Professional Documents
Culture Documents
Absolute Judgement
Abstract
This report conveys the findings of an experiment that used a series of pitch absolute judgment
tasks to determine the affect loudness equalisation had on participant’s ability to distinguish
between a set of audio tones varying in frequency. A secondary goal of the experiment was to
verify the channel capacity for pitch-related absolute judgement tasks. While inconclusive due to
experimental limitations both hypotheses were generally confirmed. To elaborate, it was
determined that due to the participants ability to judge each audio tone on the basis of pitch and
volume when loudness equalisation was disabled, they were able to perform the tasks more
effectively. Furthermore, contrary to expectation, in both the disabled and enabled cases, the
data did not exceed the 2.5-bit channel capacity.
Introduction
An absolute judgement task is the action of making a judgement about a stimulus based on a
single trait or dimension. Absolute judgement tasks often require the user to distinguish a
stimulus along a continuum that has three or more options [1]. Absolute judgment also has its
applications in what George A. Miller calls T he Magical Number Seven, Plus or Minus Two [ 2]
which can be interpreted as the range of information the average human can hold in their
working memory. Therefore in an absolute judgement task, on average, a subject can
accurately distinguish stimuli if the number of options falls within this 7 ± 2 range.
In the realm of information theory, studies have found that certain stimulus dimensions
(brightness, loudness etc..) have varying channel capacities [3], which similar to 7 ± 2 range, are
measures of the amount of information in bits that can be reliably transmitted through a certain
channel. This experiment will investigate the channel capacity of a pitch absolute judgement
task while also determining if loudness equalisation effects the upper limit of a subjects channel
capacity.
Hypothesis Formation
Loudness equalisation is used to combat the impression that higher frequencies appear louder
to the human ear. This is achieved by adjusting the amplitude of certain frequencies such that
lower frequencies are easier to hear and higher frequencies are less loud [4]. In a pitch absolute
judgement task, loudness equalisation is integral to ensuring the subject uses only the pitch of
the audio stimuli, not the loudness. Therefore without loudness equalisation a subject would be
able to more easily distinguish between frequencies because there exist two dimensions on
which to judge a stimulus. As a result we expect the channel capacity of an absolute judgement
task to be higher without loudness equalisation.
As seen in Figure 1, a secondary goal of the experiment is to affirm the proportionality between
the input information H(s), and the transmitted information H(t) of a pitch absolute judgement
task. We are particularly interested in verifying the findings of Pollack that suggest an
asymptotic performance limit at channel capacity [5]. We also will attempt to verify that the
channel capacity with loudness equalisation occurs at 2.5 bits of transmitted information and
that information transmission increases but at a gradually decreasing rate [3].
Figure 1- An example of the asymptotic performance limit near 2.5 bits from data collected by
Pollack [5].
Method
Using the software provided, participants were asked to identify a set of tones to the best of
their ability. During each trial, the range of tones varied based on the amount tones in each set.
For example, a 4 tone trial would present the participant with tones at the following frequencies:
50 Hz, 100 Hz, 150 Hz, 200 Hz. As per the example, regardless of the number of tones in each
trial, the lowest frequency tested was 50Hz and subsequent frequencies would increase by
50Hz.
Each trial commenced with a mandatory learning phase where each tone would be assigned a
number. Participants were asked to listen to each tone and remember the number to which it
was associated. The assigned number increased with frequency (eg. 1-50 Hz, 2-100 Hz, 3-150
Hz, 4-200 Hz). Participants were permitted to listen to the tones twice either by replaying the set
or listening to individual tones of their choosing.
In the experiment phase, participants were presented a set of tones in random order with each
frequency presented 3 times. This ensured that regardless of the number of tones presented the
experiment time would be relatively short so as not to overwork the memory of the participant.
Necessitating each tone be presented 3 times also limited the presence of the Bow Effect which
states that less accurate performance is elicited at the centre of a stimulus-set range [1]. Taking
this into account, an experimental design with randomly presented tones would yield the
possibility that centre range tones would be presented more frequently, potentially skewing the
results. Participants were required to respond to each tone by selecting the number assigned to
it from the learning phase. The number of tones in each trial was determined randomly in order
to limit the participant’s capacity for improvement. The trials were conducted in two blocks of 10,
with and without loudness equalisation.
Results
The data presented below represents aggregated data from two participants for the trial
blocks with and without loudness equalisation (see Appendix B for individual data). The
graph in Figure 2 below plots H(s), the information (in bits) present in the stimulus set
against H(t), the information (in bits) transmitted by the participants. The formulae for
these values can be found in Appendix A as well an example of a stimulus-response
matrix that was employed as a tool to organize the collected data. The H(t) values
generated from each individual matrix were averaged to arrive at each data point. The
graph also contains a line for perfect performance (H(s) = H(t)) representing 100%
information transmission and a line for the pitch channel capacity of 2.5 bits.
The results of the experiment show that as the participants were presented an
increasing amount of tones, represented by H(s), more information was transmitted.
Furthermore, despite a few anomalies, generally, when loudness equalisation was
disabled, participants could distinguish the audio tones at a greater rate. This is
especially apparent when looking at the point in which each line veers from perfect
performance. With loudness equalisation, this occurs just past H(s) = 1.5 bits. Without
loudness equalisation, this occurs around H(s) = 2.3, representing a difference of
approximately 0.8 bits.
