Your Worst Nightmare
An investigation into arousal-raising environmental sound features
David M. van der Kooij 10431543
Bachelor thesis Credits: 18 EC
Bachelor Opleiding Kunstmatige Intelligentie University of Amsterdam Faculty of Science Science Park 904 1098 XH Amsterdam Supervisor dr. F.M. Nack Informatics institute Faculty of Science University of Amsterdam Science Park 904 1098 XH Amsterdam June 26th, 2015
Abstract
This thesis addresses the question of what properties of the mechani-cal representation of sounds are associated with high arousal emotions in humans. Furthermore, I hypothesise that these mechanical proper-ties are sufficient to find a particular high arousal sound, a nightmare sound, for each user individually. This study focuses on environmental sounds and does therefore not consider speech and music.
Following Russell’s model of affect, this thesis applies a scale of arousal as the primary measure of evaluation. An auditory biofeedback loop was created incorporating machine learning, a sound database and fea-ture extraction to find that particular arousal raising sound for every user.
A user test has been conducted to assess the built system and address the question of which properties of sound are associated with the in-crease in arousal. It was concluded that sounds with a high power in the frequency range of 500Hz to 2500Hz increased the arousal of listeners the most. Furthermore, the method proposed in this thesis is concluded to be sufficient for finding high arousal environmental sounds in 77% of users.
Acknowledgements
I would like to thank the Waag Society for providing a working environment and for participating in the user test conducted in this thesis. Moreover, I want to thank Frank Kresin in particular for the initial proposal of this project and for supporting us throughout these three months.
My appreciation also goes towards all the other students and family mem-bers that took part in the user test.
Lastly, I would like to thank my supervisor Frank Nack for his insight and enthusiasm in our weekly conversations and for setting my priorities straight many times.
Contents
1 Introduction 4 2 Related work 5 3 Research Question 6 4 Approach 7 4.1 Sound Database . . . 7 4.2 Sound Features . . . 8 4.2.1 Frequency extraction . . . 8 4.3 Machine Learning . . . 94.3.1 Multiple Linear Regression . . . 9
4.4 User Test . . . 11
4.4.1 Biofeedback . . . 11
5 Results and Evaluation 13
6 Discussion 15
7 Conclusion 16
1
Introduction
The Your Worst Nightmare project aims to create an audiovisual experience that increases arousal in the user. The valence, i.e. the degree of attraction or aversion to the medium, can be either positive or negative. If the chosen valence is negative, the resulting experience can be nightmare-like for the user.
Furthermore, the project intends to achieve high immersion in the user by utilising an Oculus Rift for the visual portion and by creating an interac-tive experience in which users move through a labyrinth of rooms. Each of which will have generated visuals and sound based on biofeedback and previous user choices. A learning component was added to the system to try to learn from the user’s biofeedback and choices which results in a personal experience for each user and navigates the user towards the most disliked or liked medium available.
This document will not cover the entire project, but rather the auditory portion of the project, including a learning system based on Multiple Linear Regression.
Typical sounds, such as a high-pitched squeal of nails on blackboard are generally perceived as arousal-raising and unpleasant. An explanation as to why these particular sounds are perceived that way has two mutually inclu-sive components; it may have evolutionary origins such that these sensations are evoked by the mechanical properties of sound, or it may have associative origins whereby sounds become associated with particular objects, situations or events that in turn give rise to arousal-raising or unpleasant sensations. This thesis will attempt to discover which component has more influence over the increase in arousal. Furthermore, I hypothesise that the mechan-ical properties of sound should be enough to find the potential nightmare experience, assuming a real-time system and no previous knowledge of the user.
The remainder of the document is structured as such. The related work done on this topic will be explored first. Then the unaddressed problems in this research field will be investigated while posing the research question. Moreover, this document will cover an approach to creating an auditory biofeedback loop, which incorporates a sound database, feature extraction, biofeedback obtention and machine learning in a computer program. In addition, a user test has been conducted to determine the effectiveness of the system. The obtained results are then reported on and evaluated. Lastly, the caveats of the project and the computer system will be discussed and a conclusion on the effectiveness of sound feature modelling for the pur-pose of raising arousal levels is drawn.
