2016 International Conference on Micro-Electronics and Telecommunication Engineering

Mood Based Classification of Music by Analyzing

Lyrical Data Using Text Mining
Nirbhay Kashyap1, Tanupriya Choudhury2,Dev kumar Chaudhary3,Roshan Lal 4
Amity University UttarPradesh, Noida – INDIA
1
nkashyap@amity.edu,2tchoudhury@amity.edu,3dkchaudhary@amity.edu,4rchhokar@amity.edu

Abstract - This research task looks at the part lyric content and 2) Mood Cataloging: There is consistently a civil quarrel on
additionally audio features can play in enhancing audio music whatever sort of feeling a pleasant-sounding piece can prompt
mood arrangement. With the developing measure of computerized seen by us. Nevertheless investigation work likewise give the
music and human different requests to music data recovery,
programmed music conclusion investigation is turning into a
premise to mood classification.[2]
critical and vital assignment for different framework and 3) Acoustic Cues: In the given work we have managed decoded
applications, for example, music association, melody choice in cell sound information which it is very hard to translate mood
phones, music proposal et cetera. This undertaking is to naturally specifically. Numerous model are proposed earlier in removing
stamp a melody utilizing emotional marks as a part of an emotion critical highlights comparable mel-frequency cepstral
set determined by analysts. In the most recent couple of years, it coefficients, motion strength, zero Crossing rate and so forth,
has pulled in more consideration and extensive varieties of related from the acoustic sign. In this way, to manufacture an effective
researches have been completed. The proposed system obtained model for mood recognition in music, we have to concentrate
85% accuracy from the trained mood mapper, which might lead acoustic highlights to speak to the premise of different moods.
to a more accurate understanding of the music mood in the mood
mapping process.
In our system, power (vitality), resonance and mood (beat)
highlights are separated, from the acoustic sign to guide through
Keywords –music, data mining, music mood, audio feature. specific classification of mood.

I. INTRODUCTION: II. MUSIC AND DATA MINING

With the quickly expanding innovation and the new approach A. Overview of data mining
in most recent sight and sound devices, music has come to each Data mining is a field which manages the extraction of
individual's close to home contraptions , be it be tablet, music fascinating, non-trifling, implied, already obscure and
player or mobile phone. The connection between musical conceivably helpful examples or learning from colossal
sounds and their impact on the audience members' feeling has measure of data. Data mining is frequently termed as learning
been very much considered and is clear from the abundantly disclosure in databases . An ordinary information disclosure
commended papers, for example, that of Hevner and procedure can be seen as delineated.
Farnsworth. As of now we can store, sort and recover our B. Music Mood Taxonomy
digital music taking into account different conventional music There are two mood model sorts of how human see feeling:
classification labels like Artist, Band (Group), Album, Movie category model and dimensional model.
and Genre[1]. Then again, picking a melody or music piece 1) Category model: It comprises of various separate
suiting our mood from a substantial database is troublesome fundamental moods, for example, glad, pitiful, outrage,
and time intensive, since each of the specified parameter can't loathsomeness thus on however these moods are shifted
sufficiently pass on the enthusiastic viewpoint connected with rapidly, since the essential moods are irregular.
the tune. What we need is an extra parameter or rather hunt 2) Dimensional model: James Russells characterizes the two
channel here - Mood – which means the feeling of that specific dimension valence (negative-positive) and excitement (high-
music piece. low), relating to Thayer's. What's more, this definition is
A. Mood detection Challenges utilized broadly now. In the 1990s, Thayer set forward a two-
1) Mood Perception : Properties of music may give general dimensional model. This model speaks to moods as focuses in a
signs to enthusiastic importance. Yet, it is hard to evaluate this two-dimensional space: Stress (cheerful/restless) and Energy
theory, on the grounds that most research on the theme has (smooth/vitality) vector plane. In this model, a wide range of
included requesting that audience members judge the passionate moods are constant, the smooth and progressive move among
significance of harmony, instruction and discrete identity. Thus, different moods can be accomplished and the comparability or
for the identical musical part, distinctive people may have contrast can be measured from the separation in the space[3].
diverse recognitions. Further scientists in demonstrate the C. Mood models
change in impression inalienable in a stanza inside social
connection may also be insignificant. Along these lines, it is Model is broken into different models:
conceivable to manufacture a mood identification framework in Model 1: Information characterization calculations keeping in
a certain connection mind the end goal to learn, prepare and test the model speaking

