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

net/publication/351996220

North Indian Classical Music Tabla Tala (Rhythm) Prediction System Using
Machine Learning

Chapter · January 2021

DOI: 10.1007/978-981-33-6881-1_16

CITATION READS

1 137

2 authors, including:

Shambhavi Shivraj Shete

6 PUBLICATIONS 4 CITATIONS

SEE PROFILE

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

Analysis of Effect of Indian Music( Classical and other all forms) on Human Emotions and to establish relationship between Ragas,Musical Note Patterns and Change of
Human Emotions. View project

All content following this page was uploaded by Shambhavi Shivraj Shete on 18 August 2022.

The user has requested enhancement of the downloaded file.

 North Indian Classical Music Tabla Tala
(Rhythm) Prediction System Using Machine
Learning
Shambhavi Shivraj Shete
Career Point University,
Kota (RJ) India – 324005.
shambhavishete2020@gmail.com Commented [t1]: Check the mail address here?
Saurabh Harish Deshmukh
Maharashtra Institute of Technology,
Aurangabad (MH) India - 431005.
saurabh.h.deshmukh@gmail.com

Abstract—Music Information Retrieval (MIR) and its applications have gained

popularity after the advancements in machine learning techniques. In various appli-
cations such as, genre detection, song recognition, automatic musical score genera-
tion, and music transcription, the identification of rhythm, tempo, stroke type, etc.
is necessary. For Western Music, various applications have been developed to serve
the purpose. HoweverH however; , very few researches have been done in North
Indian classical Music (NICM). The rhythm instrument used in NICM called Tabla,
and its rhythm patterns (Talas) are considered in this research. The system proposed
here predicts the Tala (Rhythm). ‘Timber’ is a Non-nontangible audio attribute (de- Commented [Professor2]: Keep consistent. Non-tangible or
scriptor) of sound. The descriptors that more clearly define a timber such as, Zero nontangible.
Cross Rate, Roll-off, Roughness, Brightness, and Irregularity) are combined with
Mel Frequency Cepstral Coefficient to extract the audio features. Overall, 366 audio
excerpts containing sound samples of 9 basic Tabla strokes are used to train the
system. Machine learning algorithms, Multi-Support Vector Machine, and Feed
Forward Back Propagation Neural Network are used to identify the Tabla strokes,
and the frequency of occurrence of each basic stroke is calculated. The novel
method proposed in this paper predicts the Tala with the help of identified strokes,
the total number of strokes, and the frequency of occurrence of each stroke. The
result shows that the Tala could successfully be predicted even if, not all strokes
from the Tala are correctly identified. The Tala prediction accuracy of 93.33% is
obtained and for the Tabla stroke identification accuracy of 96.85% are obtained.are
obtained. The correlation factor of +0.77 between Tabla stroke identification and
Tala prediction calculated proves that, even if, not all Tabla strokes are identified
correctly, the Tala could be stillstill be predicted using the basic structure of
NICM’s Tabla Tala.

Keywords—Machine Learning, North Indian Classical Music, Tabla, Tala

(Rhythm), Timbre
2 Error! Use the Home tab to apply title to the text that you want to appear here.2

Introduction

Music Information Retrieval (MIR) is a rapidly growing branch of sound engi-

neering that has various applications in the music industry, film, and television in-
dustry and in the fields of music recognition and synthesis. With the advancements
in machine learning, Music Information Retrieval has gained substantial popularity
and interest across the world.
North Indian Classical Music (NICM) is one of the ancient classical music
traditions in the world. Along with vocal recitation, in the NICM different
accompanying musical instruments such as Tanpura, Harmonium, Sarangi, Tabla,
Pakhawaj, etc. are used. The requirements and the theoretical structures of the use of
accompanying musical instruments are explained in ‘Natyashastra’[1]. This etiology
explains the structure of the instrument, the material to be used, the tuning frequency
of the instrument, and methods to play these accompanying musical instruments.
Tabla is one of such basic accompanying rhythm instrument used popularly by
the vocalists during their vocal performances in NICM. Tabla is made up of two
drums of different sizes and shapes. The left drum is metallic, hollow, dome-
shaped, tightly wrapped with animal skin. The right drum is made up of teak wood
and smaller than the left drum. It is laced with hoops and wrapped with animal skin.
Both the drums have dowels and hoops to tighten the drums. A paste mixture of rice
flour and iron filings is applied on both the drum surfaces. This dried paste gives a
defined pitch to the drums. This paste is applied to the skin in several layers [2]. The
right drum is usually tuned with the musical note of the fundamental frequency of
the singer’s scale(C4# for male singer and F4# for female).

