Fairbanks, Grant. Experimental Phonetics – T07

Auditory Comprehension
in Relation to Listening Rate
and Selective Verbal Redundancy * 1

Grant Fairbanks
Newman Guttman
Murray S. Miron ** 2

In a previous investigation of the
effects of listening rate upon comprehension
of connected speech (3), two
long technical messages were read at
141 words per minute, recorded, and
compressed instrumentally in time (1)
by amounts which varied progressively
over a suitable range. Factual
comprehension, as measured by tests
after presentation, was found to vary
inversely and significantly with listening
rate, but not linearly. At 201 and
282 wpm, corresponding to time compressions
of 30 and 50%, respectively,
of the original speaking time, the
decrement in comprehension was small
in comparison to the reduction in
listening time. This suggested that the
time ‘saved’ might be used to add
reinforcing material to the compressed
message within the time of the original
message, and that some such arrangement
might produce an increase in
comprehension. Accordingly, a second
experiment (2) was performed to explore
one possibility for the use of
50% compression in this manner. Subjects
heard the complete message twice
in succession and the results were
positive.

Procedure

The general plan for the stimulus
materials is shown in Table 1. The
augmented version of the message
(long version) contained the complete,
unaltered, original message
(short version), and expanded both
the words and the time of the latter
by 43%. The message time of the
long version when compressed by
30% equalled that of the uncompressed
short version. From Table 1
it will be seen that the latter differed
from the uncompressed long version
in words and time, but not in rate;
that it differed from the compressed
short version in time and rate, but
not in words; that it differed from the
compressed long version in words
and rate, but not in time.

In preparing the long version it was
decided, as has been mentioned, to
augment the content corresponding to
30 of the 60 items in the combined
test of the two messages. These will be
termed experimental items; the remainder,
identical in both versions,
will be termed control items. Equal
numbers of both types were selected
from both messages. The two sets of
items were approximately matched by
pairing and both were representative
of the range of response. Each of the
five Message Effectiveness Levels, as
defined by the basic experiment (3),
contributed six items to each set. Data
from the uncompressed condition of
the basic experiment were sub-scored
for the 30 experimental and 30 control
items; the respective means were
18.03 and 20.25 items correct per subject.
Corresponding means for the
57test-only condition (test administered,
but no message presented) were 5.43
and 6.99, so that the respective differences,
estimates of message effectiveness,
were 12.60 and 13.26. In
order to compare the sets of items
with respect to degree of association
between test items and content, the
uncompressed message-test and test-only
conditions were intercorrelated
separately, using items as N and number
of correct responses as the
measure. The obtained coefficients
were low and approximately equal;
.23 and .22 for experimental and control
items, respectively.

An attempt was made to compare
the relative number of words devoted
to the experimental and control content
in the two versions. In general,
this involved identifying the content
which supported each item. Study of
the problem revealed that item density
was so high and relatively constant
that virtually none of the message
could be said to be unrelated to the
test. Since all items were factual, however,
the location of the specific content
corresponding to each could be
readily determined, and afforded a
basis for comparison of experimental
and control items on the assumption
that the remainder of the text supported
the two sets of items equally.
The specific content was defined as
the minimum continuous portion of
the text consisting of whole sentences
which, if isolated, would supply the
correct answer to the test item, subject
to the qualification that said portion
include the antecedent of any
pronoun. The mean lengths of these
specific content sections for the experimental
and control items were 29
and 26 words, respectively. Four pairs
of experimental and control items
overlapped by a total of 95 words.
For purposes of assessing the division
of the total messages between the two
sets of items, this overlap was equally
divided, as was the remainder of the
text outside the specific content sections.
The 3127 words of the two
basic messages thus were estimated to
be divided 1625 and 1502, 52 and
48%, between the experimental and
control content, respectively, or with
mean amounts of 54 and 50 words
per item. The subjective impressions
were that the latter values are good
approximations to the general length
of the content supporting comprehension
of the average item, and to the
difference between experimental and
control items.

In preparing the long version, the
augmentation of the short version by
43% amounted to adding 1358 words,
or 45 words per experimental item.
The total addition was satisfied exactly,
as were the different 43% additions
to the two messages required by
their slightly different lengths. No attempt
was made to equate the lengths
of the augmentations of specific experimental
items, which ranged from
25 to 64 words, although minor adjustments
were made to satisfy the
total requirements. The 4485 words
of the long version were estimated as
dividing 2754 and 1731, or 61 and
39%, between experimental and control
content. Thus, the long version
provided 92 and 58 mean words per
experimental and control item, respectively,
in comparison to the 54
and 50 words in the short version. For
these estimates the augmentations were
regarded as adding only to the experimental
content, except in the instances
of overlap mentioned above, where
equal division again was practiced.

