A COMPARATIVE STUDY OF INDIAN AND WESTERN MUSIC FORMS

Parul Agarwal1 , Harish Karnick2

Bhiksha Raj
Indian Institute of Technology, Kanpur, India
Carnegie Mellon University, USA
parulagarwal89@gmail.com1
bhiksha@cs.cmu.edu
hk@cse.iitk.ac.in2

ABSTRACT multiple genres. For example Hindustani music has Dhru-

pad, Khayal, Tarana as classical genres and Thumri, Dadra,
Music in India has very ancient roots. Indian classical Tappa, Bhajan as semi-classical genres - amongst others.
music is considered to be one of the oldest musical tra- Popular music has multiple folk genres based on region,
ditions in the world but compared to Western music very film music and adhunik or modern music which is influ-
little work has been done in the areas of genre recognition, enced by multiple Indian and western genres.
classification, automatic tagging, comparative studies etc. At the base of classical and semi-classical Indian mu-
In this work, we investigate the structural differences be- sic is a raga(s) which is described as a mood or senti-
tween Indian and Western music forms and compare the ment expressed by a microtonal scale form [28]. A raga
two forms of music in terms of harmony, rhythm, micro- can be sung in many Indian genres like Dhrupad, Khayal,
tones, timbre and other spectral features. To capture the Thumri etc with its corresponding microtonal scale based
temporal and static structure of the spectrogram, we form on a natural harmonic series. Popular Indian music, on
a set of global and local frame-wise features for 5- genres the other hand, is not necessarily based on ragas and can
of each music form. We then apply Adaboost classification have tonal elements that can differ from the traditionally
and GMM based Hidden Markov Models for four types of accepted classical norms. Almost all Indian music is based
feature sets and observe that Indian Music performs bet- on melody - single notes played in a given order. Western
ter as compared to Western Music. We have achieved a music has strong harmonic content i.e. a group of notes
best accuracy of 98.0% and 77.5% for Indian and Western called chords played simultaneously. Unlike Indian tonal
musical genres respectively. Our comparative analysis in- system, the Western tonal system is divided into twelve
dicates that features that work well with one form of music equal intervals. Each Indian genre has its own well-defined
may not necessarily perform well with the other form. The structure which makes it different from other forms of mu-
results obtained on Indian Music Genres are better than the sic. To capture the inherent structure of Indian music and
previous state-of-the-art. use it for genre classification is a challenging problem.
Through our work we try to explore this area and focus
on the following research questions:
1. INTRODUCTION
Due to technology advances the size of digital music col- • What are the structural differences between Indian
lections have increased making it difficult to navigate such and Western musical forms?
collections. Music content is very often described by its
genre [1], though the definition of genre may differ based • How well can the audio features used for genre clas-
on culture, musicians and other factors. Considerable anal- sification in Western music, capture the characteris-
ysis has been done on western music for genre classifica- tics of Indian genres?
tion [2] [3] and content analysis [4].
Although, Indian music, especially classical music, is • Is there a feature set that can be used for cross-cultural
considered to be one of the oldest musical traditions in music genre classification? In this case, for Indian
the world, not much computational analytical work has and Western music.
been done in this area. Indian music can be divided into
two broad categories classical and popular. Classical mu- 2. RELEVANT WORK
sic has two main variants Hindustani classical prevalent
largely in north and central India and Carnatic classical 2.1 Study on Genre classification by humans
prevalent largely in the south of India. Each variety has
Research has been done on the human ability to classify
music genres. Perrot et al. [16](1999) reported that hu-
Permission to make digital or hard copies of all or part of this work for mans with little musical training were able to classify gen-
personal or classroom use is granted without fee provided that copies are res with an accuracy of about 72% on a ten-genre dataset of
not made or distributed for profit or commercial advantage and that copies a music company, based on only 300 milliseconds of audio.
bear this notice and the full citation on the first page. Soltau (1997) conducted a Turing-test in which he trained
c 2013 International Society for Music Information Retrieval. 37 people to distinguish rock from pop music and found
that his automatic music classifier was similar to their per- 2.4 Contributions of the paper
formance on the same dataset. Lippens et al. [15](2004)
There is no study, to the best of our knowledge, that com-
conducted an experiment with six genres and observed that
pares classification of Indian musical genres with West-
people were able to classify with an accuracy of around
ern musical genres. Our work makes the following con-
90%. Considerable work has been done in the area of
tributions: first, this paper introduces a new research prob-
automatic music genre classification and with increasing
lem on cross-cultural music genre classification by doing a
improvement in classification methods there is a distinct
comparative study on Indian and Western musical forms in
possibility that they may outperform humans in the near
terms of their structure and characteristics. Second, while
future.
there are many standard datasets on genre classification