Although inconclusively due to sample size, the data confirms that 2.5-bit channel
capacity is not breached for pitch absolute judgement tasks. However, contrary to
Pollack’s findings, as H(s) increases, there is no evidence of asymptotic performance
near channel capacity. Rather, in the case when loudness equalisation is enabled, the
data simply increases at lesser rate towards channel capacity. In the disabled condition,
the data fluctuates violently leading to less definitive conclusions about the presence of
asymptotic performance.
Figure 2 - Data depicting the results of the absolute judgment tasks for pitch.
Discussion
Based on the results of this experiment the data generally support the hypothesis that
participants are able to better perform pitch absolute judgement tasks when loudness
equalisation is disabled. The findings also affirm that the 2.5-bit channel capacity acts
as a limit to pitch absolute judgement tasks. However, it was inconclusive that as H(s)
increases, participant performance approaches this limit asymptotically. Figure 2 also
supports the hypothesis that for a pitch absolute judgement task, the information
present in the stimuli is directly proportional to the information transmitted.
We are able to confirm the first hypothesis because performance when loudness
equalisation was disabled, performance improved, as evident by higher H(t) values. It is
difficult to comment on the exception occurring at H(s) = 3.32 and although it is contrary
to the conclusion stated, we suspect this result is due to the lack of a statistically
significant sample size and therefore can be dismissed as an outlier. Overall, the
improved performance can be attributed to the fact that with loudness equalisation
disabled, participants could use both the pitch and volume of the tone to determine its
assigned number. The process by which the results were collected did not account for
the presence of more information in the disabled case, therefore there was an inherent
difference in the H(s) values used between conditions in this experiment. Had this been
accounted for, the disabled data likely would have been flatter.
With all data values appearing under the H(t) = 2.5-bit channel capacity the experiment
affirmed that this acts as a limit to the amount of information that can be transmitted. It
is interesting that despite having effectively two dimensions on which to judge the audio
stimuli in the disabled case, the data does not exceed the channel capacity. We suspect
that had the experiment been designed such that the tone frequencies were spaced
further than 50Hz the effects of loudness equalisation would have been more
pronounced.
Conclusions
The experiment conducted in this lab attempted to investigate the effects of loudness
equalisation for pitch absolute judgement tasks. It affirmed the hypothesis that disabling
loudness equalisation allowed for more information to be presented in the stimuli leading to
improved performance when compared to the enabled case. In both cases, the 2.5-bit channel
capacity was not exceeded cohering to Miller’s 7 ± 2 empirically observed “magical” range.
Unlike Pollack’s findings that saw asymptotic performance approaching the channel capacity,
the findings only show a decreased rate of information transmission as more information is
presented.
Limitations
For reasons explained above the experiment necessitated that trials would present each tone 3
times in random order. However, an astute participant would be able to simultaneously while
performing the instructed task, remember the quantity of certains tones previously presented.
This would effectively allow them to “rule out” which tones would not be presented again. This
would be especially prevalent for trials with 5 tones or fewer as it would be easier to remember
which tones were presented.
Another limitation of the experiment was due to the apparatus used. In an effort to concurrently
perform the trial blocks. The two participants did not use the same hearing devices. This was
further compacted by the fact that the two devices differed significantly in quality and type
(Sennheiser HD 598 Cs Over Ear Headphones, Sony standard ear bud).
Lastly, the external environment during which the experiment took place was noisy and was
conducive to distractions potentially affecting the participant’s effectiveness in conducting the
experiment.
Implications
The results of this report highlight the limitations of absolute judgement tasks. Designers of
information sources must be particularly aware of these limitations at the risk of frustrating their
readers. For example, take a colour coded map and corresponding legend. The legend is
similar to the learning phase of this experiment because it allows the reader to assign colours to
certain values when viewing the map, much like assigning tones to numbers. However, if the
map designer employs too many colours, the reader will likely be unable to remember the
colours and their associated values and thus will be forced to look back and forth between the
map and legend. This is more time consuming and can potentially lead to frustration.
An example of a poorly designed map is seen in Figure 3. The map designer has overloaded
the reader’s working memory by using too many colours and does not rely on the use of shade
as the percentage increases or decreases.
Figure 3 - An example of a poorly designed map. [7]
References
[1] A. Yee, 'Information theoretic models - Part 3', Toronto, 2017.
[2] G. Miller, "The magical number seven, plus or minus two: Some limits on our capacity
for processing information.", Psychological Review, vol. 101, no. 2, pp. 343-352, 1994.
[3] W.N. Dember, J.S. Warm, "Psychology of Perception." NY:Holt, Rinehart & Winston
1979.
[5]I. Pollack, "The Information of Elementary Auditory Displays. II", T he Journal of the
Acoustical Society of America, vol. 25, no. 4, pp. 765-769, 1953.
[6] The world factbook 2009. Washington, D.C.: Central Intelligence Agency, 2009.
Appendix A
Calculations
The calculations for H(t) for each experiment was done using the stimulus-response matrix (see
below). The stimulus - response matrix was filled by probabilities of every stimulus-response
combination possible for every experiment to use the formulas shown above. The probabilities
were estimated by counting the number of times the participants had pressed that
corresponding response key for a respective stimulus presented on screen (frequency) and
dividing that number by the total number of responses for each experiment.
Figure 4: Example of a stimulus-response matrix tool used to calculate the information
values for a 4-stimuli set.
Appendix B