Each of the 30 augmentations was
a section of content designed to be
inserted as a block into the short
58version and fit smoothly into the text.
The location of each was immediately
after the basic content section for the
item in question defined as described
above. Twenty-three augmentations
consisted of two sentences, six of
three, and one of one. The general
orientation during preparation of each
augmentation was to increase comprehension
as much as possible by a
brief addition which included at least
one paraphrase or restatement of the
content, but which in no case involved
complete repetition of the wording of
either content section or test item. It
is to be stressed that each augmentation
was viewed as a natural-sounding
portion of the long message which did
not signal its existence in any particular
way. The intention, in other
words, was to produce a resultant
which also could be regarded as a
reference message, in comparison to
which the original message was a
briefer version from which 30 redundant
blocks of content had been
excised.

The augmentations were prepared
in spoken form by having them recorded
separately and splicing them
into the original recordings. This procedure
was considered more likely to
provide good experimental control
than the other two possibilities, namely,
having the long version recorded
and either excising the augmentations
to produce a new original, or comparing
it to the independently recorded
original. With the complete augmented
text before him, the speaker
listened repeatedly to his original recording
before and after the point of
insertion, re-reading the original in
concert in an attempt to duplicate his
previously used rate, intonation, etc,
and then read the augmentation. The
recorded augmentation was immediately
cued in with the original, played,
monitored, and either accepted or rejected.
The 30 augmentations were
then spliced into the original and the
transitions judged in the continuous
message. Certain augmentations were
re-recorded, substituted and checked
again. Finally, minor adjustments of
the total durations were made to meet
the requirements mentioned above.
These involved only a few seconds
altogether and were readily distributed
through the pauses at points of insertion.

Given the long version in recorded
form, the procedure followed in all
respects that previously described (3).
All subjects were trainees at Chanute
Air Force Base and technical aptitude
(Stanine level) was equalized between
groups. Presentation of recordings was
over a headset distribution system of
excellent quality in a conventional
classroom. Each message was followed
immediately by its test, time for the
latter not being limited.

Two separate experiments were performed.
For reasons that will be explained,
the first experiment was not
considered to be unequivocal. A second
experiment was carried out under
improved conditions and its results
analyzed.

Results

First Experiment

Unfortunately for this economical
plan, the weather which prevailed on
the days of the experiment happened
to be unusually hot and humid, extremely
uncomfortable for the subjects
in contrast to the favorable conditions
for the first experiment.

Table 2 shows the means for the six
conditions, each sub-test being based
on 30 items. If the four sub-test means
within each of the three compression
conditions are examined, it will be
observed that the highest value is invariably
that for the experimental content
in the long version, as would be
expected. When the three pairs of
means for control content are compared,
the value is lower in the long
version in all cases, but it was felt
that this reduction might be a reflection
of the uncomfortable conditions
under which the long version was presented
and that it could not be attributed
confidently to the stimulus
materials. The impression that discomfort
might have confounded the
experiment was reinforced by the fact
that the experimenters had expected
a net increase in response to the long
version, since they considered that
the augmentations had improved the
original message.

Several other explanations were advanced,
among them the possibility
that augmentation had reduced the
message-test reliability. For the uncompressed
short version the Kuder-Richardson
reliability coefficient was
.87, as previously reported (3). For
the 0, 30 and 50% compressions of the
long version, coefficients derived from
the new data were .87, .82 and .90,
respectively, which seemed to be satisfactory.
Consideration was also given
to the possibility that relative item difficulty
had been upset by the augmentations.
Since no attempt had been
made to insure augmentations that
were equally effective, this would be
likely for the experimental items, but
whether or not it would also be true
of the control items was the issue. The
intercorrelation of response (subjects
correct) to the control items in the
short and long versions, essentially an
estimate of item reliability, was .81;
for the experimental items it was .73.
All in all, it seemed advisable to repeat
the experiment under conditions that
would permit more confident interpretation.

Second Experiment

The means are presented in the
left section of Table 3 and comparison
with the data of Table 2 will show
that the two experiments yielded results
that were essentially identical,
namely, small differences between
total means for the four stimulus conditions
and consistent interaction within
compression as described above.
Apparently the misgivings about the
effects of weather upon the first experiment
had not been warranted, but
the confirmation is reassuring.

An analysis of the version by type
of item interaction appeared to be an
appropriate test of the significance of
the observed internal variations, and it
was considered sufficient to handle the
two compression levels in parallel as
separate problems. The results are
summarized in Table 4. Both interaction
Fs are large and significant
beyond the 0.1 % level, and the F for
type of items is likewise significant
at both compressions. Attention was
then directed to the differences between
the specific sub-test means
within compression level shown in the
left section of Table 3, which were
tested by t. Since the two values for
type of items within a given version
were based upon the same 38 subjects,
the within-subjects errors of Table 4
were used for the experimental-control
differences; conventional ts were computed
for the four between-versions-within-items
differences, each derived
from two groups of 38 subjects
each.