available for Western music like GTZAN dataset [5], 2005
2.2 Genre classification of Western Music MIREX dataset, Magnatune [21], USPOP [22] etc. There
is no standard dataset for Indian genre classification. For
Many supervised and unsupervised methods have been used our study, we built a dataset of 5 Indian genres- Dhru-
to automatically classify recordings into genres. Commonly pad, Thumri, Carnatic, Punjabi and Ghazals. There are
used supervised classification methods are k-nearest Neigh- 100 samples in each genre where each sample is an audio
bours (KNN) [5] [17], Support Vector Machines (SVMs) clip of 30 seconds length. This dataset can be used for
[19] [17], Gaussian Mixture Models (GMMs) [5] [20] [17], future work on Indian music genre classification. Third,
Linear Discriminant Analysis (LDA) [17] [18], non-negative we have incorporated short-time and long-time features to
Matrix Factorization (NMF) [2], Artificial Neural Networks capture the static and temporal structure of the spectro-
(ANNs) and Hidden Markov Models (HMMs). In the un- gram and have analysed both cultural forms. Fourth, we
supervised domain, there is work using the k-means ap- propose a novel approach of using timbre and chromagram
proach and hierarchical clustering. features with Gaussian Mixture Models(GMM) based Hid-
To the best of our knowledge, there is no work in the den Markov Models to identify the patterns in Indian music
literature that analyses which features are best for music which gave an overall accuracy of 98.0%, which is better
genre classification. There is work using Fourier analy- than the previous state-of-the-art [12] which also worked
sis, cepstral analysis and wavelet analysis [17] [24]. Li with 5-genres.
et al. [17] and J. Bergstra et al. [3] worked with a set of
parameters like spectral centroid, spectral rolloff, spectral 2.5 Outline of Paper
flux, zero crossings and low energy. Y.Panagakis et al. [25]
used sparse representation-based classifiers on audio tem- In section 3 we discuss the dataset used for the experiments
poral modulations of songs for genre classification. Sala- and the structure and characteristics that make an Indian
mon et al. [26] used a combination of low level features genre distinct from other genres. Sections 4 and 5 give a
with high level melodic features derived from pitch con- detailed explanation of the features chosen and classifiers
tours and showed that the combination outperforms the re- used. Section 6 discusses the experimental results. Section
sults from using only low level features. 7 outlines possible future work.
Tzanetakis et al. [5] proposed to classify musical genres
using K -Nearest Neighbours and GMMs based on timbral 3. DATASET
texture and rhythmic features. E. Benetos et al. [2] used
In this section, we look at the structural differences be-
non-negative tensor factorization (NTF) with GTZAN data
tween Indian and Western music forms and discuss the
set [5]. J. Bergstra et al. [3] used GMM based Adaboost
characteristic features of the Indian genres used in the ex-
classification with aggregate features dividing frames into
periment.
segments. Hidden Markov Models were used by T. Lan-
glois et al. [10] for genre classification based on timbre
3.1 Indian Music
information.
For the experiments we have considered five popular gen-
res - Hindustani (Dhrupad + Thumri), Carnatic, Folk (Pun-
2.3 Genre classification of Indian Music
jabi) and Ghazal. We constructed our own dataset because
Compared to western music very little work has been done no standard dataset is available. Each genre contains 100
in Indian musical genre classification. S.Jothilakshmi et audio recordings each 30 seconds long extracted randomly
al. [12] used kNN and GMM on spectral and perceptual from recordings available on-line. All samples are vocals.
features and achieved an accuracy of 91.23% with 5 In-
3.1.1 Indian classical Music
dian genres- Hindustani, Carnatic, Ghazal, Folk and Indian
western. P. Chordia et al. [13] used pitch profiles and Pitch- Indian classical music is based on melody without the pres-
Class Distributions for raga identification. A. Vidwans et ence of any major harmonic structure. It has seven basic
al. [14] extracted melodic contours to classify Indian clas- notes with five interspersed half- notes, resulting in a 12-
sical vocals. They also performed a listening test in which note scale. In Indian music any frequency can be set as
around 85% samples were correctly identified by people the base frequency, known as swara(note) ’Sa’ with other
without any training. notes taking frequency values following their fixed ratios
with respect to Sa. Indian classical music has three im- the raga without any rhythmic accompaniment. There are
portant characteristics- raga (melodic aspect of music) and niruval and kalpana swara(notes) that provide opportuni-
taal (cycle of fixed number of beats repeated over and over ties to improvise. In the dataset, about 55% of the samples
as shown in Figure 1) and a drone(a sustained note). Indian are sung by female singers.
popular music have the the melodic (though not raga) and Hindustani and Carnatic music differs mainly in the fol-
taal/beat components but do not usually have the drone. lowing characteristics-
• Hindustani music like Dhrupad gradually builds up
tempo from very slow to very fast while in Carnatic
music there is a constant and fairly fast tempo through-
out.
• In Hindustani music notes are held longer and are
mostly approached and released by ornaments or small
ornamental phrases whereas in Carnatic music notes
are not held for long and usually released by a char-
acteristic oscillation using indeterminate pitch.
• Hindustani music has more improvisation around the
raga unlike Carnatic which has a some what more
rigid structure and is composition based.

