Fairbanks, Grant. Experimental Phonetics – T17

Auditory Detection of the Presence and Absence of Signals in Noise *1 **2

Grant Fairbanks, Speech Research Laboratory, University of Illinois, Urbana, Illinois
and
Arthur S. House †3 and Jay Melrose, ‡4 Control Systems Laboratory, University of Illinois, Urbana, Illinois
(Received March 23,1956)

An experiment is described in which detection of the absence of signal remained high and relatively
constant while detection of intermixed signals varied over the threshold region as a function of signal voltage.
Statistical signal detection thresholds of observers with set to detect absence of signal were approximately
2 db lower than those of observers with set to detect signal.

This paper reports an experiment concerned with
auditory detection of signals in noise in which the
variable of chief interest was detection of the absence
of signal. One purpose was to explore the variation of
such detection as signal detection varies around
threshold. A second purpose was to compare the effects
of the set to detect absence of signal with the opposite,
commonly used set to detect signal. 15

Procedure

The basic stimulus unit was a 3-sec interval of thermal
noise, essentially flat from 100 to 4000 cps, presented
throughout the experiment at approximately a 60-db
sensation level. One second after the onset of the
stimulus unit a 1500-cyde pure-frequency signal either
was (signal sample) or was not (no-signal sample)
presented for a period of 1 sec Each sample was
followed by a silent 3-sec judgment interval, so that
the stimuli cycled periodically at the rate of 10 per
min. Samples were presented in sets of 100, each consisting
of 50 signal and 50 no-signal samples, randomly
ordered, with both noise and signal levels constant
throughout a given set. Three different sets, one for
each of three signal levels, were administered to all
observers. Three orders of presentation of 100 stimuli
were derived from a table of random numbers and
rotated systematically among signal levels and observers.
Presentation was bilateral over FDR-10 earphones
with 1505 cushions.

By informal trial with this particular noise it was
found that a signal level of 0.007 v at the earphones
would yield high but not perfect signal detection, and
that additional steps, 2 and 4 db lower, would extend
the range downward suitably. These three levels will
be referred to as 0, -2, and -4 db, respectively.

Figure 1 shows the instrumentation. The signal
source was a 90-in. loop of magnetic tape driven continuously
counterclockwise at 15 in./sec over two
reproduce heads, as diagramed at the left, and
consisting of a 15-in. section of recorded signal spliced
to 75 in. of blind tape. Reproduction of the recorded
signal by Head 1 triggered an electronic time control
adjusted for a 3-sec interval, supplying the basic
stimulus unit of noise. One second after the onset of
noise the recorded signal passed over Head 2, located
15 in. from Head 1, and mixed with the noise when the
manual switch was closed. The switch was operated
during the judgment intervals, according to the
schedules mentioned, and was inaudible. The loop
moved at rated speed with no detectable flutter;
periodicity and low harmonic content of the reproduced
signal were verified oscillographically. The upper
portion of Fig. 1 shows the arrangement for recording
signal samples and responses on an ink-writing oscillograph.
Headsets and signal keys were located in a sound-treated
room, the remaining equipment in an adjacent
control room.

The observers were 36 young, male university
students with normal hearing bilaterally from 250 to
8000 cps, as determined by audiometric sweep check
at 5-db hearing loss. They were divided into two equal
subgroups and scheduled in pairs, one from each subgroup.
The two members of each pair were instructed
separately and did not communicate. After brief explanation
of the mechanics of the situation the positive
observer was instructed as follows: “Please listen
carefully to each sample and during the silent period
immediately following press the signal key if the noise
contained a tone” (negative observer: “… did not
contain a tone”). These instructions were presented
formally in writing, and read silently by the observer
as the experimenter read them aloud. Positive or
negative set was imparted entirely by the foregoing
words and “Please listen carefully” was the only
mention of effort. The observers were not informed
about the proportions of signal and no-signal samples
or about the success of their judgments until after the148

Table I. Mean number of responses.

experiment. All communication was deliberately factual
and minimal. After instruction the two observers were
brought together and seated facing away from each
other at a distance of about 6 ft. The signal keys could
be easily depressed with a finger, had a stroke of about
2 mm, and were noiseless. A brief familiarization routine
was then followed, consisting of a series of signal and
no-signal samples in which the signals were at first
prominent and then gradually reduced to levels approximating
those used in the experiment. Between
sets of stimuli short rest periods were scheduled, during
which the observers moved about freely without
adjusting earphones. Orders of the three sets of stimuli
and the three random orders for stimuli within set were
varied among the pairs of observers after the manner
of a systematic Greco-Latin square. Two headsets were
alternated between positive and negative members of
successive pairs of observers.

