CTLF Corpus de textes linguistiques fondamentaux • IMPRIMER • RETOUR ÉCRAN
CTLF - Menu général - Textes

Fairbanks, Grant. Experimental Phonetics – T05

Effects of Time Compression
Upon the Comprehension
of Connected Speech *1

Grant Fairbanks
Newman Guttman
Murray S. Miron **2

This article reports an experimental
investigation of the relationship between
comprehension of the factual
details of extended spoken messages
and the rate at which they were heard.
The matters of greatest interest were
the variation of comprehension as a
function of rate and the efficiency of
comprehension, the amount per listening
time. In addition, the interactions
of rate, content difficulty, and listener
aptitude were explored.

Control of message time, or of
listening rate, was accomplished by
means of the device for automatic
time-frequency compression-expansion
described by Fairbanks, Everitt
and Jaeger (2). When used for time
compression as in the present experiment,
the device displays recorded
material in less than the original time
without essential alteration of the
other dimensions of the signal. In a
sample of speech, the time base is unselectively
and uniformly compressed
throughout with the rime proportions
of the signal units undisturbed. The
essentials of the phonetic units, vocal
pitch, stress, etc., are preserved. The
amount of compression, which may
conveniently be thought of as a percentage
of the original time, is variable
and a given message may be presented
to listeners at a variety of rates. In the
experiment, for example, a message
which was originally spoken at HI
words per minute was compressed by
selected amounts. The largest of these
was 70%, which yielded a listening
rate of 470 wpm.

Auditory comprehension and speaking
rate have been studied separately
many rimes, but investigations of their
interrelationships have not been numerous
and the data are not extensive.
No evidence of association was found
43by Miller (5), who studied classroom
lectures delivered at comparatively
slow rates. Harwood (4) and Nelson
(6), both of whom included moderately
fast rates among their conditions,
reported that the effect of rate was
discernible, but not significant. At fast
rates, Goldstein (3) obtained significant
reduction of comprehension.

In considering the results of these
experiments, however, it is important
to observe that the objective in each
instance was to study the effect of the
rate of the total speaker-listener situation,
and that the independent variable,
appropriately, was speaking rate.
It is well known that live variations
of speaking rate are accompanied by
correlated non-rate changes which are
large, systematic, and obviously important
in the complete, live speaker-listener
system. Examples of such accompaniments
are the deterioration of
articulation and the disproportionate
reduction of pause time which ensue
as speaking rate is increased. Variations
of this kind were not at issue
in the present investigation; it has been
a study of listening rate.

Procedure

Experimental Materials

Experimental Materials. The basic
stimulus materials were two technical
messages on the subject of meteorology.
One was concerned with
meteorological instruments and the
other with weather forecasting in support
of flying. They were written for
this special purpose and consisted of
straightforward expositions of factual
information, including descriptions of
instruments, definitions of concepts,
explanations of procedures, etc. Attempts
were made to present content
in which pre-knowledge would be
small, dependence upon background
information and skills (e.g., technical
terminology and mathematics) minimal,
and levels of abstraction low. The
messages were relatively long, 1554
and 1573 words, respectively. They
bore close general similarity to classroom
lectures, or to speeches such as
a technical expert might prepare for
presentation to a motivated lay audience.

The messages were read for original
uncompressed tape recordings by an
experienced, mature, male speaker
who was paced at a target rate of 140
wpm by light flashes at one-minute
intervals. This rate was selected on the
basis of the data available in the literature,
the results of a pilot experiment
with another message, and a number
of measurements made while the
speaker practiced. The obtained rate
was 141 wpm for each message, the
total times being 11.04 and 11.18
minutes, respectively. The overall impression
was that of a rate unquestionably
close to the central tendency
which would be expected for skilled
speakers reading to communicate the
content in question. The master recordings
were made on Magnecord
PT6 equipment at 15 ips, with an
Altec M11 microphone system.

