See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228960849

Raga Identification of Carnatic music for music Information Retrieval

Article · April 2009

CITATIONS READS

53 5,085

2 authors:

Rajeswari Sridhar T.V Geetha

National Institute of Technology Tiruchirappalli Anna University, Chennai
43 PUBLICATIONS   154 CITATIONS    118 PUBLICATIONS   479 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cross Lingual Information Access View project

Cross-Lingual Information Access View project

All content following this page was uploaded by T.V Geetha on 21 May 2014.

The user has requested enhancement of the downloaded file.

SHORT PAPER
International Journal of Recent Trends in Engineering, Vol. 1, No. 1, May 2009

Raga Identification of Carnatic music for Music

Information Retrieval
Rajeswari Sridhar1, and T.V. Geetha2
1, 2
Department of Computer Science and Engineering, Anna University, Chennai 600025, India
1
Email: [email protected], 2Email: [email protected],

Abstract— In this work we propose a method to identify the ascending order of the arrangement of notes and
raga of a Carnatic music signal. Raga is analogous to melody in avarohanam is defined as the descending order
Western music but is much richer and stronger in its arrangement of notes. A raga is classified into melakarta
representation. The main motive behind Raga identification is raga (parent) and janya raga (child). The various notes in
that it can be used as a good basis for music information
retrieval of Carnatic music songs or Film songs based on
Carnatic music are S, R, G, M, P, D, N which can be
Carnatic music. The input polyphonic music signal is analyzed thought of as analogous to C, D, E, F, G, B, A. Every
and made to pass through a signal separation algorithm to note has a frequency associated with it as given in [2]. A
separate the instrument and the vocal signal. After extracting the melakarta Ragam will have all the seven swaras and a
vocal signal we segment the vocal signal using our proposed janya ragam will have atleast 5 of the seven swaras [1].
segmentation algorithm. Using our proposed singer Every song in Carnatic music has an associated Talam.
identification algorithm we determine the singer to know the There are essentially 175 talams as explained in [1]. In
fundamental frequency of the singer. The frequency Carnatic music any song will be of time duration which is
components of the signal are then determined and we map these an integer multiple of a Talam. A talam is analogous to
frequency components into the swara sequence and thereby
determine the Raga of the particular song, which could be used
the beat of European music and the seven basic talams is
to index the songs and further for retrieval based on the Raga. discussed in [2]. It indicates the pacing of the music and
the placement of syllables in the composition. In Carnatic
Index Terms—Carnatic music, Raga, Music information music, the singer indicates the talam using hand gestures
retrieval [2].
Usually songs, which are based on South Indian
I. INTRODUCTION Carnatic music, will belong to one of the several Ragas or
a combination of Ragas and one of the Talam. Hence it is
Music information retrieval is the need of the hour due a good idea to index songs, which are based on the
to the availability of large amount of music on the
Carnatic music characteristics using these components,
Internet. Music information retrieval can be thought of as Raga and Talam. Hence, in this work we have proposed
part of Multimedia information retrieval. A lot of work an algorithm to determine the ragam and talam of a
has been done in the other components of multimedia like
Carnatic music song and have proposed that these
text, video and the one that is yet to be fully developed is components could be used as features for indexing and
audio. Audio processing involves processing speech and later for retrieval.
music. In this paper we discuss music processing which
could be used as the basis of music information retrieval
3. EXISTING WORK
using music characteristics. This paper is organized as
follows: Section 2 discusses about the Characteristics of Literature work in Carnatic music retrieval is on a slow
Carnatic music, Section 3 discusses on Existing work in pace compared to Western music. Some work is being
Raga and melody identification, Section 4 discusses on done in Swara identification [2] and Singer identification
the proposed System Architecture, Section 5 discusses on [3] of Carnatic music. In Hindustani music work has been
Results and Evaluation and Section 6 concludes the done in identifying the Raga of Hindustani music [4]. In
paper. [4] the authors have created a HMM based on which they
have identified two ragas of Hindustani music. The
2. CHARACTERISTICS OF CARNATIC MUSIC fundamental difference between Hindustani Raga pattern
and Carnatic Raga pattern is that in Hindustani we have
Indian music is broadly classified into South Indian R1, R2 as against R1, R2, R3 in Carnatic. Similarly G, D,
Carnatic music and North Indian Hindustani music. Both N all has three distinct frequencies in Carnatic music as
the systems of music are rich in their own style and compared to two frequencies in Hindustani [5]. This
Carnatic music is much more complex in the way the reduces the confusion in identifying the distinct
notes are arranged and rendered [1]. Indian music in
frequencies in Hindustani music as compared to Carnatic
general is based on Raga and Talam. Raga and Talam can music. The authors have not used polyphonic music
be thought of as equivalent to melody and rhythm in signal and have assumed that the input music signal is a
Western music. Raga is much more complex than melody
voice only signal. The fundamental frequency of the
and scale in western music. Raga can be thought of as the signal was also assumed and based on these features the
sequential arrangement of notes that is capable of raga identification process was done for two Hindustani
invoking the emotion of a song. A note is called as a
ragas. On the western music aspect, melody retrieval is
Swara in Carnatic music. Raga is characterized by being performed by researchers. The one proposed by [6]
Arohanam and Avarohanam. Arohanam defines the
571