2
Related work
The perception of sound is researched in the field of psychoacoustics. “Psy-choacoustics is [...] the branch of science studying the psychological and physiological responses associated with sound (including speech and mu-sic)” (Wikipedia, 2015). Human sound perception is partially based on the mechanical properties of sound, such as frequency, intensity, direction and distance. However, the inner ear transforms those properties into neural action potentials that then travel to the brain. Therefore, it is important to remember that the nervous system and specifically the brain is involved in the perception of sound. Consequently, a sound will trigger other associated memory structures, such as other sounds, events or images, which in turn influence the perception of the original sound.
Kumar et al. (2008) map the unpleasantness of environmental sounds to the auditory representation of those sounds. Kumar’s research is closely related to this thesis, however, Kumar measures the unpleasantness i.e. the valence of the sounds, while this thesis will focus on the arousal induced in the user. In his conclusion, Kumar infers that the frequency range of 2500 to 5500 Hz corresponds to an increase in the unpleasantness of a sound. Gygi et al. (2007) ran four experiments regarding environmental sound simi-larity and categorisation. They created a 3D model for sound categorisation which resulted in a clear grouping of continuous sounds, discrete impact sounds and harmonic sounds on the first two dimensions. The third dimen-sion distinguished between vocalised and non-vocalised sounds. Although sound categorisation is not employed in this thesis, the categories estab-lished by Gygi and his colleagues will help to define sound features for the system.
Dingler et al. (2008) studies the learnability of types of sound. They use dif-ferent types of sonification techniques to represent common environmental features. An auditory icon, for instance, the sound of wind going through leaves is a sound that is very recognisable from experience. An earcon, how-ever, is a non-vocal synthetic sound that were decomposed in five dimensions, namely, rhythm, pitch, timbre, register and dynamics. These sounds proved to be more difficult to learn. Since in part learning is associative, a sound that has less associated concepts is more difficult to learn. Furthermore, such a sound will presumably also elicit a less emotional response.
Russell (1980) proposes a circumplex model of affect in which the two axes represent arousal (y-axis) and valence (x-axis). The model can be utilised to classify emotions on these scales, for instance, annoyed can be classified as a negative valence, medium arousal emotion while excited is a positive valence high arousal emotion.
that there are only a small number of emotions. Ekman (1992) for exam-ple, states that there are six core emotions; anger, disgust, fear, happiness, sadness, and surprise. These emotions represent discrete categories in which varying degrees of that emotion are possible. This thesis does not discuss the validity of these models, however the circumplex model by Russell is more appropriate because of the arousal scale present in that model. Guo and Li (2003) study audio classification and retrieval using a Sup-port Vector Machine (SVM). They find that an SVM can achieve a lower error rate than other popular audio classification methods (Euclidean dis-tance, decision-tree). They also study audio retrieval based on a technique they call distance-from-boundary (DFB), in which the boundaries are also learned with SVMs.
Guo’s article shows that an SVM is a good technique in audio classification and retrieval. Although SVMs will not be utilised in this thesis, a Support Vector Machine is still a technique worth considering when handling audio.
3
Research Question
This thesis will investigate the influence of sound on arousal in humans. To discover this relation, sounds will be characterised by their features. Fea-tures in this context mean the properties of a sound, such as the frequencies present in the sound, the intensity or amplitude of a sound, the timbre of a sound, and the direction and distance of a sound. All these properties can have an impact on how the sound is perceived and on the arousal of the lis-tener. Furthermore, associations of sound with particular events, locations and memories can also have an impact on the user’s perception and arousal. Which features are most relevant in influencing arousal is still unclear, how-ever, Russell’s model (Russell, 1980) seems to provide a link between those features and the experiences they can trigger. Hence the research question for this thesis becomes which features of environmental sounds increase the listener’s arousal the most?
4
Approach
In order to obtain an answer to the research question, a computer pro-gram was created. Users of this system listen to environmental sounds and are enquired about the arousal they experienced during the sound. The system then learns which sounds increase arousal for the user and tries to present sounds with increasingly more arousal. Incorporated in the system are a sound database, a feature extraction algorithm, a machine learning algorithm, and biofeedback obtention, among others. Figure 1 shows an overview of the system and the data flow.
The entire system was programmed in Python because of the extensive
Figure 1: An overview of the computer system library of modules, particularly in the field of machine learning.