Figure 1. Repetition of beats (taal) in Indian Classical 3.1.2 Folk- Punjabi

Punjabi music has a smaller range which seldom extends
Dhrupad: Dhrupad is a very old form of Hindustani beyond an octave. It is accompanied by a lively rhythm
music that is primarily devotional in content. It has two (chaal) and is usually based on kaherava taal, a cycle of
major parts- Alap (sung without words) and Dhrupad (fixed eight beats. Greater emphasis of the music is on the lyrical
composition part) accompanied by Pakhawaj, a two headed nature of the songs. The characteristic which makes it dif-
barrel drum. A performance begins slowly with alap where ferent from other music forms is the use of Dhol- a barrel
usually the syllables of the following mantra is recited: shaped drum that gives a bass sound. Samples are taken
”Om Anant tam Taran Tarini Twam Hari Om Narayan from Folk N Duet 2010 Album contributed by 16 singers
Anant Hari Om Narayan”. In alap an artist develops each which includes 3 female singers.
note and uncovers the personality of the chosen raga. The
tempo gradually increases and an alap ends after exploring 3.1.3 Ghazal
three octaves. Then begins the dhrupad in which the artist Ghazal is defined as a poetic genre and consists of a se-
is joined by the Pakhawaj. It is usually set to the following ries of rhymed verses, each symmetrically divided into two
talas/taal: tala chautal (12 beats), tivra (7 beats), sulfak (10 half-verses. The structure of a Ghazal is defined by its me-
beats), matt (9 beats), or farodast (14 beats). In the dataset tre and rhyme which are maintained throughout the poem.
samples were extracted from dhrupad performances where Samples were extracted from ghazal performances where
about 25% of them were by female artists. 19% of total pieces were sung by female singers.

Thumri: Thumri is a semi-classical form of Hindustani 3.2 Western Music

music which is romantic and devotional in nature. The text
revolves around a girl’s love for Lord Krishna. The lyrics For western music we considered the GTZAN dataset [5].
are written in dialects of Hindi called Awadhi and Brij It consists of 10 distinct genres- Blues, Classical, Country,
Bhasha. Thumri has greater flexibility and more liberties Disco, Hip hop, Jazz, Metal, Pop, Reggae and Rock. Each
are taken with the raga as compared to Dhrupad or Khayal. genre is represented by 100 samples of 30 seconds length
The compositions are usually of 8 beats (kaherava), 14 each. Since we have 5 genres for the Indian music, we
beats (dipchandi) or of 16 beats (addha tala). Samples considered only top 5 genres (Classical, Hiphop, Jazz, Pop
were extracted from 10 Thumri performances where 3 were and Reggae) from the GTZAN dataset based on the results
by female artists. of [5].

