomized in their order of appearance in the eight
tests, each test would be expected to have approxi-
mately the same number of disagreements. The prob-
ability of getting the sequence of numbers of total dis-
agreements which was obtained would be somewhat less
than 0.05 if there were no learning trend or other non-
random effect.
It was also found that the listening position had an
effect upon the scores obtained. The observers were
arranged in 9 rows in the auditorium, and the listeners
in the back 4 rows had a significantly greater number of
disagreements with the speakers than did the listeners
in the first 5 rows. The effect of a listener's position
•4 H. Fletcher and R. H. Galt, J. Acoust. Soc. Am. 22, 93
0950).
6O0
400 [-
60
• 50-
uJ 40-
,,:I: 3O-
j 2..0-
0 10-
0
ALL OBSERVERS
I ,
(b) OBSERVER
I,I
3O
(C) OBSERVER 013
I ! , ,
i I I• • a 3 u u. ^ •
Fro. 4. Observer disagreements in listening tests.
within an auditorium upon intelligibility has been ob-
served previously and is reported in the literature. 1•
ACOUSTIC MEASUREMENTS
Calibrations of Equipment
A rapid calibrating technique was developed for
checking the over-all performance of the recording and
analyzing systems. This depended on the use of a test
tone which had an envelope spectrum that was essen-
tially flat with frequency over the voice band. The
circuit used to generate this test tone is shown sche-
matically in Fig. 5. It consists essentially of an overload-
ing amplifier and pulse sharpening circuit. The wave
shapes which may be observed at several different
points in the test tone generator are indicated in Fig. 5.
The test tone generator may be driven by an input
sine wave signal of any frequency between 50 and 2000
cycles. Figure 6(a) shows a section of the test tone with
a 100 cycle repetition frequency, which had been re-
corded on magnetic tape in place of the word lists by
the speaker, and then played back into the sourid
spectrograph. The departure from uniform frequency
response of the over-all systems is indicated by the
shape of the envelope enclosing the peaks of the 100
•* V. O. Knudsen and C. M. Harris, Acoustical Designing in
Architecture (John Wiley 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
250 V
FIG. 5. Schematic of calibrating tone generator.
w.E. 400 A
GERMANIUM
---- DIODE
cycle harmonics. With the 100 cycles from the Labora-
tories standard frequency oscillator as the drive signal,
the frequency calibration of the systems may be checked
very readily by comparison of the harmonic spacing on
the section with the template scale. The amplitude
scale in 6(a) is obtained by inserting a pure tone at the
spectrograph in 5 db increments. The frequency scale
for spectrograms may also be calibrated as shown in
Fig. 6(b). The horizontal ines here are representations
of the harmonics of the test tone when the test tone
generator is driven by a 500 cycle standard frequency.
These lines further afford a means of checking the
amount of speed irregularity or wow in the over-all
mechanical system. A calibration of the time scale may
be obtained by using the test tone generator with 100
cycle drive and making a broad band spectrogram as
shown in Fig. 6(c). The spacings between vertical stria-
tions in this case correspond to one-hundredth of a
second intervals.
In the process of recording some of the word lists,
it was arranged to substitute the calibrating test tone
circuit for the microphone circuit, and record a few
seconds of test tone between the lists of words. When the
word lists were analyzed with the spectrograph, the ac-
companying test tone sections provided a means of
checking the over-all frequency response of the recorder
and analyzer, and the frequency scale of the sectioner.
The effect of speed variations in either the recorder or
the sound spectrograph is to change the frequency scale.
A series of measurements with the 100 cycle test tone
showed that the tape recorder ran approximately one
percent slower when playing back than it did on
recording.
The speed variations on the sound spectrograph were
measured with the test tone applied directly, and the
maximum short time variations were found to be :t=0.3
percent. Such direct calibrations of the frequency scale
of the spectrograph, during a period of four weeks when
most of the spectrographic analysis was done, showed
maximum deviations of +30 cycles at the 31st harmonic
of the 100 cycle test tone. During that period a control
chart •6 of the measurements of the 3100 cycle compo-
nent of the test tone showed a downward trend of about
10 cycles, which was attributed to changes in the elec-
tonic circuit components of the spectrograph. As a
result of these calibration tests, it was concluded that
the frequency scale of the sound spectrograph could be
relied upon as being accurate within :t=1 percent.
Formant Measurements
Measurements of both the frequency and the ampli-
tude of the formants were made for the 20 words re-
corded by each of the 76 speakers. The frequency posi-
tion of each formant was obtained by estimating a
weighted average of the frequencies of the principal
components in the formant. (See reference 4 for a dis-
cussion of this procedure.) When the principal compo-
nents in the formant were symmetrically distributed
about a dominant component, such as the second
formant of ['A• hud in Fig. 2, there is little ambiguity
•6 "A.S.T.M. manual on presentation of data," Am. Soc. Test-
ing Materials (Philadelphia, 1945), Appendix B.