4.1 Sound Database
The sound database consists of 45 environmental sounds ranging from na-ture sounds to traffic sounds and other sounds one would ordinarily come across in the world. All sounds have a length of six to eight seconds. The first six sounds are identical for all participants. These sounds are
‘an-gryCrowd.wav’, ‘animals1 long.wav’, ‘blackboard nails long.wav’, ‘fork glass long.wav’, ‘ocean waves.wav’ and ‘office.wav’. They represent the sound database
be-cause of their varying frequencies, amplitudes, and expected arousal levels. The biofeedback collected from those first sounds is used to determine the initial coefficients for the machine learning algorithm. The majority of the sounds were gathered from the IADS database (Bradley & Lang, 2007), the other sounds were collected from various sources on the web, such as SoundBible.com (SoundBible, 2015). Sounds gathered from the web have the extension ‘ long’ in their name. These sounds have been lengthened or
shortened to the desired six to eight seconds, while the sounds collected from the IADS database already had the desired length.
4.2 Sound Features
Each sound in the database was characterised by five acoustical features. In the field of psychoaccoustics, the frequency and amplitude of a sound are considered to be critical factors in sound perception (Plack, 2013). There-fore, this thesis utilises four features based on the frequency distribution of a sound and one feature based on rapid increases in amplitude. The first four features are build on Kumar’s findings that the frequency range of 2500Hz to 5500 Hz correlates with an increase in unpleasantness (Kumar et al., 2008). To be able to compare that range with other frequency ranges, the remain-der of the frequency spectrum up to 15000 Hz was covered by three more frequency ranges; 0Hz to 500Hz, 500Hz to 2500Hz and 5500Hz to 15000Hz. The last feature utilises a hard-wired evolutionary trait in people. A sound that suddenly increases in amplitude is generally considered as arousal-raising since it is unexpected and might indicate a dangerous situation. For this last feature, every 0.1 seconds of audio is summed up in bins. The program then keeps track of the number of times the amplitude of a bin is half the amplitude of the highest bin higher than the previous bin. This amount is applied as a feature by the program.
4.2.1 Frequency extraction
For the features that utilised the frequency distribution of a sound, the Discrete Fourier Transform (DFT) was implemented to compute this distri-bution. The DFT transforms the digital audio signal into a list of coefficients of complex sinusoids, ordered by their frequencies. Essentially it translates the sound from the time domain into the frequency domain. The definition of the Discrete Fourier Transform is:
Xk ≡ N −1 X n=0 xn· e−2πik n N (1)
in which x0, x1, . . . , xN −1 is the audio source and X are the output
coeffi-cients of the complex sinusoids.
With a complex coefficient for each frequency, the amplitude in that fre-quency range for that sound is calculated with
where a is the real part of the coefficient and b is the imaginary part. The resulting values are divided into four bins. Each bin represents a por-tion of the entire frequency spectrum. The sum of all values per bin is then equivalent to the total power or amplitude in that frequency range. When divided by the total sum of frequency power over all bins, the percentage of power per bin is calculated. The frequency ranges used for the begin and end points of the bins are 0 Hz to 500 Hz, 500 Hz to 2500 Hz, 2500 Hz to 5500 Hz and 5500 Hz to 15000 Hz. The specific ranges are derived from (Kumar et al., 2008), which states that the 2500 Hz to 5500 Hz frequency range correlates with an increase in unpleasantness. To discover the impact of the remainder of the frequency spectrum, the other three ranges are used to cover the entire spectrum to 15000 Hz. Thus, the percentages of power in the four frequency ranges are utilised by the system directly as features to characterise a sound.
4.3 Machine Learning
The aim of machine learning is to learn from the user which sounds increases the arousal the most in that specific user. For that purpose, the sound fea-tures are matched to the biofeedback from the user via Multiple Linear Regression (MLR). Although other machine learning techniques might be more fitting for this project, MLR was chosen because of the level of insight in the inner workings of the algorithm during and after the execution. A neural network was considered, but disregarded because it comes with lit-tle control over its internal workings. Furthermore, when learnt to predict biofeedback from sound features, predicting sound features from biofeedback with the same values in the nodes of the neural network was not possible with the available python modules.