Carnatic: Carnatic music is the classical music of South- 4. FEATURE EXTRACTION

ern India. Like Hindustani music, it is based on ragas and
For feature extraction, we used the MIRToolbox [27], an
taals. Performances in Carnatic music are divided into
open source toolbox for musical feature extraction. For
a number of sections. It starts with a varanam, literally
the experimental set-up we used three types of feature sets
meaning ”a description”, which unfolds the important fea-
as described below:
tures of the raga being performed. Then comes the fixed
composition in the raga, known as kritis. The main com-
4.1 Feature set 1- Frequency of chromagram notes
position is alapana, which explores the raga and uses the
sounds like aa, ri, na, ta etc to elaborate the notes. This A chromagram is a redistribution of the spectrum energy
starts slowly and finally builds a complicated exposition of along the different pitch levels (or chromas). We have con-
sidered 12 pitch classes- C, C#, D, D#, E, F, F#, G, G#, A, delta-delta-MFCC, 11 spectral features (zero-crossing rate,
A# and B. We used a framesize of 100 milli-seconds with rolloff, brightness, centroid, spread, skewness, kurtosis,
50% overlap. The choice of a large framesize is to cover flatness, entropy, roughness and irregularity). For the chro-
for the semitone energies in lower frequencies. For each magram features, we first perform a logarithm transform
frame, we assigned the note that has the maximum energy and then a discrete cosine transform (DCT) which essen-
in the frame’s 12-dimensional chromagram vector; we call tially decorrelates the features. We use these modified chro-
this the ”dominant note”. The use of the dominant note in magram features with the above mentioned features to get
each frame captures the melodic information as the melody a vector of 62 features for each frame. A sample is now
note is the one that typically dominates the feature. Con- represented by a 2-D matrix of size 62 × f rames, where
catenating the dominant note in each frame gave us a note frames are the number of frames in the sample. These fea-
sequence representing the sample. We then formed a 12- tures carry the temporal information of the spectrogram.
dimensional vector of the normalized frequencies (counts)
of the notes in the note sequence. This feature set is moti- 5. EXPERIMENTAL RESULTS
vated by the repetition of notes in Indian music and the use
of chromagrams in [6] and [7]. We performed four experiments based on the above three
feature sets. For each run, we randomly divided the dataset
4.2 Feature set 2- Global features into 80% train and 20% test.

We extract a set of global features based on the main di- 5.1 Experiment 1: Adaboost on Feature set 1
mensions of music which includes melody, rhythm, timbre
and spatial features as described in [8] [9] and [10]. We We train One-Vs-One Adaboost classifiers. A sample is
used the following features to form a vector of 86 global assigned the class which gets the maximum votes. In case
features- of a tie, the class with the lower index is the predicted la-
• Energy features- mean and variance of the energy bel. Table 1 shows the confusion matrix for 10 runs of the
envelope, Root Mean Square (RMS) energy, low- experiment for Indian Music and Western Music. Over-
energy rate(percentage of frames having less than all the accuracies obtained for Indian Music and Western
average energy). Music are 58.70% and 46.90% respectively.
• Rhythm features- mean and variance of notes on-
set time (successive bursts of energy indicating es- Table 1. Confusion Matrix (in %) for Exp 1
timated positions of the notes), event density (aver-
(a) Indian Music
age frequency of events, i.e., the number of note on- Ca Dh Gh Pu Th
sets per second), tempo, pulse clarity [11](strength Ca 66.50 9.00 5.50 6.50 12.50
of beats). Dh 10.50 64.00 6.00 9.00 10.50
• Pitch features- mean and variance of pitch. Gh 16.50 6.50 64.50 7.00 5.50
Pu 18.50 12.50 14.00 47.50 7.50
• Tonality features- 12-chromagram pitch class ener-
Th 19.50 6.50 13.00 10.00 51.00
gies, 24- key strength major and minor (cross cor- Legend: Ca:Carnatic, Dh:Dhrupad, Gh:Ghazal, Pu:Punjabi, Th:Thumri
relation of the chromagram), 6- dimensional tonal
(b)Western Music
centroid vector from the chromagram (corresponds Cl Hi Ja Po Re
to projection of the chords along circles of fifths, of Cl 60.00 8.00 13.50 12.50 6.00
minor thirds, and of major thirds). Hi 3.50 58.50 12.00 14.00 12.00
• Timbre features- mean and variance of attack time Ja 8.50 21.00 48.00 13.00 9.50
of notes onset, rolloff frequency, brightness (percent- Po 9.50 18.50 22.00 32.00 18.00
Re 14.00 17.50 16.50 16.00 36.00
age of energy above 1500Hz), 13 Mel-Frequency Legend: Cl:Classical, Hi:Hip-hop, Ja:Jazz, Po:Pop, Re:Reggae
Cepstral Coefficients (MFCCs), roughness(sensory
dissonance), irregularity (degree of variation of the
successive peaks of the spectrum)
• Spectral-shape features- zero-crossing rate, spread, 5.2 Experiment 2: Adaboost on Feature set 2
centroid, skewness, kurtosis, mean and variance of The Adaboost classification technique is used in a manner
Inverse Fourier Transform of logarithm of spectrum, similar to Experiment 1. The confusion matrices for Indian
flatness, mean and variance of spectrum, mean and music and Western music are shown in Table 2. The overall
variance of spectral flux, mean and variance of spec- accuracies obtained are: Indian music- 87.30%, Western
trum peaks. music- 74.50%.
These features are taken to capture the static structure
of the spectrogram. 5.3 Experiment 3: Adaboost on Feature set 1+2
Adaboost classification technique is used similar to Exper-
4.3 Feature set 3- Frame-wise features
iment 1. Here we take a combination of feature set 1 and
We analyse each song using a framesize of 100 milli sec- feature set 2 for the training and testing. Table 3 shows the
onds with 50% overlap and extract the following features: confusion matrices. The overall accuracies achieved are:
12-chromagram features, 13 MFCC, 13 delta- MFCC, 13 Indian music- 87.80% , Western music- 77.50%.
 Table 2. Confusion Matrix (in %) for Exp 2 Table 4. Confusion Matrix (in %) for Exp 4
(a) Indian Music (a) Indian Music
Ca Dh Gh Pu Th Ca Dh Gh Pu Th
Ca 89.50 1.50 5.00 0.50 3.50 Ca 95.00 0 3.00 2.00 0
Dh 5.00 90.00 2.50 0 2.50 Dh 0 99.00 1.00 0 0
Gh 5.50 6.50 81.00 1.50 5.50 Gh 1.00 0 98.50 0.50 0
Pu 3.00 0 2.00 95.00 0 Pu 0 0.500 0 99.50 0
Th 10.50 5.00 3.50 0 81.00 Th 1.00 0 1.00 0 98.00
Legend: Ca:Carnatic, Dh:Dhrupad, Gh:Ghazal, Pu:Punjabi, Th:Thumri Legend: Ca:Carnatic, Dh:Dhrupad, Gh:Ghazal, Pu:Punjabi, Th:Thumri