Tala Beats Rhythm Pattern Tabla Stroke Frequency

Dha Dhin Dhin Dha |
Dha Dhin Dhin Dha |
Teental 16 Dha-6, Dhin-6, Tin-2, Ta-2
Dha Tin Tin Ta |
Ta Dhin Dhin Dha ||
Dadra 6 Dha Dhin Na | Dha Tin Na || Dha-2, Dhin-1, Na-2, Tin-1
Rupak 7 Tin Tin Na | Dhin Na | Dhin Na || Tin-2, Na-3, Dhin-2
Dhin Na | Dhin Dhin Na|
Jhaptal 10 Dhin-5, Na-4, Tin-1
Tin Na | Dhin Dhin Na ||
Dha Ge Na Tin |
Kerva 8 Dha-1, Ge-1, Na-3, Tin-1, Ka-1, Dhin-1
Na Ka Dhin Na ||

Table1. Basic Tabla Talas and Rhythm Pattern

Tabla is also played in a solo form. Tabla player plays different forms and
combinations of basic Tabla strokes. These combinations of Basic Tabla strokes are
performed by keeping the type of Tala (Rhythm) and tempo intact. The Tala is a
rhythmic cycle of the beat. From the North Indian Classical Music Tala structure, it
is observed that each basic Tala contains a fixed number of Tabla strokes that are
 Introduction 33

dissimilar for different types of Tala. The basic Tabla Talas and their rhythm patterns
with the frequency of occurrences of Tabla strokes are shown in Table 1. There are
several methods to identify the stroke of a rhythm instrument. The percussion
instrument used in music differs for the type of music. African music makes use of
percussion instruments that are made up of animal skins. Western music and popular
music make use of drums. Each of these instruments has different names of the
rhythms and the basic strokes to perform the rhythm. The Drum stroke Identification
[3] from polyphonic music requires separation of sound originating from different
frequency bands dedicated to the part of the drum set being played. Apart from this,
classification of the percussive sound [4], Automatic Transcription of Drum strokes
[5] [6], Tonic Independent Mrudangam stroke Transcription [7], Akshara Transcrip-
tion of Mrudangam [8], Tabla stroke Recognition and Resynthesize [9], Polyphonic
Musical Instrument Recognition [10] are some of the major researches carried out in
music information retrieval domain, where the basic strokes of the instrument and
rhythm patterns are identified. The typical process of identification of strokes from
rhythm instrument requires onset detection, audio segmentation, feature extraction,
and classification. After detecting the basic strokes from the rhythm instrument, it is
easy to recognize the rhythm pattern.
Similar to other world music, in NICM, the rhythm of the Tabla instrument is
periodic. A Tala is a combination of a fixed number of total strokes, played in cycles.
In this research, five basic Talas, namely Teental (16 strokes), Jhaptal (10 strokes),
Rupak (7 strokes), Dadra (6 strokes), and Kerva (8 strokes) are considered. The se-
lection of these Talas is based on their variety in the total number of beats as well as
a variety of Tabla strokes being used to play the Talas. These Talas are selected that
have a different number of total strokes and are some of the basic Tabla Talas used
in NICM. The strokes of Tabla also have a pattern of repetition for some strokes,
while some strokes are used only once in a cycle of the rhythm. The frequency of
occurrence of Tabla strokes in the Talas could be counted by identifying the basic
Tabla strokes. For example, Teentala is composed of 16 basic strokes, out of which
‘Dha’ and ‘Dhin’ strokes are repeated 6 times, and ‘Tin’ and ‘Ta’ strokes are repeated
2 times throughout the entire one cycle of the Tabla recording.
Prediction of Tala from the identified strokes is possible with the help of basic
information provided to the system regarding the number of strokes in each Tala,
types of strokes, and frequency of occurrence of these basic strokes in that Tala. For
the learners of Tabla instruments, it is necessary to know the Tabla strokes that are
being played by the Tabla performer. For the systems of automatic music recognition
and transcription, it is important to identify the Tala. For the learned listeners/practi-
tioners of NICM, it is a bit difficult to recognize the basic strokes and their repetition
of the pattern to predict the Tala accurately. For common listeners, it is very chal-
lenging to recognize the Tala based on the basic Tabla strokes being played and the
total number of repetitions of the strokes. Machine Learning becomes a useful tech-
nology to train a system to automatically identify Tabla strokes, tempo, and predict Commented [Professor3]: Capital?
the Tala.
In this research, we propose a North Indian Classical Music Tala (Rhythm) predic-
tion system based on thetype and number of Tabla strokes being identified using
Machine Learning. The System uses a Machine Learning algorithm to identify the
Tala based on the detection of the Tabla stroke, frequency of occurrence of Tabla
4 Error! Use the Home tab to apply title to the text that you want to appear here.4