4.3.1 Multiple Linear Regression
MLR tries to fit a linear model on the data. In this case, it tries to find coeffi-cients that, when multiplied by the sound features, result in the biofeedback from the user. In a formula:
a1x1+ a2x2+ · · · + anxn= B (3)
where n is the number of features, a1. . . an is the coefficient vector and
x1. . . xnis the sound feature vector. Multiplied they result in B, the
biofeed-back for the sound with those sound features.
multiplication becomes a matrix multiplication: x11 x12 · · · x1n x21 x22 · · · x2n .. . ... . .. ... xm1 xm2 · · · xmn × a1 a2 .. . am = b1 b2 .. . bm (4)
where m is the number of sets of features, and n is the number of features. To calculate the coefficients vector A, the MLR algorithm implements a least squares fit algorithm. Once computed, the coefficients vector represents a user and which sound features have the most impact on the arousal of that user.
As the aim of the system is to increase the arousal, a predicted (or aimed) arousal statistic is used. The predicted arousal is a function of the arousal the user experienced on the previous sound and is always an increase. With the predicted arousal a new set of features is computed through the Multi-ple Linear Regression algorithm. This time the coefficients vector and the predicted arousal (predicted biofeedback) are used to approximate a vector of sound features. The vector multiplication then looks like:
a1 a2 · · · an × x1 x2 .. . xn =bp (5)
where the sound feature vector X is unknown as opposed to the coefficients vector A in equation 4.
Vector X then holds the sound features that are anticipated to induce the aimed arousal. Because a sound with exactly those features will not be available, the closest sound in the featurespace (i.e. the feature vectors of the remaining sounds) is found using the Euclidean distance measure. In the case that the experienced arousal (biofeedback) is lower than the predicted arousal, the system computes new coefficients using the following formula:
(N − 1)(c1o. . . cno) + (c1c. . . cnc)
N = c
1
n. . . cnn (6)
where c1o. . . cno are the old coefficients, c1c. . . cnc are the coefficients of the fit where the arousal was lower than predicted (current coefficients) and N is the number of played sounds up to that point. Thus, the new coefficients, c1n. . . cnn, follow from the weighted average of the old coefficients and the current coefficients.
gener-a predetermined number of loops gener-are finished.
4.4 User Test
For the purpose of evaluating the system and answering the research ques-tion as posed in secques-tion 3 a user test was conducted. 22 subjects took part in this test, of which seven were female. Their ages ranged from 16 to 55 (average 32) years old. Participants were seated in a quiet office setting and wore headphones with a noise cancelling technique. A single test plus some preparation took approximately fifteen minutes.
The test involved participants listening to a range of twenty sounds. The first six sounds were the same for every subject and served as the initial val-ues for the Multiple Linear Regression algorithm. These sounds are listed in section 4.1. The remaining fourteen sounds were selected by the learning algorithm from a pool of 39 sounds. After listening to all 20 sounds, the participants were asked whether their nightmare sound was played and if this was the case, which sound it was.
4.4.1 Biofeedback
Measuring the arousal experienced by users while listening to sounds is a vital part of the system. From the wide range of measuring techniques the electroencephalogram (EEG) was chosen as the most appropriate for the Your Worst Nightmare project. An EEG can most directly measure one’s well-being and hence also indicate an increase in arousal. Alpha waves, for instance, can be useful in measuring arousal since they are generally consid-ered to indicate a state of relaxation and low tension (Andreassi, 2000). However, the integration of this measuring technique was not finished be-fore the user test started. Therebe-fore users were instead enquired to indicate their arousal on a scale, ranging from 0 to 100, during the test. After every sound the system waited ten seconds for the subject to indicate their arousal level for that sound. In the case that the user responded quicker, the sys-tem would still wait ten seconds to ensure the subject’s echoic memory was clear of any stored sounds. Echoic memory holds sounds as a ‘holding tank’, since one sound might not be comprehensible on its own, while it can be understood when joined with the succeeding sound. According to a study by Glucksberg and Cowen (1970), echoic memory can hold a sound up to twenty seconds in the absence of interference. In this test however, the sub-jects were tasked to indicate their arousal during those ten seconds which interfered with their ability to hold the sound in echoic memory. Therefore, the waiting time between sounds was lowered to ten seconds.
expected biofeedback was recorded by the system. Lastly, they were asked to indicate the sound they found to be the ‘worst’ sound as a qualitative measure.
5
Results and Evaluation
The results generated by the 22 tests were analysed and the findings are detailed below.