(b)Western Music (b)Western Music

Cl Hi Ja Po Re Cl Hi Ja Po Re
Cl 87.50 0 10.50 0 2.00 Cl 85.00 0 10.00 0 5.00
Hi 0.50 69.50 0.50 14.00 15.50 Hi 0 75.00 5.00 15.00 5.00
Ja 9.00 2.00 78.50 0 10.50 Ja 25.00 15.00 55.00 5.00 0
Po 1.50 10.50 3.00 74.50 10.50 Po 0 20.00 5.00 65.00 10.00
Re 4.50 18.50 4.00 10.50 62.50 Re 0 15.00 0 30.00 55.00
Legend: Cl:Classical, Hi:Hip-hop, Ja:Jazz, Po:Pop, Re:Reggae Legend: Cl:Classical, Hi:Hip-hop, Ja:Jazz, Po:Pop, Re:Reggae

Table 3. Confusion Matrix (in %) for Exp 3 offer a variety of cues that do not seem to be present in
(a) Indian Music Western music.
Ca Dh Gh Pu Th
Ca 90.00 1.50 4.00 0.50 4.00
Dh 4.50 91.00 2.00 0 2.50
Gh 6.50 5.50 81.50 2.00 4.50
Pu 3.00 0 2.00 95.00 0
Th 7.50 3.50 7.50 0 81.50
Legend: Ca:Carnatic, Dh:Dhrupad, Gh:Ghazal, Pu:Punjabi, Th:Thumri

(b)Western Music
Cl Hi Ja Po Re
Cl 90.50 0 9.00 0.50 0
Hi 0.50 79.00 1.00 12.50 7.00
Ja 10.50 2.00 78.00 0.50 9.00
Po 0.50 15.00 1.50 72.50 10.50
Re 3.50 11.50 9.50 8.00 67.50
Legend: Cl:Classical, Hi:Hip-hop, Ja:Jazz, Po:Pop, Re:Reggae