strokes, and a theory of the basic structure of a given Tala that formulates the com-
bination of the stroke into the Tala.

State of the Art

Music information retrieval is a branch of Sound Information Retrieval that deals

with extracting meaningful musical information from a given piece of music. This
musical information contains a low-level audio signal description. It is useful to
identify a Tabla stroke. The descriptors that describe this low-level information of a
sound are called low-level descriptors (LLD) [11]. There exist many Digital Signal
Processing (DSP) methods to extract musical information and represent it in various
ways.
The features extracted from an audio play a prominent role in the identification
of Tabla strokes. The audio features, also called Audio Descriptors, are categorized
using different perspectives. The taxonomy of the audio descriptors differs with re-
spect to their applications. There are Global audio descriptors that are computed over
the entire audio signal and instantaneous descriptors that are computed for each time
frame. Temporal audio descriptors are computed to generate signal energy envelop,
and Spectral audio descriptors are computed by computing different transforms [11].
Sound is a physical phenomenon that is expressed in the form of Pitch,
Amplitude, Duration, and Timbre. Pitch is expressed in term of Hertz (Hz),
Amplitude is expressed in terms of decibels (dB), and Duration is expressed in terms
of Seconds (Sec). However, other than these three measurable dimensions of sound,
the Timbre is a multidimensional nontangible entity. Timber cannot be expressed in
terms of any unit. There are many ways with which researchers have explained the
Timber [12]. Timbre is expressed using Attack time, Attack slope, Zero crossing
Rate, Roll-off, Brightness, Roughness, and Irregularity, along with MFCC [13]. The
zero-cross rate is used to detect the number of times the signal crosses the x-axis. It
is a measure of noise. Roll-off is used to represent the shape of a high-frequency
energy signal. It is a frequency below which 85% energy signal is present. Brightness
represents the spectral energy above the threshold frequency of 1500 Hz. Roughness
is a sensory dissonance between all peaks of the spectrum of the signal. Irregularity
is the degree of variation of the successive peaks of the spectrum of a signal. Mel-
frequency Cepstral Coefficient (MFCC) is a set of the power spectrum which de-
scribes the shape of the sound [14].
It is very difficult to select appropriate audio descriptors that describe the
timber of a sound accurately. There exist hundreds of audio descriptors and their
variants. Hybrid Selection Method of wrapper approach promises to develop a subset
of audio descriptors useful for given music information retrieval applications [15].
Research has proved that these audio descriptors are useful for Singer Identification
[16] [17], Gender Recognition of a Singer [18], Tabla stroke Identification [19], and
Musical Instrument Identification [20].
There are many challenges to identify a Tabla stroke. In general, it is difficult to
differentiate between the strokes ‘Te’ and ‘Ti’ without making use of a language
 Introduction 55