Comprehension was sampled by
means of two tests, each consisting
of 30 five-alternative multiple-choice
items confined to the factual content.
An attempt was made to provide a
substantial range in difficulty among
the items. A further goal, for each test
considered individually, was to construct
it so that the mean score of
unselected subjects would approximate
chance without the message and a
point midway between chance and
perfect score immediately after having
heard the message, i.e., 20 and 60%
correct, respectively. The final tests
were achieved after a series of trial
presentations and revisions, which involved
44carrying six different recorded
messages and their tests through all
stages, and selecting the two best message-test
units. At the last preliminary
presentation, to 49 subjects, these two
units yielded mean scores of 59 and
62%, Kuder-Richardson reliability coefficients
of .77 and .81, and a pooled
reliability coefficient of .87. When the
tests were administered without the
messages to other subjects, the mean
scores were 23 and 18%.

An additional consideration in test
construction was provision for treatment
of message-test difficulty. After
various attempts to separate content
difficulty from item difficulty it was
decided, in view of the purposes of
the experiment, to manage the two factors
jointly and empirically in terms
of the increment in comprehension
produced by the message, or the
message effectiveness. It was proposed
to estimate the message effectiveness
of each content item, the particular
portion of the message corresponding
to a given test item, as the difference
between the percentage of correct responses
to the test item when message-test
and test-only conditions were administered
to independent groups of
unselected subjects. This method
seemed warranted since the intention
was not to deal with individual items,
but to provide a rough means of
stratification so that the tests could be
sub-scored. As will appear below, five
Message Effectiveness Levels were
specified in this manner, each with 12
items distributed through the two
message-test units.

Experimental Procedure

Experimental Procedure. The plan
involved assignment of independent
groups of subjects to five message-test
conditions corresponding to time compressions
of 0, 30, 50, 60 and 70%,
and an additional test-only, or 100%
compression, condition. All subjects
were young adult, male trainees at
Chanute Air Force Base not engaged
in weather training. Thirty-six were
assigned to each of the message-test
conditions. 44 to the test-only condition.
Each group was composed of
four equal sub-groups formed according
to Stanine levels on the Technical
Specialist Aptitude Index of the Airman
Classification Battery. Stanines
Five through Eight from the upper
part of the distribution, with aptitude
increasing in that order, were represented
by random selection from available
men for the various conditions.
The plan further involved determination
of the five Message Effectiveness
Levels from the 0 and 100% conditions,
as described above, and an analysis
of variance with the four 30-70%
conditions. The analysis was of
a mixed factorial type, Compression
by Stanine by MEL.

At each session which entailed message
presentation the procedures were
as follows: standardized, extemporaneous
instructions; a two-minute
sample of a similar message appropriately
compressed (considered to cover
adaptation to the particular rate); first
message; first test; second message;
second test. Message-test order was
counterbalanced for each group and
sub-group, and similar counterbalancing
was employed with the test-only
condition. Except where specified below,
the 60-item yield from the two
tests was pooled for each subject.
Testing rime was not limited.

Experimental sessions were conducted
in a typical classroom located
at Chanute Air Force Base. Ambient
noise was appreciable, but not objectionably
high or variable. Presentation
was bilateral over Permoflux PDR10
earphones with 1505 cushions.
Twenty-four such headsets were arranged
in a distribution system and45

Table 1. Mean number of items correct at the
various experimental conditions; 60 total items.

tableau compression | rate | mean

fed by the Magnecord equipment
through a matching transformer. The
level was comfortable for listening and
approximately equal by VU meter for
all conditions. The various compressed
versions of the messages were prepared
in advance by methods described
elsewhere (2) and it is necessary
to note here only that a discard
interval of 0.02 sec. was arbitrarily
selected. Since the process involves
periodic time sampling, the discard
interval, or the segment between successive
sampling intervals, should be
short in relation to significant units of
the signal. Although the optimum interval

image relative message effectiveness | relative message time | words/min

Figure 1. Variation of relative message effectiveness
with relative message time. Upper
scale: rate in words/minute.

for connected speech is not
known as yet, the value chosen was
sufficiently short to avoid impairment
of intelligibility by fragmentation of
words.

Results

General Effects of Time Compression

General Effects of Time Compression.
Table 1 presents the mean number
of items correct for the six conditions.
The scores at 0 and 100%
compression represent 63.8 and 20.7%,
respectively, of the 60 items, close to
the range mentioned above as a target
for the materials. This range, 25.9
in terms of items correct, is a useful
estimate of the effectiveness of the
message under the most favorable condition,
and is taken as 1.0 for purposes
of the ordinate in Figure 1, which
shows relative message effectiveness
as a function of relative message time.
The latter is the decimal fraction of
the original message time remaining
after compression. Corresponding values
of rate in wpm are shown along

image rate learned (items/min.) | rate presented (items/min.)