978-1-5090-3411-6/16 $31.00 © 2016 IEEE 289

287
DOI 10.1109/ICMETE.2016.65
to the moods of these audio tunes and built up an open source Excitement, Exciting, Exhilarating, Thrill, Ardor
structure for the same The field of content mining looks to Exciting
concentrate helpful data from unstructured literary information Words
through the ID and investigation of fascinating examples. Dreamy Dreamy, Excellent, Pensive, Misty, Relaxing
Model 2: Another strategy is proposed to assemble an Words
expansive ground truth set of melodies and 18 mood
Pessimism, defeatist, despairing, downbeat, hopel
classifications in light of social labels in order to mirror a
Pessimism ess
practical, client loped viewpoint. The greatest accomplishment
Words
lyric list of capabilities was likewise contrasted with a main
Romantic, Fantastic, Glamorous, Exotic, Colorful
song related framework. In consolidating lyric and audio Romantic
sources, half and half capabilities constructed with three diverse Words
highlight determination systems were additionally analyzed. Happy Cheerful, Delighted, Glad, Joyful, Lively, Merry
Model 3: The outcomes show designs at chances with Words
discoveries in past studies: audio features don't generally beat Angst Agony, Dread, Nervousness, Blues,, Misgiving
words geographies, and joining words and song features can Words
enhance execution in numerous mood classifications. For
content mining characterization, the venture makes utilization of D. Naïve Bayes Classifier used in text mining work
Naive bayes classifier. Naive Bayes is a candid approach for evolving classifiers:
representations that assign class names to matter existences,
spoke to as paths of highlight assets, wherever the session
Mood Words inscriptions are haggard from certain incomplete set. It is a
Category cluster of procedures attractive into account a characteristic rule
Name not a secluded procedure for making such classifiers,: all
Stop A, An, The, This, That, He, She, Is, Am, Are, classifiers of Naive Bayes receive that the approximation of a
Words Then exact highpoint is free of the approximation of certain additional
Sad Dejected, Sorry, Sad, Sorrowful, Wistful, Hurting highpoint, assumed the class variable.
Words
In utilization knowledge, classifiers of Naive Bayes are a
Angry Annoyed, Bitter, Angry, Heated, Furious, Hateful
assortment of open probabilistic classifiers considering applying
Words
Bayes' theory with hard independence qualms amongst the
Aggression, Attacking, Barbaric, Advancing,
Aggressio Martial
highpoints [4].
n Words For instance, a natural product may be thought to be an apple on
Average, Garden, Familiar, General, Regular the off chance that it is red, round, and around 3" in width. A
Average naive Bayes classifier considers each of these highlights to
Words contribute freely to the likelihood that this organic product is an
Brooding, Contemplative, Meditative, Reflective apple, paying little respect to any conceivable correlations
Brooding between the shading, roundness and breadth highlights. The
Words standard formula for inverting conditional probabilities is
Calm Calm, Comfort, Quiet, Chill out, Serene P(C|A)=P(A|C)P(C)/P(A) ---------------Equation1.
Words We've looked at ailments and symptoms, but the method can be
Cheerful, Cheer up, Festive, Jolly, Jovial used whenever we need classifications of cases described in
Cheerful terms of attributes [5].
Words The more general version of Bayes rule deals with the case
Confident, Encouraging, Encouragement where C is a class assessment and the attributes are A1,A2,A3
Confident to upto An.
Words P(C|A1,A2,A3….An)=ς௡௜ୀଵ ܲሺ‫݅ܣ‬ȁ‫ܥ‬ሻܲሺ‫ܥ‬ሻ /P(A1,A2,….An).
Depressed, Blue, Dark, Depressive, Dreary A Naive Bayes Classifier is a program which forecasts a class
Depressed value given a set of set of attributes.
Words For each known class value,
Desire Desire, Hope, Hopeful, Aspiration, Wish Calculate probabilities for each attribute, conditional on the
Words class value.
Earnest Earnest, Close, Firm, Fixed, Grave, Intent Use the product rule to get a joint conditional probability for the
Word attributes.