model for each of the strokes. This is because the Timbral descriptors of ‘Te’ and
‘Ti’ are much similar to each other, and a Tabla player plays these short duration
strokes interchangeably. Also, a fall of accuracy for the identification of Tabla stroke
is reported due to the mixture of the previous stroke’s ringing sound with the
current stroke. For example, Tabla stroke ‘Na’ could be misclassified due to the
ringing sound of the previous stroke ‘Dha’,which is generated by hitting
both the drums simultaneously. The ‘Dha’ is produced using ‘Ga’ (Left Drum) +
‘Na’ (Right Drum). Thus, it becomes difficult to separate this ‘Na’ with the next
stroke ‘Na’ [21]. Tabla stroke identification is based on the overlapping of the Tabla
strokes. Spectral centroid and ZCR features are useful to identify the power spectrum
and overlapping between the Tabla strokes. The Tabla strokes ‘Dha’, ‘Dhin’, ‘Din’,
and ‘Tin’ produce more overlapping due to the sustained sound.

The Method

Fig. 1. Tabla Stroke Identification and Tala Prediction System

6 Error! Use the Home tab to apply title to the text that you want to appear here.6

An automatic Tabla stroke identification and Tala prediction system are as shown
in Fig. 1. The audio database is recorded through a professional Tabla player in a
studio environment. Different Talas and individual basic Tabla strokes are recorded
with multiple instances of each Tabla stroke. The Tabla instrument used is tuned to
C4# note. The database consists of the audio excerpts of 9 basic Tabla strokes with
25 samples of each stroke. The Tabla strokes considered here are ‘Na’, ‘Ta’, ‘Te’,
‘Tun’, ‘Ga’, ‘Ka’, ‘Dha’, ‘Dhin, and ‘Tin’. The Tabla strokes are digitized using a
sampling frequency of 44.1 kHz with 16-bit PCM .wav file format
The Tabla stroke location in an audio file is determined by identifying the onset
locations. Onsets are the locations where the signal energy is high [22] [23]. Each
audio excerpt is processed to identify the onsets. After identifying the onsets, the
audio file is segmented. Each segment contains one Tabla stroke. The accuracy of
the identification of the onsets and subsequent segmentation plays an important role
in the Tabla stroke identification system. After the segmentation of all the audio
excerpts, a total of 366 Tabla strokes are separated. Out of the total Tabla strokes
segmented, 70% is used for training, and 30% for testing.
Timbral audio descriptors, namely Zero Crossing Rate (ZCR), Roll-off,
Brightness, Roughness, Irregularity, and Mel Frequency Cepstral Coefficients
(MFCC), are extracted from the Tabla stroke samples [13].
We have considered two supervised learning algorithms for identification of the
Tabla strokes, namely Multi-Support Vector Machine (Multi-SVM) and Feed For-
ward Back Propagation Neural Network (FFBPNN). In general, the SVM is useful
for binary classification; however, it could also be extended to be used for multi-
class classification. In one vs. one configuration, the output is binarized such that the
target class represents 1 while other classes as 0. This configuration requires voting
to identify a stroke that is a class [7]. Both machine learning algorithms are super-
vised and are useful in many music information retrieval applications.
Feed Forward Back Propagation Neural Network (FFBPNN) has two passes for-
ward pass and backward pass [24]. The output of each neuron is calculated by the
sum of the product of each input and its associated weights, followed by a squashing
function. At the output, layer error is calculated by comparing the output and the
target. In the backward pass, this error is then back-propagated from the output layer
to the input layer in the form of a change of weights. Feed Forward Back Propagation
Neural Network (FFBPNN) is trained by 25 audio features extracted from audio ex-
cerpts.
Based on the types of Tabla strokes, the total number of Tabla strokes, and their
frequency of occurrences, the Tala is predicted. The performance of the system is
evaluated based on the accuracy of the identification of the Tabla strokes and Tala.