Figure 2. Variation of rate learned with
rate presented. A, experimental. B, items
correct unchanged by redaction of message
time. C, items correct proportional to message
time.46

the top edge of the figure. The curve
is seen to be typically sigmoid, with
major inflection between 30 and 50%
of the original message time. The message
was 50% effective when the
relative message time was approximately
40% of the original, rate being
higher than 350 wpm. At 50% compression,
with a rate of 282 wpm,
the reduction in message effectiveness
was small. In this condition the costs
in rime of the large amount learned
and the small amount not learned were
equal.

The negatively accelerated portion
of the curve of Figure 1 may be interpreted
as a reduction in message
efficiency, or in learning per rime.
This view led to the re-plotting of the
data shown in Figure 2, where the
coordinates are rate learned and rate
presented, both given in items per
minute. The ordinate will be recognized
as the reciprocal of the mean
amount of message rime per correct
item; the abscissa is expressed in terms
of the rate of item presentation, that
is, the rate at which the content corresponding
to the 60 test items was
presented. The points show ordinate
values for the five message-test conditions.

The slope of the broken line B is
that which would obtain if there were
no reduction in amount of learning as
message rate was increased, while C
would be the case if number of items
correct were found to be directly
proportional to message rime. The
empirical curve, A, is seen to be higher
than C over much of the range (the
30, 50 and 60% compression conditions),
and to be fairly close to B at
the left, where it reaches its maximum
at the 50% condition. Thus message
efficiency, as measured by the amount
of factual comprehension per stimulus
time, increased up to a message rate
of 282 wpm, at which rate the reduction
in absolute amount of comprehension
was not large (Figure 1).

It seems reasonable to suggest that,
for learning of this type, the original
message as recorded was relatively
inefficient. If the original message is
representative of expository lectures
on factual topics, as it is believed to
be both in content and delivery, then
the implications for training procedures
are direct, namely, that when
message rate is slowed beyond a certain
point the increment in learning
becomes relatively costly in rime, and
that the cost should be weighed in
terms of the objectives of the learning.

In the present experiment both messages
were displayed to the subjects
in the 50% condition in the mean
time used for one message in the uncompressed
condition, or 11.11 minutes.
In the 50% condition that
amount of time yielded an increment
of 34.8 - 12.4 = 22.4 in mean number
of items correct (Table 1); in the uncompressed
condition the same amount
of time yielded a mean of (38.3 -
12.4) /2 = 13.0 items correct. Thus,
if the standard of learning were set by
the amount effected by the two messages
in the 50% condition, then the
use of the same time to display one
message at one-half the rate would be
little more than one-half as efficient.
In short, the data appear to justify the
conclusion that acceleration of message
presentation beyond rates ordinarily
practiced effects a given amount
of learning of the type in question in
less time, or a greater amount of learning
in a given time.

Message Order and Serial Position

Message Order and Serial Position.
Since, as has been explained, the plan
of the experiment involved pooling
results from the two messages and
sub-scoring according to variations in
message effectiveness among the test
47items, the general effects of order of
messages and of serial position of
content within messages were studied.
The experiment was not designed to
investigate either effect; study of them
was for the purpose of supporting the
intended procedures. It will be recalled
that precautionary counterbalancing
of order was practiced.
When the two messages were scored
separately for the various conditions,
no evidence of order effect was found.
Special attention was given to the
large compression conditions with the
idea that experience with the ultra-rapid
rate during the first message
might have placed the second message
in a more favorable position. If such
a situation in fact obtains, the present
measures were not sensitive to it.

In writing the messages no attempt
was made to control the frequency
of occurrence of the portions
of the content to be sampled, and the
portions also were allowed to vary in

image n items correct | n items presented | % message time

Figure 3. Cumulative number of items correct
during message at various experimental
conditions, rates labelled in words/minute.
Broken line, reference curve for test-only
condition.