As part of the research question, this document intended to discover the impact mechanical features have on increasing arousal. Specifically, five fea-tures were chosen to characterise a sound. To demonstrate the effect of each feature separately, the features of each sound were multiplied by the arousal indicated by the users for that sound. If the arousal was higher than 50 (more arousal raising than neutral) the product is added to a total, when lower than 50, the product is subtracted from the total. A higher figure then translates to more impact on arousal. The calculated results are listed in table 1.
Feature description Feature impact Frequency range: 0Hz-500Hz 47.47 Frequency range: 500Hz-2500Hz 685.10 Frequency range: 2500Hz-5500Hz -508.96 Frequency range: 5500Hz-15000Hz 599.74
Rapid differences in amplitude 115.60
Table 1: Impact of features on arousal, per feature
Table 1 indicates that the frequency range of 500Hz to 2500Hz has the most influence over the increase in arousal. Following closely is the frequency range from 5500Hz to 15000Hz. Although Kumar et al. (2008) found a corre-lation between the frequency range of 2500Hz to 5500Hz and unpleasantness in sounds, the results in table 1 indicate the contrary.
Table 2 presents statistics on the utilised sounds. It displays the ten sounds with highest average arousal over all tests, the minimum and maximum arousal assigned to them and the standard deviation per sound.
Name Avg. arousal Min. arousal Max. arousal Std. Dev. maleVomitting 79.8 55.0 99.0 13.0 dentist drill long 78.8 50.0 95.0 13.8 baby crying long 72.5 25.0 99.0 18.2 blackboard nails long 71.1 35.0 95.0 15.0 siren long 69.8 50.0 85.0 9.1 fork glass long 65.4 22.0 97.0 19.6
belch 65.0 50.0 95.0 17.6
siren2 63.0 63.0 63.0 0.0
bees 62.2 55.0 76.0 7.8
carBrakes 61.0 30.0 80.0 13.0
Table 2: Ten sounds ordered by highest average arousal
Three more result sets were obtained by analysing the test data. The first set consists of the positions of the highest arousal sound per participant. In four out of 22 cases, the highest arousal sound was the first sound. In one case, the last sound was considered the worst, on average the sound had position 5.8 in the series and the median was 6.5.
Associated with each of the fourteen testing sounds was an expected arousal value. The expected arousal served as input for the machine learning algo-rithm, as detailed in section 4.3.1. Unfortunately, the arousal as indicated by the users was sometimes lower than the expected value. The second result set comprises the amount of times this was the case for each user. For three users the actual arousal was lower than the expected arousal five times, the smallest value, for one user eleven times, the highest value. On average, the amount of times the arousal was lower than expected was 7.2 times, which is slightly more than half the time.
The third result set is composed of the sounds that users indicated to be the ‘worst’ sound from the sounds they listened to. Nine users pointed out the sound ‘maleVomitting.wav’, five users stated the sound ‘baby crying long.wav’ was the worst for them. The remainder of the users named ‘siren long.wav’ and ‘office.wav’, among others. Two users stated none of the sounds in-creased their arousal significantly.
As stated in the introduction, this thesis hypothesised that the mechani-cal properties of a sound should be enough to find the potential nightmare sound. In the user test 77% (17/22) of the users indicated an arousal of 80 or higher and 59% (13/22) of the users indicated an arousal of 90 or higher at least once.
6
Discussion
The goal of this thesis was to determine which acoustical/mechanical au-dio feature has the most influence over increasing arousal. Moreover, this document aimed to establish whether mechanical properties of audio suffice to find a potential nightmare sound. A machine learning algorithm centred around Multiple Linear Regression was used to model the relation between the properties of a sound and the potential arousal induced by that sound. Together with a sound database, feature extraction and data analysis, the machine learning algorithm was compiled into a computer program with the aim to answer the research question. A user test was conducted on 22 participants and the resulting statistics are detailed in section 5. Although the research question was answered by these results, some questions remain unaddressed.
Foremost, result set two indicated that the actual arousal of users was 51% (7.2/14) of the times lower than the expected value. Moreover, during the tests the arousal often oscillated instead of the desired gradual increment. Multiple explanations could elucidate this unexpected result. A possible cause might be that the five chosen features do not model the sounds suf-ficiently causing the machine learning to predict the incorrect sound. A different, though mutually inclusive explanation, might be that the associ-ations evoked by the sound play a prominent role in the arousal it induces. Since the properties used to characterise the sounds are purely mechanical, they do not capture the associated memories a sound might elicit. Although this explanation was already put forward in the introduction, it is important keep in mind when conducting similar research.