288
290
Use Bayes rule to derive conditional probabilities for the class • Beat histogram: This feature helps to identify the
variable. strength of different rhythmic periodicities in a signal. This is
E. Features of audio calculated by taking the RMS of 256 windows and then taking
1) Feature list: Following features are there the FFT of the result.
• Root Mean Square (RMS): RMS is designed on a • Strongest frequency via Zero Crossings: It represents
apiece space basis. the maximum frequency of the signal existing at the Zero
• Rhythm: Social mood retort is uttered by the tempo crossing point. This is originate by mapping the fraction in the
and rhythm periodicity existing in a melodious piece. zero-crossings to a frequency in Hertz.
Systematic rhythm with wanton rhythm may be supposed as • Mel Cepstral coefficient: This feature constitutes the
communicating energy, while lopsided rhythm with deliberate coefficients derived from the Cepstral illustration of audio
measure transports uneasiness signal such that the frequency bands are equally spread out on
• Magnitude Spectrum: This article summaries the Fast the Mel scale approximating the human auditory system's
Fourier Transform degree gamut from a set of acoustic response more closely. MFCCs are more usually and widely
illustrations. It delivers a noble idea nearby the extent of used as features in speech recognition schemes. In latest times
different occurrence workings inside a opening. The range field these features are progressively finding uses in Music
is found by main conniving the FFT with a Hanning opening. information retrieval, audio similarity measures, genre
The magnitude gamut value for both holder is found by first classification etc..
summing the squares of the real and imaginary components. • Linear Predictive Coding coefficients: This feature
• Power Spectrum: It extracts the FFT power from a set helps in demonstrating the spectral envelope of an audio or
of audio samples. It stretches a decent impression about the speech signal.
influence of different incidence mechanisms within a window. • Spectral smoothness: This feature is intended by
• Spectral Roll-off- Point: The phantom roll-off- evaluating the log of a partial minus the average of the log of
opinion is the portion of bins in the control range at which 85% the surrounding partials and is based upon Stephan
of the power is at lower frequencies. It denotes the amount of Macadam’s Spectral Smoothness. This feature assistances in
the right-skewness of the influence band[6]. identifying the peak based control of the evenness of an audio
• Spectral Centroid: It is a quota of the "centre of mass" signal.
of the power gamut. It is obtained by computing the mean bin • Relative difference function: It characterizes the onset
of the supremacy spectrum. The outcome resumed is a amount detection and is intended as the log of the derivative of the
from 0 to 1 that signifies at what fraction of the total number of Root Mean Square value.
bins this central frequency is. • Mood: This is the class attribute that needs to be
• Spectral Flux: It measures the amount of spectral inhabited during the training and which is detected routinely
change of a signal by calculating the difference between the while testing the classifier against a new audio file.
current value of each magnitude spectral bin in current window In the given list of audio features some features have a single
and the corresponding value of the magnitude spectrum of the dimension for instance, Strongest Beat, which has just one
previous window. Each of these differences is then squared, value and on the contrary, some features have variable
and the result is the sum of the squares. dimensions for instance Beat Histogram, which has a series of
• Spectral Variability: It stands for the standard values showing the histogram. The variable measurement
deviation of the magnitude spectrum of the audio signal. however depends on the window size, which in case of this
• Fraction of low energy window: This procedures the work has been kept constant to 32 for every 30second audio
noiselessness of the signal comparative to the rest of a signal clip in the dataset. Hence including the class attribute, the
and is premeditated by taking the mean of the root mean square dataset consists of a total of 330 feature vectors [8].
of the last hundred windows and finding what fraction of these 2) Classification using Support Vector Machine algorithm: The
hundred windows are below the mean[7]. solution is to utilize pertinence criticism, particularly Support
• Zero Crossings: This feature supports recognize the Vector Machine dynamic learning, to take in a classifier for
pitch as well as the noisiness of a signal. It is premeditated by every inquiry. An inquiry is then a mapping from low level
finding the number of times the signal changes sign from one sound features to more elevated amount ideas, altered. To start
sample to another crossing the zero value. a search, the client presents the framework with one or more
• Strongest Beat: This feature finds the strongest beat in cases of melodies of interest or "seed" tunes. The framework
a signal. then repeats between preparing another classifier on named
• Beat sum: It is calculated by summing up the beat melodies and requesting new marks from the client for
values of a signal and gives a measure of how important a role instructive samples. Hunt continues rapidly, and at each stage
regular beats play in a piece of music. the framework supplies its best gauge of suitable melodies.
Since it takes a lot of time to listen to every tune returned by a