‘difficulty’ without regard to place of
occurrence within the messages. However,
the planned procedure of subscoring
the results according to variations
in response would have been
open to question had such variations
been significantly associated with
serial position. Study of this possibility
also gave negative results. Representative
findings are presented in Figure 3.
The abscissa is percentage of the total
message time for the condition in
question, the locations of content portions
within that rime being shown
along the top by item number. The
latter points were determined by
measuring the cumulative time for
presentation of the content measured
by the first six, the first 12, etc., items
in the two messages and computing
the mean. For example, the content
corresponding to the first six items
was presented in slightly less than
20% of the total message time.

It will be noted that the occurrence
of sampled content was not completely
periodic and that the density
was reduced in the last one-fifth of
the message particularly, where one-fourth
of the time was used to present
one-fifth of the content items. For
a given point on the abscissa, the ordinate
is the mean number of items correct
per message. Curves are plotted
for the various message rates; the test-only
curve has been added as a control
since neither message time nor position
within message was a factor. Given
periodic occurrence and equal difficulty
of items, a straight line from
zero to maximum would have been
expected in the absence of positional
effects. Deviations will be seen to be
attributable largely to the lower item
density in the last part of the message
and to the test-only variations. It was
concluded that serial position might
be disregarded since, as will be shown48

image percentage correct | words/min. | time compression ratio | easy items | difficult items | stanine

Figure 4. Mean percentage of items correct
for the various experimental conditions.
Stanine sub-groups five and eight; easy and
difficult items.

below, effects were small in comparison
to those of the groups of specific
items formed in sub-scoring. It is of
interest to observe the similarities of
the shapes of the five message-test
curves in Figure 3 to each other and
to the curve for the test-only condition;
interaction of compression and
serial position was clearly small.

Aptitude of Listeners and Message Effectiveness Level

Aptitude of Listeners and Message
Effectiveness Level
. Figure 4 shows
the general dimensions of the combined
effects of compression, stanine
and difficulty upon comprehension.
The ordinate is percentage of items
correct; compression increases to the
right along the abscissa, the compression
ratio being the percentage of
compression divided by 100. For purposes
of this illustrative graph, the 60
test items were arranged in descending
order on the basis of number of
correct responses in the uncompressed
condition. The highest ranking 12
items have been designated in Figure
4 as ‘Easy’ and the lowest 12 as ‘Difficult.’
The solid and broken lines
show mean scores for the Stanine
Eight and Five sub-groups. The outstanding
features are the large overall
effects of all three factors, and the
similarities in the shapes of the four
curves as inflected by time compression.

The mean scores expressed in percentage
correct for the five stanine
sub-groups at each of the six experimental
conditions are shown in Table
2. The progressions are as would be
expected from the pairs of stanine
curves in Figure 4, except for a few
minor reversals. The distribution of
items according to message effectiveness,
estimated as described above
from the results of the uncompressed
and test-only conditions, was entered
to yield five graduated sub-sets of 12
items each, referred to herein as Message

Table 2. Mean percentage correct for stanine sub-groups and total groups at the various experimental
conditions.

tableau % compression | stanine | total49

Table 3. Mean percentage correct at uncompressed
and test-only conditions and mean difference
between conditions for fire Message
Effectiveness Levels.

tableau MEL | % items correct | uncompressed | test-only | mean difference

Effectiveness Levels. Table 3 presents
the mean percentage correct at
each MEL for the two conditions
mentioned, which will be observed to
be negatively correlated, and the set
of mean differences, which range
roughly from 20 to 70%. In preparation
for the statistical analysis the
responses of each subject were subscored
for the five MELs, and each
sub-score was adjusted by subtracting
therefrom the appropriate mean score
from the test-only condition; this procedure
may be thought of as correcting
for pre-knowledge and test item
difficulty at each MEL. The means
of these adjusted scores expressed in
percentage correct are presented in
Table 4. It will be seen that the sets
of five means decrease progressively
in all message-test conditions, and that
the same is generally true from condition
to condition within MEL.

Table 4. Mean message effectiveness at various experimental conditions for five MELs; see text.

tableau % compression | MEL

Table 5. Summary of analysis of variance;
Compression, Stanine, MEL.

tableau df | ms | F | between subjects | compression (C) | stanine (S) | C x S | error (b) | within subjects | MEL (M) | M x C | M x S | M x C x S | error (W) | total

* <.001

The results of the analysis of variance
are summarized in Table 5. This
analysis estimated the effects of compression,
stanine and MEL, and their
interactions, employed the adjusted
response scores as measures, and was
confined to the 30, 50, 60 and 70%
message-test conditions. As will be
seen in Table 5, all three Fs for the
main effects are large and significant
beyond the 0.1% level; the evidence
for interaction is negative except for
MEL by compression.