In table 2 the ten sounds with highest average arousal are presented. No-table in this No-table is difference between the minimal and maximal arousal assigned to the sounds. In the case of ‘fork glass long.wav’ this difference is as large as 75. One user apparently found this sound quite relaxing, while for others this was their worst sound. The standard deviation confirms this contrast. A presumable cause is the difference in interpretation of the meaning of the sound. Different users assigned different meanings to the sound which caused them to associate the sound with different memories, especially when the sound was ambiguous to some degree. One participant, for instance, associated this particular sound with two people having inter-course on a squeaky bed, which resulted in an arousal value below 50. Furthermore, the manner of obtaining biofeedback in this thesis, namely through self-reporting arousal, can be viewed as unreliable because of the subjective experience they report. However, according to Mauss and Robin-son (2009), a self-report on emotions currently experienced is considered more valid as opposed to emotions experienced some time ago. Moreover, all participants used the same paper scale during the user test, which had clear indications of the sought-after emotions. Even though the validity
of the self-report improved with these adjustments, the problem cannot be fully addressed and should be considered when making use of these results. Lastly, although all sounds were selected with care, the sound database utilised in this thesis could contain a bias based on culture or gender that could influence the results if conducted on different participants or in a non-western society.
7
Conclusion
In this thesis a machine learning algorithm centred around Multiple Lin-ear Regression was used to model the relation between the properties of a sound and the potential arousal induced by that sound. Five mechani-cal sound features were chosen to characterise a sound based on work done by Plack (2013) and Kumar et al. (2008) and the arousal scale from Rus-sell’s circumplex model of affect (Russell, 1980) was implemented as the main measure of evaluation. After these segments were constructed into a computer program, a user test was conducted from which the findings are described in section 5.
The conclusion drawn from these findings is that the method applied in this thesis is sufficient for finding high arousal environmental sounds. Moreover, the amount of power in the frequency range from 500Hz to 2500Hz of a sound correlates most positively with the perceived arousal evoked by that sound.
Nevertheless, the approach proposed here is not optimal and should therefore be developed further. Future work on this topic could consider a different machine learning technique, such as a neural network or a Support Vector Machine or examine the effectiveness of an electroencephalogram (EEG).
References
Andreassi, J. (2000). Psychophysiology: Human behavior and physiological response. Lawrence Erlbaum.
Bradley, M. M., & Lang, P. J. (2007). The international affective digitized sounds (iads-2): Affective ratings of sounds and instruction manual. University of Florida, Gainesville, FL, Tech. Rep. B-3 .
Dingler, T., Lindsay, J., Walker, B. N., et al. (2008). Learnability of sound cues for environmental features: Auditory icons, earcons, spearcons, and speech. In Proceedings of the 14th international conference on auditory display, paris, france (pp. 1–6).
Glucksberg, S., & Cowen, G. N. (1970). Memory for nonattended auditory material. Cognitive Psychology , 1 (2), 149–156.
Guo, G., & Li, S. Z. (2003). Content-based audio classification and retrieval by support vector machines. Neural Networks, IEEE Transactions on, 14 (1), 209–215.
Gygi, B., Kidd, G. R., & Watson, C. S. (2007). Similarity and categorization of environmental sounds. Perception & psychophysics, 69 (6), 839–855. Kumar, S., Forster, H. M., Bailey, P., & Griffiths, T. D. (2008). Mapping
unpleasantness of sounds to their auditory representation. The Journal of the Acoustical Society of America, 124 (6), 3810–3817.
Mauss, I. B., & Robinson, M. D. (2009). Measures of emotion: A review. Cognition and emotion, 23 (2), 209–237.
Plack, C. J. (2013). The sense of hearing. Psychology Press.
Russell, J. A. (1980). A circumplex model of affect. Journal of personality and social psychology, 39 (6), 1161.
SoundBible. (2015). Soundbible.com — free sound clips, sound bites and sound effects. Retrieved from http://soundbible.com ([Online; accessed 22-June-2015])
Wikipedia. (2015). Psychoacoustics — wikipedia, the free encyclopedia. Retrieved from
https://en.wikipedia.org/w/index.php?title=Psychoacoustics&oldid=660060100 ([Online; accessed 17-June-2015])