© 2009 ACADEMY PUBLISHER

SHORT PAPER
International Journal of Recent Trends in Engineering, Vol. 1, No. 1, May 2009

is based on identifying the change in frequency in the 4.2 Segmentation

given query. The query is received in the form a The segmentation algorithm we have proposed
humming tune and based on the rise and fall in the pitch involves determination of onset and offset followed by a
of the received query, the melody pattern that matches two level segmentation process. The modules are
with the query’s rise and fall of pitch is retrieved. In the explained below:
work proposed by [7] for melody retrieval the authors
have identified melody based on features like distance 4.2.1 Onset and Offset detection
measures and gestalt principles. The approach we have Usually, a Carnatic song is associated with one Ragam.
followed is based on low level signal features and we Sometimes, there is a special song which is normally
have identified the raga by considering a polyphonic sung in a combination of Ragams where, every 4 or 5
signal as input to our system. lines will be sung in one Ragam the next 4 or 5 lines in
another ragam and so on. Such a song is called as a
4. SYSTEM ARCHITECTURE Ragamalika, which means composed of many Ragas.
This is not a usual scenario and hence in our discussion
The proposed architecture of the system is given in
we are trying to identify songs that are based on only one
Figure 1. The various modules of our system are signal
Ragam. Every song is associated with a Talams as
separation, segmentation, feature extraction, frequency
indicated [2]. The Talam’s first count starts at the
mapping based on singer identification and finally raga
beginning of the song and ends when the song completes.
identification. The various blocks are explained in the
In other words, any song is an integral multiple of a pre-
following sections.
specified Talam. This gives an indication that in order to
I/P
Music identify the swaras of the song it is a good idea to find out
signal Signal Segmentation Feature the beginning and the end of the Talam. This requires us
Separation Extraction to identify the onset and offset of a song. Onset refers to
the beginning of a musical note or other sound, in which
the amplitude rises from zero to an initial peak.
Approaches to onset detection can operate in the time
domain, frequency domain, phase domain, or complex
Talam domain, and include looking for the following changes:
1. Increases in spectral energy
Frequency
Singer 2. Changes in spectral energy distribution
Mapping based
Identification on fundamental (spectral flux) or phase
frequency 3. Changes in detected pitch
4. Spectral patterns recognisable by machine
learning techniques such as neural networks.
Simpler techniques such as identifying increase in
Singers Indexed Raga
time-domain amplitude can typically lead to an
- DB Raga Mapping
Database unsatisfactorily high amount of false positives or false
negatives [11]. The technique that we adopted is based on
O/P identifying change in spectral energy. Hence, we take a
copy of the signal and convert it into frequency domain
Figure 1: System Architecture.
and observe the change in spectral energy. We then map
this point to the time domain signal and mark that as the
4.1 Signal Separation onset. We detect the offset in a manner similar to
detecting onset where we identify the point at which the
The performances of two signal separation algorithms spectral energy starts dying as the offset. Both the onset
proposed by [8, 9] have been compared and the authors of and offset points are transformed to the time domain
[10] have concluded that the one proposed by [8] is better signal and then the time domain signal between the points
suited for Carnatic music. The algorithm proposed by [8] of onset and offset is taken for the purpose of raga
is slightly modified to efficiently suit Carnatic music. The identification.
threshold value of the harmonic is chosen so that on
repeated iterations the signal gets separated. During the 4.2.2 First Level Segmentation
first iteration the threshold value for the harmonic is as After determining the onset and offset the signal
given by [9] and the signal is separated into voice and between these two points are considered for the next level
music. During the next iteration the threshold value is of segmentation. Our assumption is that, this segmented
chosen depending on the extracted output of the first signal must be an integral multiple of one of the Talams
iteration. After the signal is separated into voice and as indicated in [2]. A database of 175 Talams with their
instrument, the signal is segmented to identify the time duration is maintained. This database is sorted based
frequency components present in the signal for swara on decreasing frequency of usage of the talams. We try to
identification yielding to raga identification. map the talams time duration with the signal. A fitness
function is written based on time duration to assign the

© 2009 ACADEMY PUBLISHER 572

SHORT PAPER
International Journal of Recent Trends in Engineering, Vol. 1, No. 1, May 2009

time duration of the Talam to the segmented input frame. TABLE I.