The non-significance of interactions
involving stanine should be interpreted
with due regard to the restricted range
of levels in the present experiment.
Study of the means of Table 2 discloses50

image response score | words/min. | stanine | M.E.L. | time compression ratio

Figure 5. Mean response score (obtained
percentage correct minus test-only percentage)
for the various experimental conditions.
A, stanine sub-groups five and eight.
B, Message Effectiveness Levels 1 and 5.

that the differences between
stanine levels decrease with increasing
compression, although small differences
obtain, as would be expected,
even in the test-only condition. However,
if the test-only mean is subtracted
from the remaining means in
each column to explore the differential
effect of the message as if ‘pre-knowledge’
were equal, the result is as illustrated
in Figure 5, A, which shows
curves for the extreme stanine subgroups.
Figure 5, B illustrates the
necessity of specifying MEL in a
statement about the effects of large
time compressions.

Summary

A pair of independent message-test
units, each consisting of an extended
exposition of technical information
and a corresponding test of factual
comprehension, were developed. The
messages were read by an experienced
speaker at 141 wpm, recorded, and
compressed automatically in time by
various amounts. Independent groups
of subjects, ail Air Force trainees,
were assigned to five experimental
conditions which represented a series
of compressions ranging from 0 to
70%, and to a sixth test-only condition
in which no message was presented.
Listener aptitude was controlled
by forming equal sub-groups
for each condition at four different
levels. The effect of message-test difficulty
was assessed by sub-scoring the
results of the tests according to five
message effectiveness levels which
were based upon differences in response
to test items in the 0% compression
condition and the test-only
condition.

The curve of comprehension as a
function of message time was characteristically
sigmoid. Response was approximately
50% of maximum when
message time was 40% (60% compression,
353 wpm). When message
time was 50% (282 wpm), the response
was slightly less than 90% and
efficiency, response per time, was
maximal. Analysis of variance indicated
that time compression, listener
aptitude and message effectiveness all
affect factual comprehension significantly,
and afforded evidence that interaction
of time compression and
message effectiveness in the expected
direction is significant.

References

1. Brown, J. I. and Carlsen, G. R. Brown-Carlsen
Listening Comprehension Test,
Form AM
. Chicago: World Book Co.,
1953.51

2. Fairbanks, G., Everitt, W. L., and
Jaeger, R. P. Method for time or
frequency compression-expansion of
speech. Trans. I.R.E.-P.G.A., 1954, AU-2,
7-11.

3. Goldstein, H. Reading and listening
comprehension at various controlled
rates. Teach. Coll. Contr. Educ., 1940,
No. 821.

4. Harwood, K. A. Listenability and rate
of presentation. Speech Monogr., 22,
1955, 57-59.

5. Miller, E. C. Effect on learning of
variations in oral presentation. Ph.D.
dissertation, University of Denver, 1955.

6. Nelson, H. E. The effect of variation
of rate on the recall by radio listeners
of ‘straight’ newscasts. Speech Monogr.,
15, 1948, 173-180.52

1* Reprinted from the Journal of Speech and Hearing Disorders, Vol. 22, 1957, pp. 10-19.

2** Grant Fairbanks (Ph.D., State University
of Iowa, 1936) is Professor of Speech at
the University of Illinois. Newman Guttman
(PhD., University of Illinois, 1954) is a
member of the research staff of the Bell
Telephone Laboratories. Murray S. Miron
(M.A., University of Illinois, 1956) is a Research
Assistant in Speech at the University
of Illinois. This research was supported in
part, by the United States Air Force under
Contract No. AF 18(600)-1059, monitored
by the Training Aids Research Laboratory,
Air Force Personnel and Training Research
Center, Chanute Air Force Base, Illinois.
Permission is granted for reproduction,
translation, publication, use and disposal in
whole and in part by or for the United
States Government. Special acknowledgment
is made to Dr. A. A. Lumsdaine and
Dr. Arthur J. Hoehn of Training Aids Research
Laboratory for valuable assistance
and counsel.