Experiments and Results

The experiments are carried for 9 basic Tabla strokes. The test samples are se-
lected from 5 basic Tala. The audio features are categorized into two sets consisting
of only MFCC and Timbre audio descriptors with MFCC. The Feed Forward Back
 Introduction 77

Propagation Neural Network (FFBPNN) and Multi-SVM are trained and tested to
identify the Tabla strokes. When presented 15 test samples, 3 from each tala, it is
observed that FFBPNN supersedes in performance over Multi-SVM for the identifi-
cation of Tabla strokes. The Automatic Tabla stroke identification system using
FFBPNN gives a stable system than using Multi-SVM. The results show that the
generalization capability of a trained neural network accurately identifies Tabla
strokes from unknown test samples whereas, being binary (one vs. one) classifier,
Multi-SVM does not guarantee a stable Tabla stroke identification system. Also, au-
tomatic Tabla stroke identification accuracy is highest for Timber audio descriptors
than MFCC when FFBPNN is used.
It is observed that the sound produced from 'Ga' is closed, and its sustain is small
. Thus the accuracy of identification of Tabla stroke ‘Ga’ is 100%. Timbral attributes
of the sound produced by stroke ‘Ga’ are unique and have no match with any other
Tabla stroke present in the database. On the other hand, for the Tabla stroke ‘Tin’,
that generates overtones merged with its fundamental frequency. From this mixed
harmonic structure, it becomes difficult to correctly identify the Tabla stroke ‘Tin’.
The Tabla stroke ‘Tin’ is produced by hitting an open stroke on the right-hand drum.
In contrast to this, when the Tabla stroke 'Na' is produced, the Tabla player damps
the overtones produced after it. Thus, the ‘Na’ stroke has high onset energy while
sharp sustain. We have compared the Tabla stroke originated from Left, Right, and
Both the drums with respect to their identification accuracy. It is observed that the
Right drums Tabla stroke identification accuracy is 96%, the Left drum 84%, and for
the strokes originated from both the drums simultaneously, it is 88%. The reason
behind these differences could be explained with respect to the A (attack), D (Decay),
S (Sustain), and R (Release) of each stroke.
The method proposed here ensures that, even if the system does not detect all
the Tabla strokes accurately, it is possible to predict the Tala to some extent. This is
because few misclassified Tabla strokes do not affect Tala identification. In NICM,
there exist Talas that contain the same number of total strokes, but the number of
occurrences of the strokes and the locations of the strokes in the entire cycle of the
Tala is different for different Talas. Thus, it is possible to predict the Tala even if one
or more Tabla strokes are misclassified.
Fig. 2 shows the results obtained for Tabla stroke Identification and
corresponding Tala Prediction. As shown in the table, the Tabla stroke identification
accuracy for FFBPNN using Timbral audio descriptors is 100%, and the correspond-
ing Tala is also predicted correctly. However, for a few cases of Teental, even if not
all Tabla strokes are correctly identified, the system predicts the Tala accurately. In
the case of Rupak Tala, though, the Tabla stroke identification accuracy is 71.43%
the Tala could not be predicted correctly. This is because the total number of basic
Tabla strokes identified is more than the basic strokes of Rupak Tala. In such a situ-
ation, the system could not predict Tala accurately.
When MFCC alone was used as an audio descriptor, the Tabla stroke identifica-
tion accuracy has dropped to some extent. But still, the Tala could be predicted ac-
curately. In this case also, for Rupak Tala, which has only three basic Tabla strokes,
it is observed that, if the identified strokes belong to the basic Tabla strokes, the Tala
could be predicted. However, if the identified stroke is not from the category of basic
Tabla stroke of the given Tala, then the probability that the system may predict the
8 Error! Use the Home tab to apply title to the text that you want to appear here.8

Tala accurately would decrease. In Fig. 2, for test audio samples R_2 and R_3, the
Tala could be predicted for the former while it could not be predicted for the
latter, even though the Tabla stroke identification accuracy for both is 85.71%. Commented [Professor4]: Double check.
When Multi-SVM is used with MFCC and Timber, the results were not promising,
even though the Tabla stroke identification accuracy is high. This is due to the mis-
classification of the Tabla strokes into some other category.