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
in choosing the formant frequency. When the distribu-
tion is asymmetrical, however, as in the first formant of
[-zr'] heard in Fig. 2, the difference between estimated
formant frequency and that assigned by the ear may be
appreciable.
One of the greatest difficulties in estimating formant
frequencies was encountered in those cases where the
fundamental frequency was high so that the formant
was poorly defined. These factors may account for some,
but certainly not all, of the differences discussed later
-:..:•-..:.: -.½ .., .- -. ..... ., .½ ,• , ,v. ,,' ,' .... ½ 1;.:::';•:'½;.;
'. "o.-:: ,"'-": • ', ;- .:'"::'. *;',: .... ":' ' •':" .' "'.'" ;;'7•:':;•"
_--' '. l: :w .... i ..... : -:';-'.x;. ' .. ... ;';;;':--"*½/-*".'?:•;•57
ß --.•, ß - --w-:-:
-" 4000 .... '; ..... ; .*.;
0 ' :- -"' '::;L ' .... ,•: .:-'-":-::-:'" -..:-,,' ...... '-: ':--'- ....
t-t-l:- : ................
ß • . . ... : ...... . ..................
.• ......... ..:: --.:-:-} _.;•.:;::::•:**; -
.• .... •. . ..... . . .:. .
...........
.• ...... ._ :.: .
• :- :.: : .........
• -:: _ .......
• -½•:.** . :•;..•; ........ -•,-• -.• _••• .•• ...;' --',• .•. • ..... ----
• .: ....... ,.-- ß ½ ........... ".½: . .....
.:•. 0 ,*-' -. ;•*•=•,• ' ' -•• :•'•--:" -- ' ': - .".' *•-• -.•,•
(b).. NAR.;:BOW BAND .SP:ECTROGRAM
,t.- . .. 0 F 500:-CYCLE CAL:!BRATiNG TONE.--
:% ...... ... -:.- ...... : ....
.._-•:- •:::..j;:
400
380
36O
//
Fro. 6. Spectrograms and section of calibrating tone.
181
//// / ß // ß
o-• //// ß .J //
• 340 //// ß / ß o© // k.) //
Z / ß / .//// -- 320- //
300 // /e /
ZO 280' .o' ß ///
• _ / ß /// R OF DIFFE'RENCES=I?.2
,,- 240 - '/ // ESTIMATED ½ OF
LL •/// // DIFFERENCES: 15.3 220
220 240 260 280 300 320 340 360 380 400 420
F 1 OF FIRST CALLING IN CYCLES PER SECOND
Fro. 7. Accuracy-precision chart of first formant frequencies
of [i] as•spoken by 28 women.
between vowel classification by ear and by measured
values of formant frequencies.
Amplitudes were obtained by assigning a value in
decibels to the formant peak. In the case of the ampli-
tude measurements it was then necessary to apply a
correction for the over-all frequency response of the
system.
The procedure of making duplicate recordings and
analyses of the ten words for each of the speakers
provided the basis for essential checks on the reliability
of the data.
One method by which the duplicate measured values
were used is illustrated by Fig. 7. This is a plot of the
values for the first formant frequency F, of [i] as in
heed, as spoken by the 28 female subjects. Each point
represents, for a single speaker, the value of F, measured
for the heed in the first list, versus the value of F, for the
heed in the second list. If the F, for the second list or
calling was greater than that for the first calling, the
point lies above a 45-degree line;if it is less, the point
lies below the 45-degree line. The average difference R
between the paired values of F, for first and second
callings, was 17.2 cycles. The estimated standard evia-
tion • derived from the differences between pairs of F,
values was 15.3 cycles. The dotted lines in Fig. 7 are
spaced +3 • cycles from the 45-degree line through the
origin. In case a point falls outside the dotted lines, it
is generally because of an erroneous measurement.
Each of the three formant frequencies for each of the
10 vowels was plotted.in this way. There were 760 such
points for each formant, or a total of 2280 points plotted
on 90 accuracy-precision charts like Fig. 7. Of these 2280
points, 118 fell outside the +3 • limits. On checking
back over the measurements, it was found that 88 of
the points were incorrect because of gross measure-
ment errors, typographical errors in transcribing the
data, or because the section had been made during the
influence period of the consonants instead of in the
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
o 2oo ,too eoo coo Iooo
FREQUENCY OF Ft IN CYCLES PER SECOND
Fro. 8. Frequency of second formant versus frequency of first
formant for ten vowels by 76 speakers.
steady state period of the vowel. When corrected, these
88 points were within the =1= 3 • limits. Of the remaining
30 points which were still outside the limits, 20 were the
result of the individuals' having produced pairs of
sounds which were unlike phonetically, as shown by
the results of the listening tests.
