Chapter 3:
Frequency selectivity, masking and the critical
band
-
Introduction.
-
Frequency selectivity refers to the ability of the auditory
system to resolve sinusoidal components in a complex sound and plays a
role in many aspects of auditory perception.
-
Frequency selectivity is often demonstrated and measured
by studying masking.
-
Masking is the process by which the absolute threshold
for one sound (the signal) is raised by the presence of another sound (the
masker).
-
The usual measure of masking is the dB increase in threshold
produced by the masker (dB of masking = 10 log10(IM /
I), where IM is the threshold for the sound in the presence
of the masker, and I is the threshold in its absence).
-
It has been known for years that a signal is most easily
masked by a sound having frequency components close to those of the signal.
-
Led to the idea that our ability to separate the components
of a complex sound depends on the frequency-resolving power of the basilar
membrane.
-
Also led to the idea that masking reflects the limits
of frequency selectivity and provides a way to quantify it.
-
Time is also important to masking.
-
We shall initially consider simultaneous masking in
which the signal is presented at the same time as the masker.
-
Later we shall consider forward and backward masking
in which the signal is masked by a preceding or a following sound, respectively.
-
The critical band concept.
-
Fletcher�s band-widening experiment, the power spectrum
model and critical ratios.
-
Fletcher measured the threshold of a sinusoidal signal
as a function of the bandwidth of a noise masker.
-
The noise was always centered at the signal frequency
and the noise power density was held constant.
-
Thus, the total noise power increased with bandwidth.
-
Results of a similar study are shown in Fig. 3.1.
-
The threshold of the signal at first increases as bandwidth
(and total noise power) increases, but then flattens off.
-
That is, addition of more noise at a greater distance
from the signal frequency produces no additional masking.
-
To account for these results, Fletcher suggested that
the basilar membrane behaves as if it contains a bank of bandpass filters
with overlapping passbands. I.e., each point on the basilar membrane corresponds
to a filter with a different center frequency.
-
When trying to detect a signal in noise, the listener
is assumed to make use of a filter with a center frequency close to that
of the signal.
-
This filter passes the signal, but removes much of the
noise.
-
Only the components of the noise passing through the
filter have any effect in masking the signal.
-
Threshold is assumed to correspond to a certain signal-to-noise
ratio at the output of the auditory filter.
-
This set of assumptions has come to be known as the
power spectrum model of masking.
-
Stimuli are represented by their long-term power spectra.
-
Relative phases of the components and short-term fluctuations
in the masker are ignored.
-
Fletcher called the bandwidth at which the signal threshold
ceased to increase the critical bandwidth (CBW). To estimate it, Fletcher
made two simplifying assumptions.
-
The shape of the auditory filter is rectangular. He
knew this was not true, but it has little effect on the calculations.
-
When the noise just masks the tone, the power of the
tone, P, divided by the power of the noise inside the critical band is
a constant, K, ( P/(W x N0)=K, where W is the width of the critical
band and N0 is the noise power density).
-
Since P and N0 can be measured, if we had
an estimate of K, we could compute the CBW as W=P/(K x N0).
-
Fletcher estimated K to be 1.0 on the basis of the threshold
for the tone in a very narrow band of noise.
-
Using this estimate for K, W=P/N0.
-
More recent work has shown K to vary somewhat with frequency,
but to have a value of about 0.4.
-
P/N0 is now called the critical ratio and
is about 0.4 times the width of the critical band.
-
Since Fletcher described the critical band concept,
many different experiments have shown that listeners� responses to complex
sounds differ according to whether the stimuli are wider or narrower than
the critical band.
-
Each of these provides an independent way to estimate
the critical bandwidth and how it varies with center frequency.
-
The fact that these various experiments provide reasonably
similar estimates adds considerable strength to the critical band concept.
-
We now shall examine some of these other kinds of experiments.
-
The loudness of complex sounds.
-
For a complex sound of fixed intensity and bandwidth
W, if W is less than CBW, then the loudness of the sound is independent
of W and is about as loud as a pure tone of equal intensity at the center
frequency of the band.