Fig. 2. Tabla Stroke Identification and Corresponding Tala Prediction Results

Overall, the system performance is calculated with respect to the number of times
the system could predict the Tala. The Tala prediction accuracy for FFBPNN using
the Timber audio descriptor is 93.33% and using Multi-SVM 73.33%. Fig. 2. shows
the comparison of the performance of stroke identification and Tala prediction. It is
observed that when the Tabla stroke identification accuracy is higher for FFBPNN,
the Tala is predicted accurately. These results are similar for both Timbral and MFCC
audio descriptors. On the other hand, for Multi-SVM, the stroke identification accu-
racy degrades the performance of Tala prediction. The correlation factor between
stroke identification and corresponding Tala identification is found to be +0.77,
which indicates a strong uphill linear relationship between stroke identification and
Tala Predication. It is obvious that if the stroke identification accuracy is 100%, then
the Tala prediction would also be accurate.
The Tala Prediction algorithm proposed here works differently for different
types of Talas. We have considered 5 basic Talas, and the logic could be extended
to any number of Talas from NICM. The proposed algorithm considers three factors
before predicting a Tala viz. The total number of Tabla strokes, the number of basic
Tabla strokes considered, and the frequency of each of the identified Tabla strokes.
The proposed algorithm considers different logic for different types of Talas. As
 Introduction 99

explained earlier in Table 1, for example, Teental consists of 4 basic strokes and 16
total numbers of strokes per cycle. The fixed frequency of occurrences of the strokes
‘Dha’ & ‘Dhin’ (6 times) and ‘Ta’ & ‘Tin’ (2 times) each ensures the presence of
Teental. The algorithm predicts the Tala based on the fulfillment of the required
frequency of strokes ‘Dha’ or ‘Dhin’. In such cases, it is not necessary to identify
all 16 strokes accurately. By checking the frequencies for these two strokes, the Tala
could be predicted accurately. Similar logic applies to each Tala, and the logic dif-
fers for different Tala. The same algorithm can be further extended for any number
of Tala present in North Indian Classical Music Tabla Talas.

Fig. 3. Comparison of Tabla Stroke Identification and Corresponding Tala Prediction

Conclusion and Future Scope

Music information retrieval has gained substantial popularity in the music re-
search domain. With the advancements of signal processing techniques and compu-
tational capabilities of neural networks, a lot of applications in the field of music
have been invented. In this research, we have proposed a method for Tala predic-
tion, based on the number and type of basic Tabla strokes identified. Total 9 Tabla
strokes and 5 types of basic Tala are considered to train the system using FFBPNN
and validated using Multi-SVM. Traditional MFCC and non-tangible Timbral at-
tributes of sound are considered for feature extraction. For FFBPNN and Multi-
SVM, Tala could be predicted accurately when Tabla stroke identification accuracy
was 100%. However, the proposed system in this research could also successfully
identify Tala even though the Tabla stroke identification accuracy was lower. Commented [Professor5]: lower?
The system proposes a method that uses North Indian Classical Music Tala struc-
ture in such a way that, for each Tala, if some basic strokes are correctly identified,
then the Tala could be predicted accurately by using count and type of the basic
strokes identified for that Tala. The Tabla stroke identification accuracy of the
10 Error! Use the Home tab to apply title to the text that you want to appear here.10

system is found to be 96.85% using FFBPNN, and the Tala prediction accuracy of
93.33%. The Tala prediction accuracy increases with the increase in Tabla stroke
identification accuracy. The correlation factor of +0.77 between Tabla stroke iden-
tification and Tala prediction proves that accurate is the Tala prediction when Tabla
stroke identification is accurate. The results show that Tala identification could be
successfully achieved even for low Tabla stroke identification accuracy as 68.75%.
In the future, the system and the proposed Tabla Tala Prediction Algorithm could
be extended to any number of Tabla Tala in North Indian Classical Music. Also, the
acceptable lower threshold for misclassified Tabla stokes that is capable
of predictingTala could be determined.