Within the short version, the difference
between the experimental and
control content in favor of the latter,
obtained for both presentations, confirms
the difference already mentioned
above in the description of the procedure.
At 0% compression it was
found to be significant at the 1 % level,
but failed to meet the 5% requirement
at 30% compression. The reversed
and larger differences between
experimental and control content in
the long version are significant beyond
the 0.1% level at both values of compression.

For the experimental content the
increase from short to long version is
significant at the 1% level with uncompressed
presentation, at the 5%
level with 30% compression. The decrease
for the control content is
significant beyond the 0.1% level at
both compressions. In terms of number
of correct responses to items in
the long version in comparison to the
short version, 21 of the 30 experimental
items increased, while 23 of
the 30 control items decreased. Therefore,
it may be asserted with considerable
confidence that augmentation of
the experimental content sharply differentiated
it from the control content,
but that the differentiation took
the form of an absolute sum of
positive and negative changes, respectively,
of approximately equal
amounts.

Special interest attaches to the lower
relative response to the control content.
It has been noted above that the
experimental and control items were
not exactly matched, according to
scores in the basic experiment and in
estimated words per item in the text.
The data from the basic experiment
were re-examined and 15 carefully
matched pairs of experimental and
control items extracted, three from
each Message Effectiveness Level. The
respective mean scores were 2.97 and
2.91 items correct in the test-only
condition, 9.54 and 9.56 in the uncompressed
original condition, and the
mean numbers of words in specific
content sections of the text were 31.7
and 30.4. The comparable long version
was that administered to the 36 subjects
across four stanines in the first
experiment; there the scores were
10.67 and 8.03. In the same condition
the remaining 15 experimental and
control pairs, less exactly matched,
62with means of 8.49 and 10.69 for the
short version, yielded means of 10.86
and 9.66, respectively.

The results for the control content
were also studied for possible differential
effects related to relative message
effectiveness, or message-test
difficulty. The inter-correlation of
response to control items in the uncompressed
short and long versions
was repeated with the new data for
the second experiment; the coefficient
was .86. The scatter was inspected and
various sub-sets formed according to
scores in both versions with uniformly
negative results. For example, the 15
highest ranking control items in the
short version yielded a mean score per
subject of 12.45 in comparison to 7.92
for the lowest 15 items. The same two
sets of items incurred respective mean
reductions of 1.45 and 1.60 in the long
version. This unselectivity with respect
to message-test difficulty suggests
that the potent factor in reduction
of response was relative verbal
redundancy, a possibility that should
be studied in a future experiment.
Various other sub-sets of control items
were formed according to location
within message, length in words, proximity
to experimental items, etc., likewise
with no positive indication. The
depressed score for the control content
in the long version appears to
have been a fairly uniform effect
across the items.

In the long version the difference
between response to experimental and
control content was in the same direction
as the difference in quantity of
the content, as has been observed. The
design of the experiment assumed that
the experimental and control content
would be approximately equal in the
short version, and it was noted above
that the word count showed this division
to have been approximately
realized; that is, the estimate for experimental
content was 52%. If the
experimental content of the short
version had been exactly 50%, and if
all augmentation, 43% of the total
short version, had been added to it
exclusively, the experimental content
would have constituted 65% of the
long version. The word count showed
that 61 % is probably a closer estimate,
as also has been stated.

The mean scores were converted to
estimated message effectiveness by
subtraction of the means for the Stanine
Seven subjects in the test-only
condition of the basic experiment, following
previous procedures (3).
These means were 5.00, 7.27 and 12.27
for experimental, control and total
items, respectively. The converted
means are shown in the right section
of Table 3. Calculation will show that
the experimental component was 62 %
of the total message effectiveness in
the long version at both compressions.
Similarly, for the first experiment,
with data based on Stanines Five
through Eight, the corresponding
test-only means are 5.43, 6.99 and
12.42. For the long version at 0, 30
and 50% compression (Table 2) the
experimental components were 60, 59
and 61%, respectively, of the totals.
Reciprocally, prediction of the experimental
component as 61 % (the estimated
percentage of experimental
words) of the total message effectiveness
for the five conditions in the first
and second experiments yielded a
maximum difference of .47 item from
the obtained mean. Since five independent
groups were involved, the
compactness of the percentages and
predictions is provocative.