289
291
hunt, our framework endeavors to minimize the quantity of field, is the issue of channel befuddles, in which the enlistment
melodies that a client must name for an inquiry sound has been accumulated utilizing one contraption and the
Active learning has two principle points of interest over test sound has been created by an alternate channel. It is
customary SVM classification. First and foremost, by imperative to note that the wellsprings of confound fluctuate
presenting the client with and preparing on the most instructive and are by and large truly confused. They could be any mix and
melodies, the algorithm can accomplish the same classification more often than not will be not constrained to confound in the
execution with less marked illustrations. Second, by permitting handset or recording device, the system limit and quality,
the client to rapidly name the arrangement of occasions, a clamour conditions, sickness related conditions, anxiety related
solitary framework may perform any number of classification conditions, move between diverse media, and so forth. Some
and recovery errands utilizing the same pre-computed methodologies include standardization or something to that
arrangement of features and classifier system. For instance, the affect to either change the data (crude or in the feature space) or
framework may be utilized to hunt down female craftsmen, to change the model parameters.
content melodies, or hallucinogenic music.
A large ground database of 100s of songs is made. The songs III. LITERATURE REVIEW:
are first converted in .wav format. The songs are then tested in G. Weiss and B. Davison, in Handbook of Technology
the feature extractor developed. Spectral Centroid is ticked and Management, John Wiley,2006, reasoned that Data mining at
rest all features are un-ticked. In the database, a table named first created a lot of fervor and press scope, and, as is regular
‘feature’ is developed which has two columns-‘feature name’ with new "advances", exaggerated desires. In any case, as data
and ‘feature value’. Xml files are generated depicting the result mining has started to develop as a train, its systems and
of feature extraction of each song. These values are than methods have turned out to be helpful, as well as have started to
extracted and stored in database manually. be acknowledged by the more extensive group of data
For testing purpose, different mood categories are given a finite investigators.[3]
range. Testing dataset are then compared with those ranges and Xindong Wu offered the leading ten data mining algorithms
the resulting category is displayed. named C4.5, k-Means, SVM, Apriori, EM, PageRank,
3). Gender and age recognition: Each individual experiences AdaBoost, kNN, Naive Bayes, and CART. These main 10
the procedure of maturing. Changes in our voices happen in algorithms are amongst the most powerful data mining
right on time adolescence and pubescence as well as in our algorithms in the exploration group. With each algorithm, we
grown-up lives into maturity. A considerable measure of give a depiction of the algorithm, examine the effect of the
acoustic features change with speaker’s age. Acoustic variety algorithm, and audit ebb and flow and further research on the
has been found in worldly and in laryngeals and algorithm
supralaryngeals adapted parts of speech Elderly individuals Shakir Khan, et al, it has been inspected that data mining
regularly talk slower than more youthful individuals; applications in security and their suggestions for protection.
notwithstanding, there is no distinction in explanation rate in They have inspected the thought of protection and afterward
the middle of youthful and old ladies amid read speech. discussed the advancements especially those on security
4 ) Research problems and obstacles :Following problems are preserving data mining. We then presented a plan for
there exploration on protection and data mining.[4]
• Children's speech is a great deal more troublesome JNeelamadhab Padhy, Dr. Pragnyaban Mishra, and Rasmita
than grown-up's speech in programmed speech Panigrahi3,2012, The survey of data mining applications and
acknowledgment. This issue is considerably more troublesome future scope,in this paper they quickly investigated the different
in light of small preparing data[8]. Notwithstanding, some data mining applications. This audit would be useful to analysts
methodologies exist which attempt to adjust for this downside. to concentrate on the different issues of data mining. In future
One remaining issue is the solid anatomic adjustment of the course, we will survey the different classification algorithms
vocal tract of youngsters inside a brief time of time. A thought and essentialness of developmental figuring (hereditary
to take care of this issue is to utilize diverse acoustic models for programming) approach in planning of proficient classification
distinctive age classes of kids. The most suitable acoustic algorithms for data mining.
model must be chosen before the programmed speech Cyril Laurie et all 2010, Multimodal music mood classification
acknowledgment can be performed. On the off chance that the using audio and lyrics checked on the different data mining
age of a kid is not known ahead of time, it can be predicted applications. This audit would be useful to scientists to
from the kid's voice[9] concentrate on the different issues of data mining. In future
• Challenging Sound One of the most critical difficulties course, we will audit the different classification algorithms and
in speaker acknowledgment comes from irregularities in the criticalness of developmental registering (hereditary
diverse sorts of sound and their quality. One such issue, which programming) approach in planning of productive classification
has been the centre of most research and productions in the algorithms for data mining.[7]