SWARAS AND THEIR RATIOS WITH ‘S’
The fitness function tries to match the segment against
the duration of a talam which is retrieved from the Swara Ratio Swara Ratio
database. The matching is based on the fact that the S 1 M2 27/20
duration of segment must be an integral multiple of the R1 32/31 P 3/2
R2 16/15 D1 128/81
talam duration. The talam that matches the segment input R3 10/9 D2 8/5
is identified as the Talam for the input song. The next G1 32/27 D3 5/3
level of segmentation is based on the identified Talam. G2 6/5 N1 16/9
G3 5/4 N2 9/5
4.2.3 Second Level Segmentation M1 4/3 N3 15/8

After obtaining the talam associated with a segment, The fundamental frequency normally corresponds to S
the segment is divided into individual talam segments by and all other swaras of R, G, M, P, D, and N will depend
observing the time duration in the Talam database. Any on this S. The ratio of the other swaras relative to S is
Talam has a fixed predetermined component pattern. listed in Table I. Using literature on Carnatic music and
Each component depending on whether it is a count – ‘|’, normal singers singing pattern we have formulated a table
thattu ‘0’, veech ‘0’, has varying duration. After of singers with their corresponding fundamental
identifying the Talam in the first level of segmentation frequency. This is populated in the singer’s database and
the pattern of the identified Talam is known. Using this the fundamental frequency of the singers is retrieved for
feature the segmented signal is further segmented at each Swara determination. Using Table I and the extracted
count, Thattu or Veechu of the given Talam. At this point frequency components raga identification is performed.
of segmentation we have individual segments which
correspond to the Thattu, Veechu or count. In Carnatic 4.4 Frequency mapping and Raga determination
music, for a count, Thattu, Veechu there can be one, two, The swara is determined by identifying the singer’s
or four swaras depending on the tempo in which the song fundamental frequency and the dominant energy based
is sung. That is, each component of a talam obtained after frequency in each segment. Using the
the second level of segmentation can be realized as one, the ratio available in Table I the swara corresponding to
two or four swaras. Hence each of our segments at this every frequency is determined using the identified
level of segmentation could correspond to either one, two dominant frequency in every segment. We then identify
or four swaras. This requires us to further segment the the individual swaras available and use these swaras and
signal. Hence we segment each talam component compare it with a database of Raga to identify the raga of
obtained as a result of second level segmentation into the given song. The Raga database is a table consisting of
four equal frames. This segmented signal is then used for the name of the raga, the arohanam, avarohanam of raga
the purpose of Raga identification. in the form of swara components. We use a simple string
matching to compare the identified swara pattern with the
4.3. Feature Extraction
Raga database and determine the Raga of the input
After identifying the smallest possible segment we Carnatic song.
assume that this could correspond to a swara. We then
use the HPS algorithm as explained in [12] to identify the 5. RESULTS & EVALUATION
frequency components in a single segment. In each
segment there is a possibility for the presence of more The input signal is sampled at 44.1 KHz. We consider
than one frequency component. In order to determine a the identification of three melakarta Ragams for the
frequency component corresponding to a swara in each purpose of evaluating this algorithm. For the purpose of
segment, we use energy as the basis for identifying the Raga identification we have considered songs sung by
distinct frequency component in each segment. After like Nithyasree, M.S.Subalakshmi, Balamuralikrishna
identifying the frequency component present in each and Ilayaraja. From music knowledge we have identified
segment, we determine the one with the highest energy as their fundamental frequencies as 400 Hz, 320 Hz, 400 Hz
the one that corresponds to the swara. We now have to and 240 Hz respectively. We considered songs that
map this frequency component to the corresponding belong to melakarta ragas, Sankarabharanam, Kanakangi,
swara. In order to perform this mapping, we need to Karaharapriya ragam and belonging to Adi talam or
identify the fundamental frequency of the signal. The Roopaka Talam. We made the signal to pass through the
fundamental frequency is identified by performing the signal separation algorithm, and segmentation algorithm.
process of singer identification. The result showing the segmentation points for one input
We extract features pertaining to identifying the is shown in Fig 2. This is the first level of segmentation
singer of the song as explained in [3]. The singer of the where the protruding lines indicate the points of
song is identified as explained in [3] and this is used as segmentation. After identifying the segmentation points
the basis for identifying the fundamental frequency of the we determine the frequency components using the HPS
song. The fundamental frequency is essential for Carnatic algorithm and tabulated the frequency values which have
music analysis, since a singer can use a frequency in the the dominant energy. The result is shown in Figure 3. The
range of 240 to 400 Hz as the fundamental frequency. distinct frequencies obtained for the signal separated song