5.4 Experiment 4: HMM with GMM based on Figure 2. Comparison of performance of Indian and West-
Feature set 3 ern Genres on four experiments
We train Hidden Markov Models on Feature set 3 with 62
hidden states and full covariance matrix. The emissions of The better performance of Indian music over Western
all states were randomly initialized and the states need not music on similar experiments answers the second question
represent exactly the underlying feature. The algorithms of our research problem which suggests that the state-of-
uses maximum likelihood parameter estimation using Ex- the-art features (spectral, timbre, harmony etc) used for
pectation Maximization. The state emissions were 62 di- Western music genre classification can be applied to Indian
mensional chroma and timbre data modelled by Gaussian music where they show higher accuracies.
mixtures with 8 Gaussians. We used the Hidden Markov In Indian music, the classical forms (Carnatic and Dhru-
Model Toolbox of Matlab by [23] for the same. The confu- pad) performed better than Ghazal. This is may be be-
sion matrix for Indian music and Western music are shown cause the Indian classical music is more structured as it has
in Table 4. Overall accuracies of 98.00% and 67.00% were 3 components of raga (melody), taal (rhythm) and drone
obtained for Indian and Western music respectively. (sustained note) and has a strong melodic structure. In last
three experiments Punjabi music has lowest confusion be-
cause its two characteristics features: lively rhythm and
6. DISCUSSION
distinct timbre due to bass sound produced by Dhol; are
Our experiments performed better on Indian music than on accounted for in features used. In Western music, Classi-
Western music for given classification techniques and sets cal and Jazz performed better than other genres, with Reg-
of features (Figure 2). The differences in performance of gae performing the worst. These results are consistent with
the two cultural forms of music can perhaps be traced to the G.Tzanetakis et al. [5].
more well-defined structural form and strong melodic con- Performance on Feature set 1 (which is just based on
tent of Indian Music. In Indian music, melody and rhythm the frequency of chromagram notes) for Indian music and
Western music is about 59% and 47% respectively. This [8] N. Scaringella, G. Zoia, and D. Mlynek, “Automatic Genre
is expected as the Indian music is based on melody, i.e. a Classification of Music Content: A Survey“ IEEE Signal
sequence of notes played in a given order. Whereas, the Processing Magazine 2006, 23(2), pp. 133141.
Western music is more harmonic in nature, i.e. group of [9] Kaichun K. Chang, Jyh-Shing Roger Jang, and Costas S. Il-
iopoulos. ”Music Genre Classification via Compressive Sam-
notes played simultaneously, which is not captured by a pling“ 11th International Society for Music Information Re-
single dominant note in Feature set 1. trieval Conference (ISMIR 2010) 2010, pp. 387-392.
In Feature set 2 we have used the 12-chromagram pitch [10] Thibault Langlois, and G. Marques ”A Music Classifica-
class energies which tries to capture the dominance order tion Method Based on Timbral Features” 10th International
of notes. The dominance order of notes is also represented Society for Music Information Retrieval Conference (ISMIR
2009), 2009, pp. 81-86.
by the frequency of chromagram notes in Feature set 1.
[11] Olivier Lartillot, Tuomas Eerola, Petri Toiviainen, Jose
Thus, there is no significant difference in accuracies of Ex- Fornari ”Multi-feature modelling of pulse clarity: Design,
periment 2 (using Feature set 2) and Experiment 3 (using validation, and optimization” International Conference on
Feature set 1+2). Music Information Retrieval Philadelphia, 2008
The highest accuracy of 98.0% for Indian music genres [12] S.Jothilakshmi,N.Kathiresan ”Automatic Music Genre Clas-
in Experiment 4 is better than S.Jothilakshmi et al. [12] and sification for Indian Music” ICSCA 2012
the state-of-the-art to the best of our knowledge. The ma- [13] P. Chordia and A. Rae ”Raag recognition using pitch-class
jor difference in accuracies between local frame-wise fea- and pitch-class dyad distributions” Proc. of ISMIR, 2007
pp.431436.
tures and global features in case of Indian music may be