-
If W is increased beyond the CBW, the loudness of the
complex begins to increase.
-
Typical results are shown in Fig. 3.2.
-
I do not think the loudness models appealed to by the
author of your text will help you understand this phenomenon very much.
-
The key to understanding why this phenomenon occurs,
however, is fairly simple.
-
The neural activity, and hence the loudness, within
a frequency channel (critical band) grows nonlinearly with increases in
intensity. E.g., a halving of intensity produces considerably less than
a halving of loudness.
-
The total loudness of a sound results from summing neural
activity across frequency channels.
-
When the energy in the stimulus is spread over more
than one channel, the loss in loudness in the center channel is thus less
than the gain in loudness in less central channels, so the overall loudness
increases.
-
This does not happen at low overall intensities, because
at low levels, loudness changes very rapidly with intensity and the loss
in loudness from the central channel is as great as the gain in loudness
in less central channels.
-
The threshold of complex sounds.
-
Early studies showed that the threshold (in terms of
total energy) of multi-tone complexes consisting of evenly spaced sinusoids
remains relatively constant as the number of sinusoids is increased, until
the overall spacing exceeds the CBW. Beyond this point the threshold increases.
-
This was interpreted as indicating that only the information
from a single critical band is used to detect the complex.
-
More recent studies have found no clear break at the
critical bandwidth and it has been shown that the ear is capable of integrating
information over more than one CBW.
-
Two-tone masking.
-
When subjects are asked to detect a narrow band of noise
in the presence of two tones with frequencies on either side of its center
frequency, threshold remains constant as the separation between the tones
is increased up to the CBW.
-
As spacing increases beyond the CBW, threshold for the
noise falls sharply.
-
Interpretation of this result is complicated by the
fact that the lower tone may interact with the noise band to produce combination
products which may be detected even if the signal is inaudible.
-
When measures are taken to mask out such combination
products, the threshold for the noise increases smoothly with increasing
frequency separation between the tones.
-
This result, like more recent results for the threshold
of complex sounds, emphasize that the auditory filter is not rectangular,
but has a rounded top and sloping sides.
-
Sensitivity to phase.
-
The phenomenon described in this section involves complex
acoustics that go beyond the scope of this course.
-
The results turn out not to be applicable to estimating
the CBW.
-
Therefore, we shall not consider this section.
-
The discrimination of partials in complex tones.
-
Subjects� abilities to hear pitches corresponding to
the individual sinusoidal components in a complex harmonic sound can be
measured.
-
One way to do this is to present two comparison tones,
one corresponding to a partial in the complex and the other one half way
between that partial and the one above or below it. The subject is then
asked which comparison was part of the complex.
-
Subjects can do this with 75% accuracy when a partial
is separated from its neighboring partials by about 1.25 times the equivalent
rectangular bandwidth of the auditory filter.
-
There may be more involved here than peripheral filtering
because musicians do better at this task than nonmusicians, even though
musicians do not show narrower CBWs in masking experiments.
-
The musicians appear to have learned to make more efficient
use of the peripheral filtering mechanism.
-
Interim summary.
-
A variety of different experiments give consistent estimates
of the CBW.
-
Most do not show a distinct breakpoint corresponding
to the CBW because the auditory filter is not rectangular.
-
We now turn to attempts to measure the shape of the
auditory filter.
-
Estimating the shape of the auditory filter.
-
Psychophysical tuning curves (PTCs).
-
Analogous in many ways to neural tuning curves.
-
The signal (usually a sinusoid) is fixed at a low level
to confine the activity it produces to a single auditory filter.
-
For each of several frequencies of the masker (usually
a narrowband noise), the level needed to just mask the signal is determined.
-
Typical results are shown in Fig. 3.7.
-
Under the assumptions of the power spectrum model of
masking, the PTC shows the level of the masker required to produce a constant
output from the auditory filter.
-
Assuming the auditory filter to be linear, its shape
is given by simply inverting the PTC.
-
One problem with this way of determining the shape of
the auditory filter is that subjects improve their performance by using
a filter whose center is offset from the frequency of the signal, in a
direction away from the masker (off-frequency listening).