Results

In Table I, which presents the various distributions
of response, the entries are mean number. The general
nature of response for the total group of 36 observers is
shown in Fig. 2; where the means have been converted
to probabilities and plotted as a function of signal
voltage. As signal detection varied over the threshold
range, null detection (correct identification of the
intermixed no-signal samples) changed comparatively

Fig. 1. Block diagram of control apparatus.

little. Chi-square was significant beyond the 1% level
for signal detection, but not significant for null detection.
As a generalization, null-detection probability
approximated 0.8 when signal detection probability
was 0.5 (i.e., at the intersection of the signal-detection
and signal-miss curves in Fig. 2). Although signal and
no-signal samples were equally numerous, it will be
noted that no-signal responses were more numerous and
proportionately more often incorrect than signal responses
at all levels.

The foregoing statements, although probably not
far from accurate for observers with neutral set, do not
hold completely for the positive and negative subgroups,
as study of Table I and Fig. 3 will reveal. If an
observer tends to resolve uncertainty of judgment by
making no active response (when-in-doubt-do-nothing),
it would be predicted that the obtained probability of
detecting stimuli opposite to the set would be enhanced
by the uncertainties. Specifically, it would be expected
that the positive observer, who makes a no-signal
response by not making a signal response, would yield
higher null detection and lower signal detection than
the negative observer, if both signal and no-signal
samples give rise to uncertainties. Figure 3 shows that

Fig. 2. Mean probabilities for total group of observers.149

Fig. 3. Signal-detection and null-detection probabilities for
positive and negative subgroups of observers.

the expectations were realized. It is interesting to note
that the difference in instructions produced a difference
of about 2 db in the statistical threshold of signal
detection, and a corresponding difference in null
detection. The differences between subgroups are large,
approximately equal for both signal and no-signal
samples, reasonably constant at the three voltage levels,
and all are statistically significant. As shown in the
lower section of Table I, the subgroups did not differ
substantially in over-all correctness of response.

Smith and Wilson 26 imparted conservative and liberal
attitudes by instruction to two groups of observers, and
obtained lower signal detection and false-alarm incidence
from the conservative group for detection of
800-cyde signals in thermal noise. For 50% detection
the difference between groups was around 4 db. Attitudes
of the observers in the present study might be
described as moderate. The instructions have been
quoted previously and contrast sharply with the forceful
instructions used by Smith and Wilson. In the earlier
experiment all observers were positive with respect to
set; that is, they responded actively only to heard
signal. It seems likely that a positive observer with
conservative attitude would yield, in terms of Fig. 3,
a still lower signal-detection curve and a still higher
null-detection curve than those obtained, while the
curves of a liberal positive observer would be displaced
in opposite directions. It also seems reasonable to
conjecture that a conservative negative observer would
yield a higher signal-detection curve and a lower null-detection
curve, etc In other words, it would be
expected that the differences produced by set would be
widened by conservative attitude and narrowed by
liberal attitude. If this is so, it should be possible to
manipulate the statistical threshold of signal detection
by various combinations of set and attitude over a
range of approximately 6 db.150

1* This work was supported in part by contract DA-36-039-SC-56695
U.S. Trust. The authors are grateful to C. W. Sherwin of
the Control Systems Laboratory, University of Illinois, for his
counsel, and to H. V. Krone of the same Laboratory for technical
assistance.

2** Reprinted from The Journal of the Acoustical Society of America, Vol. 28, 1956, pp. 614-16.

3† Now at Acoustics Laboratory, Massachusetts Institute of
Technology, Cambridge, Massachusetts.

4‡ Now at University of North Dakota, Grand Forks, North
Dakota.

51 For purposes of discussion, terms will be used as follows:
signal detection and signal miss are correct and incorrect identifications,
respectively, of signal samples, stimulus intervals consisting
of signal plus noise. Null detection and false alarm are
correct and incorrect identifications, respectively, of no-signal
samples, stimulus intervals consisting of noise only. Positive and
negative observers attempted to identify signal and no-signal
samples, respectively.

62 M. Smith and E. A. Wilson, Psychol. Monogr. 67, Whole No
359, 1 (1953).