290
292
Dan young, Won-Sook Lee,2013, Music emotion identification
from lyrics here the outcomes for mining the lyrical content of
tunes for particular feeling are promising, produce
Spectral Centroid Overall
classification models that are humanly comprehensible, and
create results that compare to practical judgment skills instincts
Standard Deviation
about particular feelings. Mining lyrics concentrated in this 50
paper is one viewpoint a piece of examination which F 40
consolidates distinctive classifiers of musical feeling, for e V 30
example, acoustics and lyrical text. a a 20
IV. PROJECT DESIGN AND t l 10
Feature Value
IMPLEMENTATION AND RESULTS u u 0

Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
Spectral…
r e
Fig. 2.5. Overview of music mood identification e
system Build ground Social TagS Feature Name
truth data set

Fig.1 Spectral Centroid overall standard deviation

lyrics Audio Dataset Spectral centroid overall deviation is shown in shown in above
figure. Spectral Centroid Overall Standard Deviation is usually
less than Spectral Centroid Overall Average.

Compare Lyrics feature

types Emotion
Compare Classification Happy Mode Featurevalue
system (SVM)
Lyrics on audio data 50
Classification based to 47
45
system (Naïve audio
40 39
39
Bayes) on lyrical 38 37
33
30 31 31
28 29 29 3030283029 30
Feature
25 27 26
2424 24 25
2423 23 24 25
Value
20 222223 23212323
22 22 23 22
Hybrid System combing 2019 20 19
19 20 19
18 17 18 17 18 17 18
audio and lyrics 161717 16
13 12 13 14
13 14
10

0 0
OF RESULT
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65
Compare audio based and lyrics based systes

Fig.2 Happy Mode representation in Graphical form

This figure2 shows feature values of happy mode. From the

figure it can be seen that it changes abruptly only for few
values, otherwise it is in constant range. The figure3 shows
feature values of sad mode. From the figure it can be seen that
it’s feature values are less than happy mode.