It is interesting that substantial,
selective increases in redundancy,
which increased response to the
selected portions of the message, were
63accompanied by decreased response to
other portions and by no change in
the total response. The present study
was not designed to explain this finding,
although a few speculations may
be useful. First, however, there is the
possibility that some unintended factor,
or combination of factors, in the
experimental materials and situation
placed a ceiling on response which
was approached by both the short and
the long versions of the message. Such
factors might reside in the verbal message,
acoustic message, instructions,
test, environmental conditions, length
of stimulus period in time or words,
characteristics of subjects, etc. This
explanation seems unlikely. In a long
and careful scrutiny of the experimental
details and results, no basis for
a ceiling effect could be discerned. On
the other hand, there is positive evidence
of significant sensitivity of response
to variations of the stimulus,
from the previous experiments which
used the short version and the same
test (2, 3), and also from the exploratory
work during preparation of
materials, in which sensitivity of the
test was one criterion.

If the experiment may be assumed
to have been satisfactory from the
above point of view, considerable interest
attaches to the fact that, although
the long version was not a
‘better’ message than the short version
in the experimental situation, it read
and sounded as if it were, as was
stated earlier. That impression was
distinct, was formed prior to the experiment,
and was confirmed by critics
other than the experimenters. The
suggestion is strong that it was based
largely upon improved ease of comprehension
of the facts in the redundant
experimental content. It may be,
then, that the redundancy of the experimental
content during display of
the long version invited a degree of
listener participation in the communication
act which was inappropriate to
the more terse control content, i.e.,
behavior analogous to an elevation of
threshold. Although standardized instructions
were used in all conditions,
it is worth noting that subjects were
allowed to seek their own levels of
effort. Cooperation was generally excellent,
but no special motivation was
attempted, and it is possible that the
long version commanded less effort
than the short version by virtue of the
stimulus itself. Along the same general
line of reasoning, it may be that
listeners come through experience to
associate redundancy with importance,
and that the relative redundancies of
the two sets of content items operated
as signals in the long version. A different
possibility is that the response
to the experimental portions of the
long version actively inhibited the response
to the control portions in a
conventional sense. In this regard it
should be observed that item density
was high in the message, one sampled
fact per 75 words, or about two facts
per minute with no compression, and
that the relationships between consecutive
blocks of content were typical
of those found in coherent messages.

The realism of the experimental
situation is believed to justify an observation
regarding application to live
situations. Although the findings do
not permit a sure explanation of the
mechanisms, they imply that the common
practice of singling out certain
points in a speech for redundant treatment
may be effective at such points,
but that its accompaniment may be
an absolute, as well as a relative reduction
of effectiveness at other
points.64

Summary

A recorded technical message, read
at 141 wpm, together with an accompanying
test of factual comprehension,
were available from a previous
experiment, which furnished data used
to construct a selectively augmented
version of the message. The 60 test
items were divided into matched pairs
according to message-test difficulty,
and the pairs split to yield two approximately
equal sets. For one set
the corresponding content of the message
(control content) remained unaltered.
For the other set the content
relating to each item (experimental
content) was augmented by a smoothly
continuous, non-repetitive restatement
of the sampled fact which was
inserted as a block into the basic text.
The augmented version was prepared
in recorded form by splicing recordings
of the augmentations, read at 141
wpm with suitable vocal continuity
by the original speaker, into a copy
of the original recorded message, so
that the resulting long version contained
the short version. By design,
the recorded augmentations, which
represented a total addition of 43%
to the short version in both number of
words and message time, constituted
30% of the words and time of the
long version. Both recorded versions
then were subjected to 30% automatic
time compression with rate increasing
to 201 wpm, for a total of
four recordings, two versions times
two compressions of 0 and 30%. In
comparison to the uncompressed short
version, the compressed long version
presented 43% more words in no
greater time.

The experiment consisted of presentation
of the four recordings to independent
groups of subjects, and analysis
of test results in relation to
version, compression, and type of
item. Because of uncertainty about the
environment at the time of first performance
the experiment was repeated
with improved control; the results of
the two experiments differed in no
essential way. Appropriate analyses of
variance were performed with data
from the second experiment.

No evidence of significant variation
in the total responses to the four
different recordings was found, but
the experimental and control components
of the response changed
significantly from version to version
within rate. Augmentation of the experimental
content resulted in a significant
increase in response which
was accompanied by an approximately
equal, unselective, significant decrease
in response to the control content.
Such internal changes obtained equally
at both messages rates, and their relative
amounts corresponded closely to
the changes in the proportions of the
number of words in the two sections
of the content.

References

1. 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.

2. Fairbanks, G., Guttman, N., and
Miron, M. S. Auditory comprehension
of repeated high-speed messages. JSHD,
22, 1957, 20-22.

3. Fairbanks, G., Guttman, N., and
Miron, M. S. Effects of time compression
upon the comprehension of connected
speech. JSHD, 22, 1957, 10-19.65