-
This produces a sharper tip on the PTC than if only
a single filter were used.
-
The notched-noise method.
-
Patterson devised a method of determining auditory filter
shape that prevents off frequency listening.
-
The signal is fixed in frequency and the masker is a
noise with a notch centered at the signal frequency.
-
The width of the notch is varied and the threshold of
the signal is measured as a function of notch width, as shown in Fig. 3.8.
-
Because the notch is symmetric about the signal, the
method is not capable of showing asymmetries in filter shape, but PTCs
indicate that at least the top part of the filters are reasonably symmetrical.
-
Off frequency listening is prevented because optimal
signal-to-noise ratio is achieved with a filter centered at the signal
frequency.
-
The amount of noise passing through the filter is proportional
to the area under the filter in the frequency ranges covered by the noise.
-
Assuming that threshold corresponds to a constant signal-to-masker,
the change in threshold with notch width shows how the area under the filter
varies with D f.
-
The first derivative of the function relating threshold
to D f gives the relative
height of the filter at each value of D
f.
-
In other word, the relative response of the filter for
a given deviation, D
f, from the center frequency is equal to the slope of the function relating
signal threshold to notch width at that value of D
f.
-
A typical auditory filter derived using this method
is shown in Fig. 3.9.
-
The filter has a round top and steep skirts.
-
A single number like CBW can not completely describe
such a shape, but summary measures such as the 3-dB bandwidth (typically
10-15% of center frequency) or equivalent rectangular bandwidth (typically
11-17% of center frequency) are used.
-
Fig. 3.10 shows equivalent rectangular bandwidths of
auditory filters, as a function of frequency, from a number of experiments
using the notched noise method.
-
Patterson�s method has been extended to conditions in
which the notch in the noise is placed asymmetrically about the signal
to allow measurement of asymmetry in the auditory filter.
-
Off frequency listening then has to be taken into account.
-
The mathematics for analyzing this case are beyond the
scope of this course.
-
Results show that the auditory filter is reasonably
symmetrical at moderate sound levels, but becomes increasingly asymmetric
at high levels, the low-frequency side becoming shallower than the high-frequency
side, as shown in Fig. 3.11.
-
Some general observations on auditory filters.
-
Although people can attend to more than one filter at
a time, one can predict whether a complex sound (e.g., an alarm) will be
detected by computing the signal-to-noise ratio for an auditory filter
at its most prominent frequency component. If that ratio is no less than
�4 dB, the signal will be detected.
-
It is important to remember that auditory filters are
continuous, not placed end to end.
-
Masking patterns and excitation patterns.
-
Many early studies of masking held the frequency of
the masker constant and varied the frequency of the signal.
-
This is not a good way to estimate the shape of the
auditory filter because, under the assumptions of the power spectrum model
of masking, subjects would use a different auditory filter for each different
signal frequency.
-
The resulting graph plotting masked threshold as a function
of the frequency of the signal is called a masking pattern or a masked
audiogram and an example is shown in Fig. 3.12.
-
Such masked audiograms show steep slopes on the low-frequency
side.
-
The high-frequency side is less steep and becomes shallower
with increasing level of the masker.
-
Since the auditory filter is shifted as the signal frequency
is altered, these masked audiograms can be used to obtain a crude estimate
of the excitation pattern produced by the masker like the one you saw in
Chapter 2 (Fig. 2.9).
-
Since the signal is detected when the excitation it
produces is some constant proportion of the excitation produced by the
masker in the frequency region of the signal, the masked audiogram should
be directly proportional to the excitation pattern of the masker.
-
This expectation can be complicated by such factors
as off-frequency listening and the possible use of combination tones produced
by the interaction of signal and masker to detect the signal.
-
The shapes of excitation patterns for the masker can
be derived using the concept of the auditory filter, as illustrated in
Fig. 3.13.
-
The excitation pattern of a sound can be thought of
as the output of the auditory filters as a function of their center frequencies.