The duplicate measurements may also be used to
show that the difference between successive utterances
of the same sound by the same individual is much less
significant statistically than the difference between
utterances of the same sound by different individuals.
An analysis of variance of the data in Fig. 7 shows that
the differences between callings of pairs are not sig-
nificant. However, the value for the variance ratio when
comparing speakers is much larger than that corre-
sponding to a 0.1 percent probability. In other words,
if the measurements hown in Fig. 7 for all callings by
all speakers were assumed to constitute a body of
statistically random data, the probability of having a
variance ratio as high as that found when comparing
speakers would be less than one in a thousand. There-
fore it is assumed that the data are not statistically
random, but that there are statistically significant
differences between speakers. Since the measurements
for pairs of callings were so nearly alike, as contrasted
with the measurements on the same sound for different
speakers, this indicated that the precision of measure-
ments with the sound spectrograph was sufficient to
resolve satisfactorily the differences between the various
individuals' pronunciations of the same sounds.
RESULTS OF ACOUSTIC MEASUREMENTS
In Fig. 3, as discussed previously, are plotted areas
in the plane of the second formant F2 versus the first
formant F•. These areas enclose points for several
repetitions of the sustained vowels by one of the
writers. It is clear that here the vowels may be separated
readily, simply by plotting F2 against F•; that is, on
the F•--F• plane, points for each 'vowel lie in isolated
areas, with no overlapping of adjacent areas, even
though there exists the variation of the measured values
which we have discussed above.
The variation of the measured data for a group of
speakers is much larger than the variation encountered
in repetitions with the same speaker, however, as may
be shown by the data for F1 and F• for the 76 speakers.
In. Fig. 8 are plotted the points for the second calling by
each speaker, with the points identified according to the
speaker's word list. The closed loops for each vowel
have been drawn arbitrarily to enclose most of the
points; the more extreme and isolated points were dis-
regarded so that in general these loops include about
90 percent of the values. The frequency scales on this
and Fig. 9 are spaced according to the approximation
to an aural scale described by Koenig, which is linear to
1000 cps and logarithmic above?
Considerable overlapping of areas is indicated, par-
ticularly between E•r-] and Ee-], E•r-] and Ev-], Ev-] and
Eu-], and Ea-] and Eo-]. In the case of the E•r-] sound, it
may be easily distinguished from all the others if the
third formant frequency is used, as the position of the
third formant is very close in frequency to.that of the
second.
The data of Fig. 8 show that the distribution of
points in the F1--F•plane is continuous in going from
sound to sound; these distributions doubtless represent
TABLE I. Classifications of vowels by speakers and by listeners. Vowels as classified by listeners.
Vowels intended by speakers
i
10267
. 6
ß
ß
I 8 • o o u u A •
4 6 ... 3 ...........
954• 694 "5 1 1 .... 56
257 9014 949 1 3 ... i.i "5 51
1 300 9919 2 2 15 39
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
•7 W. Koenig, Bell Labs. Record 27, (August, 1949), pp. 299-301.
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
TABz• II. Averages of fundamental and formant frequencies and formant amplitudes of vowels by 76 speakers.
183
i i Ig It• ct o TJ u A 3'
Fundamental frequencies M 136 135 130 127 124 129 137 141 130 133 W 235 232 223 210 212 216 232 231 221 218 (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
F2
F3
Formant amplitudes (db)
M 2290 1990 1840 1720 1090 840 1020 870 1190 1350
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
W 3310 3070 2990 2850 2810 2710 2680 2670 2780 1960
Ch 3730 3600 3570 3320 3170 3180 3310 3260 3360 2160
L• -4 --3 --2 --1 --1 0 --1 --3 --1 --5
L2 --24 --23 --17 --12 --5 --7 --12 --19 --10 --15
La --28 --27 --24 --22 --28 --34 --34 --43 --27 --20
large differences in the way individuals speak the
sounds. The values for F3 and the relative amplitudes
of the formants also have correspondingly large varia-
tions between individuals. Part of the variations are
because of the differences between classes of speakers,
that is, men, women and children. In general, the chil-
dren's formants are highest in frequency, the women's
intermediate, and the men's formants are lowest in
frequency.