[14] A.Vidwans, K.K. Ganguli and Preeti Rao ”Classification of
because the local frame-wise features are better in captur- Indian Classical Vocal Styles from Melodic Contours” Proc.
ing the characteristics of Indian music like raga notes and of the 2nd CompMusic Workshop, Istanbu, Turkey, July 12-
taal (repetition of beats) which require small size windows 13, 2012
to be analysed. An accuracy of 77.5% for Western music [15] J. Martens Lippens, S. and T. De Mulder ”A comparison of
genres in Experiment 3 is better than G.Tzanetakis et al. [5] human and automatic musical genre classification” In IEEE
on the same dataset of five genres (Classical, Hiphop, Jazz, International Conference on Acoustics, Speech, and Signal
Processing, volume 4, pages 233236, 2004
Pop and Reggae).
[16] D. Perrot and R. R. Gjerdigen. ”Scanning the dial: an ex-
ploration of factors in the identification of musical style” In
Proceedings of the 1999 Society for Music Perception and
7. FUTURE WORK
Cognition, 1999.
This work can be extended in various ways: forming a [17] M. Ogihara Li, T. and Q. Li. ”A comparative study on
‘golden-set’ of features that are genre-specific like rhyme content-based music genre classification” In Proceedings of
the 26th annual international ACM SIGIR conference on
in Ghazal, beats in Folk Punjabi etc.; recognition of pat- Research and development in information retrieval, pages
terns like taal in Indian music and chords in Western mu- 282289, 2003.
sic; expansion of classes in terms of genres and sub-genres [18] West, K. and S. Cox, ”Features and classifiers for the auto-
so that we can work with more classes in both datasets matic classification of musical audio signals” Proc. 5th In-
(GTZAN has 10-genres); studying music forms of other ternational Conference on Music Information Retrieval (IS-
MIR), 2004
cultures for example Chinese and Japanese and comparing
them with Indian and Western genres. [19] Michael I Mandel and Daniel P.W.Ellis. ”Song-level features
and support vector machines for music classification” In In-
ternational Society for Music Information Retrieval, 2005.
8. REFERENCES [20] A. Flexer Pampalk, E. and G. Widmer ”Improvements of
audio-based music similarity and genre classification” In
[1] J. J. Aucouturier, F. Pachet. ”Representing musical genre: A Crawford and Sandler, 2005.
state of the art” Journal of New Music Research, pp. 83-93, [21] www.magnatune.com (Access date: 5 May 2013)
2003.
[22] www.ee.columbia.edu/ dpwe/research/musicsim/ (Access
[2] E. Benetos, C. Kotropoulos ”A tensor-based approach for au- date: 5 May 2013)
tomatic music genre classification” Proc. EUSIPCO,2008 [23] http://www.cs.ubc.ca/ murphyk/Software/HMM/hmm.html
[3] J. Bergstra, N. Casagrande, D. Erhan, D. Eck, B. Kegl ”Ag- (Access date: 5 May 2013)
gregate features and Adaboost for music classification” Ma- [24] C. Marques, I. R. Guilherme, R. Y. M. Nakamura and J. P.
chine Learning vol. 65, no. 2-3, pp. 473-484, 2006 Papa ”New trends in musical genre classification using opti-
[4] S. Sukittanon, L. E. Atlas,J. W Pitton ”Modulation-scale mum path forest” ISMIR, 2011.
analysis for content identification” IEEE Trans. Signal Pro- [25] Yannis Panagakis, Constantine Kotropoulos, Gonzalo R.
cessing vol. 52, no. 10, pp 3023-3035, Oct. 2004 Arce ”Music genre classification via sparse representations
of Auditory Temporal Modulations” EUSIPCO,2009.
[5] George Tzanetakis and Perry Cook “Music Genre Classifi-
cation of Audio Signals” IEEE Transactions on Speech and [26] Justin Salamon, Bruno Rochay and Emilia Gomez ”Music
Audio Processing vol. 10, no. 5, pp. 293-302, 2002. genre classification using melody features extracted from
polyphonic music signals” ICASSP,2012.
[6] D. P. W. Ellis and G. E. Poliner. “Identifying cover songs with
[27] Olivier Lartillot, Petri Toiviainen ”A Matlab Toolbox for Mu-
chroma features and dynamic programming beat tracking”
sical Feature Extraction From Audio” International Confer-
ICASSP Hawaii, USA 2007.
ence on Digital Audio Effects, Bordeaux, 2007
[7] S. Kim and S. Narayanan “Dynamic chroma feature vec- [28] ”Raga” http://www.britannica.com/EBchecked/topic/489518
tors with applications to cover song identification” In MMSP (Access date: 5 May 2013)
Cairns, Australia, 2008.