-
The masked threshold for the signal at a given frequency
is an estimate of the amount of masker excitation at the output of the
auditory filter centered at that frequency.
-
Thus plotting these thresholds against signal frequency
gives a crude estimate of the excitation pattern of the masker.
-
Although the auditory filters were assumed to be symmetrical
in this derivation, the derived excitatory pattern is asymmetrical. This
happens because the bandwidth of auditory filters increase with their center
frequency.
-
This method can be extended to asymmetric filter shapes,
but whereas the lower side of the auditory filter gets less steep with
increasing level, the upper side of the masked audiogram (and the excitatory
pattern) gets less steep with increasing level.
-
The latter happens because the upper side of the excitatory
pattern is determined by the lower side of the filter and vice versa.
-
The nature of the critical band and mechanisms of masking.
-
The origin of the auditory filter.
-
The physiological basis of the critical band is not
completely understood, but it certainly involves the frequency-resolving
power of the basilar membrane.
-
The critical bandwidth or the equivalent rectangular
bandwidth of the auditory filter corresponds to a constant distance along
the basilar membrane, regardless of center frequency (about 0.9 mm).
-
The equivalent rectangular bandwidth of auditory filters
measured behaviorally in animals corresponds well with the equivalent rectangular
bandwidth of tuning curves measured in single auditory neurons in the same
species.
-
Temporal effects.
-
It is possible that some further sharpening of tuning
curves takes place after the basilar membrane due to lateral inhibition
at the level of the hair cells.
-
Data on this point are inconclusive and we shall not
consider them until later in this chapter.
-
The mechanism of masking.
-
There are two common conceptions of the mechanism by
which masking occurs.
-
The first is that masking involves swamping of the neural
activity evoked by the signal.
-
If the masker produces significant neural activity in
the auditory filter being used to detect the signal, the increment in activity
produced by the signal may be insufficient to be detected.
-
On this account masking occurs only if the masker produces
excitation in the auditory filter which would otherwise respond to the
signal.
-
A second possibility is that the masker suppresses or
inhibits activity which the signal would evoke if presented alone.
-
There is currently no clear way to distinguish between
these two possibilities.
-
Psychologists have preferred the swamping explanation
because suppression is a nonlinear process and masking is well modeled
by a system of linear filters.
-
The neural code used for signal detection.
-
What aspect of neural activity evoked by the signal
might be used for detection?
-
Similar to the question of how stimulus intensity is
neurally encoded which we considered in Chapter 2.
-
Neural firing rates is certainly part of the answer.
-
Temporal patterns of neural firing (phase locking) may
also be relevant.
-
There may be little or no phase locking to weak components
which are close in frequency to stronger ones.
-
In the case of noise, neurons fire irregularly. If a
tone is added, some neurons may phase lock to it; and hence, fire more
regularly.
-
In either case, a signal may fail to be detected if
the subject cannot detect its effect on the time pattern of neural firings.
-
Addition of a tone to noise also reduces fluctuations
in the amplitude envelope of the total sound and this may be a cue in detecting
the tone.
-
Comodulation masking release: Spectro-temporal pattern
analysis in hearing.
-
Initial demonstrations of comodulation masking release.
-
The power spectrum model of masking assumes that performance
is based on the output of the single auditory filter which gives the highest
signal-to-noise ratio and that threshold corresponds to a constant signal-to-masker
ratio.
-
This model works well in many situations.
-
In other situations it is clear that listeners make
comparisons across auditory filters and that short-term temporal fluctuations
of the masker can have important effects.
-
Hall et al. (1984) demonstrated that across-filter comparisons
could enhance detection of a sinusoidal signal in a fluctuating noise masker
if the fluctuations were coherent or correlated across frequency bands.
-
The threshold for a 1-kHz, 400-ms pure tone was measured
as a function of the bandwidth of a noise masker with constant spectrum
level (noise density).
-
The noise was centered at 1 kHz. and was either random
or comodulated.
-
A random noise has irregular fluctuations in amplitude
that are independent in different frequency regions.