These differences may be observed in the averaged
formant frequencies given on Table II. The first for-
mants for the children are seen to be about half an
octave higher than those of the men, and the second
and third formants are also appreciably higher. The
measurements of amplitudes of the formants did not
show decided differences between classes of speakers,
and so have been averaged all together. The formant
amplitudes are all referred to the amplitude of the first
formant in [a•, when the total phonetic powers of the
vowels are corrected so as to be related to each other by
the ratios of powers given by Fletcher. •a
Various methods of correlating the results of the
listening tests with the formant measurements have
been studied. In terms of the first two formants the
nature of the relationship is illustrated in Fig. 9. In this
figure measurements for all vowels of both callings are
plotted in which all members of the listening group
agreed with the speaker. Since the values for the men
and the children generally lie at the two ends of the dis-
tributions for each vowel, the confusion between vowels
is well illustrated by their data; thus the measurements
for the women speakers have been omitted.
The lines on Fig. 9 are the same as the boundaries..
drawn in Fig. 8. As indicated previously, some vowels
received 100 percent agreement much more frequently
than others.
•a H. Fletcher, Speech and Hearing (D. Van Nostrand Company,
Inc., New York, 1929), p. 74.
The plot has also been simplified by the omission of
[3.•. The [3.• produces extensive overlap in the [u•
region in a graph involving only the first two formants.
As explained previously, however, the [3.-] may be
isolated from the other vowels readily by means of the
third formant.
When only vowels which received 100 percent recog-
nition are plotted, the scatter and overlap are some-
what reduced over that for all callings. The scatter is
greater, however, than might be expected.
If the first and second formant parameters measured
from these words well defined their phonetic values;
and if the listening tests were an exact means of classi-
fying the words, then the points for each vowel of
o •oo ,•oo eoo eoo •ooo •oo •,•oo
•nE•UE•½¾ oF F• iN CYCLES PEn SECOND
Fro. 9. Frequency of second formant versus frequency of first
formant for vowels spoken by men and children, which were
classified unanimously by 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 should be well separated. Words judged inter-
mediate in phonetic position should fall at intermediate
positions in such a plot. In other words, the distribu-
tions of measured formant values in these plots do not
correspond closely to the distributions of phonetic
values.
It is the present belief that the complex acoustical
patterns represented by the words are not adequately
represented by a single section, but require a more
complex portrayal. The initial and final influences often
shown in the bar movements of the spectrograms are of
importance here. •ø The evaluation of these changing
bar patterns of normal conversational speech is, of
course, a problem of major importance in the study of
the fundamental information bearing elements of speech.
A further study of the vowel formants is now nearing
completion. This study employs sustained vowels,
without influences, obtained and measured under con-
trolled conditions. The general objectives are to de-
termine further the most fundamental means of evalu-
ating the formants, and to obtain the relations among
the various formants for each of the vowels as produced
by difference speakers. When this information has been
obtained it is anticipated that it will serve as a basis for
determining methods of evaluating and relating the
changing formants within words as produced by various
speakers.
SUMMARY
The results of our work to date on the develop-
ment of methods for making acoustic and aural meas-
•9 Potter, Kopp, and Green, Visible Speech (D. Van Nostrand
Company, Inc., New York, 1947).
urements on vowel sounds may be summarized as
follows.
1. Calibration and measurement techniques have been de-
veloped with the sound spectrograph which make possible its
use in a detailed study of the variations that appear in a broad
sample of speech.
2. Repeated utterances, repeated measurements at various
stages in the vowel study, and randomization in test procedures
have made possible the application of powerful statistical methods
in the analysis of the data.
3. The data, when so analyzed, reveal that both the production
and the identification of vowel sounds by an individual depend
on his previous language experience.
4. It is also found that the production of vowel sounds by an
individual is not a random process, i.e., the values of the acoustic
measurements of the sounds are not distributed in random order.
This is probably true of many other processes involving indi-
viduals' subjective responses.
5. Finally, the data show that certain of the vowels are gener-
ally better understood than others, possibly because they repre-
sent "limit" positions of the articulatory mechanisms.
ACKNOWLEDGMENTS
The work which we have discussed has involved the
contributions of a number of people. We should like
to acknowledge the guidance of Mr. R. K. Potter and
Mr. J. C. Steinberg in the plan of the experiment, and
the contribution of Dr. W. A. Shewhart who has assisted
in the design and interpretation of the study with
respect to the application of statistical methods. We
are indebted to Miss M. C. Packer for assistance in
statistical analyses of the data. We wish to acknowledge
also the assistance given by Mr. Anthony Prestigia-
como, Mr. George Blake, and Miss E. T. Leddy in the
recording and analysis of the sounds and in the prepara-
tion of the data.