291
293
 As mood classes in music mental models may do not have the
50 social connection of today's music listening environment, this
F 45
research inferred an arrangement of mood classifications from
40
e V social labels utilizing etymological assets and human skill. The
35
a a 30 resultant mood classes were contrasted with two delegate
25 models in music brain science. The outcomes demonstrate there
t l
20 Series2 were basic grounds between hypothetical models and
u u classifications got from observational music listening
15 Series1
r e 10 information, in actuality. While the mood classifications
e 5 recognized from social labels could at present be somewhat
0 bolstered by fantastic mental models, they were more thorough
1 6 11 16 21 26 31 36 41 46 51 56 61 and are all the more nearly associated with the truth of music
Feature Name: Sad tuning in.
Precision of Mood Plotting the music set of statistics castoff
for research the mood mapper is made up of fourteen western
Fig.3 Happy Mode representation in Graphical form classical songs. Net Support vector regression is realized
grounded on the library which gives complete functionalities
for Support vector regression training, which might lead to a
50 more accurate understanding of the music mood in the mood
F
40 42 mapping process The library comprises of diverse Support
e V 37 vector regression/support vector machine engines using
a a 30 32 compound strategy of support vector regression training
2426 26
23 222321 21 22 limitations and found optimums Name of first parameters is 2-
20 22
t l 21
2020 1819 20 1818 19 18 19 18 5 ~ 2 - 0.1 2-1.7 As a result, the proposed system obtained 85%
16 151515 17
u u
13 1155 13 14 14 16 13 accuracy from our expert mood mapper,.
r 11
e 10 11 10 11 10
REFERENCES:
e 0 [1] R. Cai, C. Zhang, C. Wang, L. Zhang, and W.-Y. Ma.,
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 "Musicsense: Contextual Music Recommendation Using Emotional
Allocation," Proc. 15th Int'l Conf. Multimedia, pp. 553-556, 2007
Feature Name: Angry,Neutral,Romantic [2] M. Tolos, R. Tato, T. Kemp. Mood-based navigation through large
collections of musical data. In 2nd IEEE Consumer Communications
and Networking Conference (CCNC 2005), 2005, 71-75
[3] D. Yang, W. Lee. Music emotion identification from lyrics.
Fig.4 Sad, angry, neutral, romantic express in graphical form Proceedings of the 2009 11th IEEE International Symposium on
This figure shows comparison of feature values of all modes. Multimedia, 2009, 624-629
[4] K. Bischoff, C. S. Firan, R. Paiu et al. Music mood and theme
V. CONCLUSION
classification-A hybrid approach. 10th International Society for Music
This research paper scrutinizes the value of content features and Information Retrieval Conference (ISMIR 2009), 2009, 657-662
audio features in music mood order on eighteen mood [5] P. N. Juslin and P. Laukka: “Expression, Perception,and Induction
arrangements or classifications got from client labels. of Musical Emotions: A Review and Questionnaire Study of
Contrasted with Part-of-Speech and capacity words, Bag-of- Everyday Listening,”Journal of New Music Research, Vol. 33, No. 3,
Words are still the most valuable highlight sort. pp. 217-238, 2004.
Notwithstanding, there is no huge distinction between the [6] P. N. Juslin and P. Laukka: “Expression, Perception,and Induction
decision of stemming or not stemming, or among the four of Musical Emotions: A Review and A Questionnaire Study of
content representations all things considered exactness’s over Everyday Listening,”Journal of New Music Research, Vol. 33, No. 3,
pp.217-238, 2004.
all classifications. Specifically lyric features alone can outflank
[7] Y.-H. Yang, Y.-C. Lin, H.-T. Cheng, I.-B. Liao, Y.-C. Ho, and H.
audio features in classes where tests are more inadequate or H. Chen: “Toward multi-modal musicemotion classification”,
when semantic implications taken from lyrics attach well to the Proceedings of Pacific RimConference on Multimedia (PCM’08),
mood classification. Likewise, consolidating lyrics and audio 2008
features enhances exhibitions on most, however not all, classes. [8] Sebastiani, F., Text categorization. In Zanasi, A., editor, Text
Trials on three distinctive highlight determination routines Mining and its Applications, pages 109–129, Southampton, UK: WIT
exhibited that an excess of content features are in fact repetitive Press, 2005.
or boisterous and joining audio with the most notable content [9] Witten, I.H., Frank, E., Data Mining: Practical Machine Learning
features may prompt higher correctness’s for most mood Tools and Techniques. Morgan Kaufmann Publishers, 2nd edition,
2005.
classifications.[9]

292
294