-
The comodulated noise was amplitude modulated at a low,
irregular rate so that amplitude fluctuations were the same in different
frequency regions.
-
Results of this experiment are shown in Fig. 3.16.
-
For the random noise, threshold increases with bandwidth
up to the critical bandwidth of the auditory filter and then levels off,
just as expected from the power spectrum model of masking.
-
For the modulated noise, threshold increases with bandwidth
up to about 100 Hz and then decreases with increasing bandwidth.
-
The latter result indicates that subjects can compare
the outputs of different auditory filters to reduce masking.
-
Another way to demonstrate comodulation masking release
is by using narrow bands of noise which inherently have slow amplitude
fluctuations.
-
One band of noise (on-frequency band) is centered at
the signal frequency and the other (flanking band) is placed outside the
critical band around the signal.
-
If amplitude fluctuations in the two bands are uncorrelated,
addition of the flanking band either worsens or does not effect detection
of the signal.
-
If amplitude fluctuations in the two bands are correlated,
addition of the flanking band improves detection of the signal.
-
This comodulation masking release occurs even if the
signal and on-frequency band are presented to one ear and the flanking
band is presented to the other.
-
This last result indicates that the brain is somehow
able to compare the amplitude fluctuations in the two bands and use this
information to improve detection of the signal.
-
The role of within-channel cues.
-
Modulation of a masker can produce a release from masking
even when the masker�s bandwidth is less than the auditory filter bandwidth.
-
This release cannot arise from comparisons of the outputs
of different auditory filters and does not represent comodulation release
from masking.
-
This results from cues available in the output of a
single auditory filter.
-
One example of such a within-channel cue is the decrease
in envelope fluctuations when a sinusoidal signal is added to a noise.
-
Within-channel cues can lead to an overestimate of comodulation
release from masking, but are easily avoided in several ways.
-
Use of brief signals.
-
Widely separating the on-frequency and flanking bands.
-
Presenting the two bands of noise to opposite ears.
-
Factors influencing the magnitude of comodulation release
from masking.
-
Occurs for a wide range of frequencies (500 to 4000
Hz) and does not vary much with signal frequency.
-
Is largest when the masker is modulated at a low rat
and covers a large frequency range.
-
When within-channel cues are eliminated, comodulation
release from masking can be as large as 11 dB.
-
When flanking band is presented to opposite ear, comodulation
release varies little with center frequency, but is larger for flanking
bands close to the signal frequency. Suggests a proximity effect separate
from the effect of within-channel cues.
-
Comodulation release tends to increase as the width
of on-frequency and flanking bands of noise is decreased, probably because
the rate of envelope fluctuations decreases.
-
If many flanking bands of noise are used, comodulation
release can be as large as 16 dB.
-
Models to explain comodulation release from masking.
-
Models fall into two general classes.
-
One possibility is that the auditory system compares
envelope modulation patterns at the outputs of auditory filters with different
center frequencies and detects disparities.
-
For a comodulated masker without a signal, the modulation
pattern would be similar for all active filters.
-
Addition of a signal would modify the modulation pattern
at the output of a filter tuned to the signal.
-
If the masker were not comodulated there would be little
similarity at the output of different auditory filters in the first place.
-
Another possibility is the dip-listening model in which
the listener monitors fluctuations in auditory filters away from the signal
frequency to determine the optimum times to listen for the signal. If the
noise is comodulated, an off-frequency minimum is predictive of an optimum
signal-to-noise ratio at the signal frequency.
-
Experimental tests of the models.
-
Richards (1987) tested the idea that we can detect disparities
in envelope modulation patterns at the outputs of different auditory filters
by asking subjects to discriminate two comodulated bands of noise from
two bands with independent envelopes.
-
Subjects were successful in this task, indicating that
across-filter disparities in modulation pattern can be detected.
-
Others have shown that this detection improves with
increasing bandwidth, which is opposite to what is found in comodulation
release experiments and suggests different mechanisms are involved.
-
A number of experiments attempting to eliminate either
across-filter disparities in envelope fluctuations, across-filter disparities
in overall level, or dip listening lead to the conclusion that comodulation
release from masking does not depend on a single cue or mechanism.