© 2009 ACADEMY PUBLISHER 573

 SHORT PAPER
International Journal of Recent Trends in Engineering, Vol. 1, No. 1, May 2009

in Figure 2 is 400 600 427 420 417 554 418 420 For the ragam Karharapriya, Nithyasree’s voice gave
418 625 623 494 416 417 418. Using the singer correct results for 6 songs out of the tested 10 songs. This
identification system, we determined that the singer is is because of the fact that, those songs had a better
Nithyasree for the considered sample and then identified frequency range compared to other songs and hence the
the ragam as Sankarabharanam. The details of the Raga result is wrong because of the assumption of the
identified based on singer and the number of songs for fundamental frequency. The same algorithm was run for
which the raga was identified is given in Figure 4 Hindustani music and we observed that the algorithm
gave correct results for two of the Hindustani Ragas.
A Since Hindustani music and Carnatic music’s note
m indication is different as explained earlier, some
p combination which is not present in Hinudstani music is
l also likely to occur which is to be omitted.
i
t 6. CONCLUSION & FUTURE
u
d This work is based on low level signal features
e for Raga identification. We have tested for three Ragas
belonging to two Talams sung by three singers. The
Time system has to be tested for different singers, all melakarta
Ragas and ten most frequently occurring talams. Our
system has to be improved with a Hidden markov model
Figure 2: Input signal separated and segmented
based approach where we combine the low level features
and HMM for better and accurate identification. This
system’s basic disadvantage is the assumption of
F fundamental frequency and hence determination of
r
e
fundamental frequency is our next task. The system has
q to be modified to use the identified raga to index songs
u and use raga as a query for music information retrieval
e
n REFERENCES
c
y [1] Prof. Sambamurthy, South Indian Music, vol. 4, The Indian
Music Publishing House, Madras, India, 1982
[2] Rajeswari Sridhar and T.V. Geetha, “Swara identification of
Carnatic music”, IEEE Computer Society press, proceeding of
Figure 3: Frequency components ICIT 2006
[3] Rajeswari Sridhar and T.V. Geetha, “Music Information
Retrieval of Carnatic Songs Based on Carnatic Music Singer
Raga Identification Identification”, IEEE conference on Computer Science, ICCEE
12
2008
10 Tested
[4] Gaurav Pandey, et al “Tansen: A system for automatic Raga
identification” IICAI, 2003
No. of Songs

6 Co rrectly [5] Ram Avtar Vir, “Theory of Indian Music” Pankaj

Identified Publications, New Delhi, 1999.
4 [6] Preeti Rao and Anand Raju, “Building a melody retrieval
2 system”, NCC 2002
[7] Maarten Grachten et al, “Melody retrieval based on the
0
Implication Realization model”, MIREX 2005
Subbalakshmi

Balamuralikrishna
Nithyasree

Nithyasree

Ilayaraja

Nithyasree

[8] Yun-Gang Zhang, “Separation of Voice and Music by

M. S.

Harmonic Structure Stability Analysis”, IEEE Proceedings,

M.

2005.
[9] Mark R Every & John E Szymanski, “A Spectral Filtering
Sankarabharanam Kanakangi Karaharapriya
approach to music signal separation”, Proceedings of the 7th
Raga and Singers International Conference on Digital Audio Effects (DAFx ’04),
Naples, Italy, October, 2004, pp 197 – 200
[10] Rajeswari Sridhar and T.V. Geetha, “Comparison of signal
Figure 4: Raga identified based on singers separation algorithms as applied to Carnatic music”, FRSM
2007
For the singer Nithyasree, the number of songs for [11] Bello, J.P., Daudet, L., Abdallah, S., Duxbury, C., Davies,
which raga identified correctly was high compared to M., Sandler, M.B., "A Tutorial on Onset Detection in Music
M.S. Subbalakshmi and Balamuralikrishna because of the Signals", IEEE Transactions on Speech and Audio Processing,
fact that Nithyasree’s fundamental frequency, which we 2005
[12] Kyogu Lee, “Automatic chord recognition from audio
have assumed is more correct compared to other singers.
using enhanced pitch profile”, ICMC 2006

© 2009 ACADEMY PUBLISHER 574

View publication stats