-
It appears to reflect the operation of flexible mechanisms
which can exploit a variety of cues depending on the specific stimuli used
and is clearly of considerable value in many real-life situations.
-
Profile analysis.
-
Even without distinct envelope fluctuations, subjects
are able to compare the outputs of different auditory filters to enhance
signal detection.
-
Green et al. have shown that subjects can detect changes
as small at 1-2 dB in the relative level of a sinusoidal signal imbedded
in a complex background with a varying overall level.
-
Such results suggest subjects can detect a change in
the shape or profile of the spectrum of the sound.
-
This should not be surprising since it has been known
for years that one of the main factors determining the timbre of a sound
is its spectral shape.
-
We discriminate various vowel sounds from each other
on the basis of their spectral shape, regardless of the level of those
sounds.
-
We identify the sound of various musical instruments
largely on the basis of the spectral shape of their sound whether they
are playing loudly or softly.
-
Non-simultaneous masking.
-
Two basic types of non-simultaneous masking.
-
Backward or pre-stimulatory masking in which the signal
precedes the masker.
-
This phenomenon is poorly understood and seems not to
occur in highly practiced subjects.
-
Could reflect interference of the masker with later
processing of the signal at some higher level of the nervous system.
-
May simply reflect some confusion of the signal with
the masker in unpracticed subjects.
-
Forward or post-stimulatory masking in which the masker
precedes the signal.
-
Must be separated from adaptation or fatigue produced
by the masker.
-
Such separation is possible because forward masking
occurs for relatively short masker presentations (usually a few hundred
msec) and is limited to signals occurring within about 200 msec of the
masker.
-
Main properties of forward masking.
-
Forward masking is greater the nearer in time to the
masker the signal occurs, as shown in Fig. 3.17.
-
The rate of recovery from forward masking is greater
for higher masker levels and always decays to zero within 100 to 200 msec.
-
Increments in masker level do not produce equal increments
in amount of forward masking as they do in simultaneous masking where threshold
corresponds to a constant signal-to-noise ratio. This is shown in the right
panel of Fig. 3.17.
-
Amount of forward masking increases with masker duration
up to at least 20 msec, with some studies finding an increase up to 200
msec.
-
Forward masking is influenced by the relation between
the frequencies of the signal and the masker, just as with simultaneous
masking.
-
Evidence for lateral suppression from non-simultaneous
masking.
-
A point of some controversy and considerable neurophysiological
complexity is the existence and nature of lateral suppression in the auditory
system.
-
Although your text provides an extended discussion of
this topic, it goes well beyond the scope of this course and we shall only
make a few basic points in this regard.
-
It may seem surprising that simultaneous masking is
well modeled by linear filters since the behavior of individual auditory
neurons is grossly nonlinear and involves lateral suppression processes
by which strong activity at one center frequency can inhibit or weaken
activity at adjacent center frequencies.
-
One would not expect to see effects of lateral suppression
in simultaneous masking because the masker and the signal are processed
simultaneously in the same channel (auditory filter).
-
If suppression does occur, one might expect to see it
in forward masking provided:
-
The neural level at which suppression occurs is not
later than the level at which most of the forward masking effect arises.
-
Suppression built up by the masker has decayed by the
time that the signal is presented.
-
Psychophysical tuning curves from forward masking are
steeper on the high frequency side than those from simultaneous masking
as shown in Fig. 3.20.
-
Several other methods of estimating frequency selectivity
also show sharper tuning in forward than in simultaneous masking.
-
At least one explanation of this difference is that
the low frequency side of the excitatory pattern for the masker is sharpened
by a suppression process.
-
Alternative explanations of this difference also involve
suppression processes.
-
It is thus likely that non-simultaneous masking results
show effects of a suppression mechanism operating close to the point of
transduction of mechanical movements to neural impulses (hair cells or
the basilar membrane).
-
Frequency selectivity in impaired hearing.
-
A variety of different types of studies show that frequency
selectivity is impaired (the critical band is widened) by damage to the
cochlea.
-
Fig. 3.21 shows an example of this by comparing filter
shapes determined by the notched-noise method for each ear of subjects
with one normal ear and one with a cochlear hearing loss.
-
The first major consequence of this is a greater susceptibility
to masking by interfering sounds which may partly account for the difficulties
experienced by those with cochlear impairments in understanding speech
in noisy situations.
-
A second major consequence is difficulty in perceptual
analysis of complex sounds such as speech or music with a resulting reduction
in timbre discrimination required to discriminate different vowel sounds
or different musical instruments.
-
A hearing aid which simply amplifies sound does nothing
to correct impaired frequency selectivity.
-
General conclusions. Your textbook gives a good summary
of the major ideas in this chapter. I shall not attempt to summarize it
further here, but suggest that you study it carefully.
-
Signal detection theory.
-
Much of Chapters 2 and 3 have been concerned with the
measurement of thresholds which were classically conceived as the intensity
of a stimulus above which it will be detected and below which it will not.
-
It has been known for a long time that this view is
not satisfactory for a number of reasons.
-
Psychometric functions are continuous, rather than discrete,
and usually have lower limits greater than zero, as shown in Fig. 3.22.
-
A subject�s performance in such situations is affected
by a variety of motivational and response factors such as instructions,
payoffs and a priori probabilities.
-
Signal detection theory provides a way of separating
factors not directly associated with the discriminability of the signal
(bias, criterion, motivation) from factors relating to purely sensory capabilities.
-
It also accounts for the fact that responses may vary
from trial to trial even when identical stimuli are presented.
-
Basic assumptions of signal detection theory.
-
Decisions are based on the internal value of a random
variable (x) whose average value is monotonically related to stimulus magnitude
and can be thought of in terms of either a sensory or a physiological dimension
(e.g., loudness or number of neural firings).
-
The value of x varies randomly from trial to trial even
though the same stimulus is presented.
-
Some of this variability may be in the stimulus (e.g.,
a white noise masker has random variations in physical intensity).
-
Some of this variability is in the nervous system of
the listener (e.g., neurons show random firings in the absence of stimulation).
-
Thus, the average value of x will be greater if a signal
is present than if it is not, but the listener cannot be certain that a
signal has occurred on a particular occasion because of the inherent variability
of x.
-
On the basis of nonsensory factors the subject adopts
some criterion value of x and reports hearing the signal if the value of
x on that trial is greater than his/her criterion.
-
The probability density function for x on trials that
do not contain the signal is usually assumed to be normal or Gaussian,
as shown in Fig. 3.24.
-
The probability density function for x on trials that
do contain the signal is also usually assumed to be normal and to have
the same variance, but a greater mean.
-
The separation between the means of these two distributions
in standard deviation units is called d� and gives a measure of the detectability
or discriminability of the signal.
-
The location of the subject�s criterion gives a measure
nonsensory effects.
-
Hit rate (probability of saying yes given the signal
was presented) estimates the area under the signal plus noise distribution
to the right of the subjects response criterion.
-
Similarly, false alarm rate (probability of saying yes
given the signal was not presented) estimates the area under the noise
alone distribution to the right of the subjects response criterion.
-
Using these estimates and tables of the normal distribution,
d� and criterion location can be computed. These computations are taught
in PSYCH 3LL3 if you wish to learn them.
-
Signal detection theory thus provides a way to separate
changes in performance due to purely sensory factors from changes due to
motivation or criterion.
-
This logic is applicable to a wide variety of problems
in psychology.
-
In order to use signal detection theory, one must measure
false alarms, as well as the proportion of hits. This information was not
obtained in many classical psychophysical studies.
-
According to this theory there is no such thing as a
threshold, only degrees of discriminability.
-
One can measure the value of the signal that produces
a particular level of d�.
-
Alternatively, one can use a task such as two-alternative
forced-choice that attempts to eliminate effects of bias or criterion and
arbitrarily define threshold as corresponding to a particular percentage
correct (76% correct in two-alternative forced-choice is equivalent to
a d� of 1.0).