BIBLIOGRAPHY

[1] G. L. Beck, Sonic Liturgy: Ritual and Music in Hindu Tradition, South Carolina: Univ of
South Carolina Press, 12-Dec-2012.
[2] Klippstein, "Construction & Playing: Tabla," https://www.india-instruments.com/, 2012.
[Online]. Available: https://www.india-instruments.com/construction.html. [Accessed 3
23 2020].
[3] K. Anssi and J. Paulus, "Drum Sound Detection in PolyphonicMusic with
HiddenMarkovModels," EURASIP Journal on Audio, Speech, and Music Processing,
vol. 2009, p. 9, 16 November 2009.
[4] D. V. Steelant, K. Tanghe, S. Degroeve, B. Baets, M. Leman, and J.-P. Martens,
"Classification of Percussive Sounds Using Support Vector Machines," in Annual
Machine Learning Conference of Belgium and The Netherlands (BENELEARN),
Belgium and The Netherlands, 2004.
[5] C. Jacques and A. Roebel, "Automatic Drum Transcription with Convolutional Neural
Networks," in 21th International Conference on Digital Audio Effects, September 2018,
Aveiro, Portugal, September 2018.
[6] R. Vogl, "Deep Learning Methods for Drum Transcription and Drum Pattern Generation,"
Johannes Kepler University Linz, Linz, 2018.
[7] A. Anantapadmanabhan, J. Bello, R. Krishnan, and H. Murthy, "Tonic Independent stroke
Transcription of the Mrudangam," in AES 53RD INTERNATIONAL CONFERENCE,
London, UK, 2014 January 27–29.
[8] J. Kuriakose, K. Chaitanya, S. Padi, H. Murthy, and S. Umayalpuram, "Akshara
Transcription of Mrudangam strokes in Carnatic Music," in 2015 Twenty-First National
Conference on Communications (NCC), Mumbai, India, 2015.
[9] P. Chordia and A. Ray, "Tabla Gyan: An Artificial Tabla Improviser," January 2010.
[10] J. J. Burred, A. Robel and T. Sikora, "Polyphonic Musical Instrument Recognition Based
on a Dynamic Model of the Spectral Envelope," in IEEE International Conference on
Acoustics, Speech and Signal Processing, Taipei, 2009.
[11] P. Geoffroy, A large set of audio features for sound description (similarity and
classification) in the CUIDADO project, 23 April 2004.
[12] T. H. Park, "Towards Automatic Musical Instrument Timbre Recognition," Candidacy,
2004.
 Introduction 1111

[13] O. Lartillot, "MIR ToolBox Manual," Finnish Centre of Excellence in Interdisciplinary

Music Research.
[14] S. Deolekar and A. Siby, "Tree-based classification of tabla strokes," CURRENT SCIENCE,
vol. 115, 10 November 2018, pp. 1724-1731, 2018.
[15] S. H. Deshmukh, "A Hybrid Selection Method of Audio Descriptors for Singer
Identification in North Indian Classical Music," in IEEE Explorer, Himeji, Japan, 2012.
[16] S. H. Deshmukh and S. Bhirud, "North Indian classical music's singer identification by
timbre recognition using MIR toolbox," International Journal of Computer
Applications, vol. 91, no. 4, pp. 1-5, 2014.
[17] S. H. Deshmukh and S. Bhirud, "Analysis of Audio Descriptor Contribution in Singer
Identification Process," International Journal of Emerging Technology and Advanced
Engineering, vol. 4, no. 2, pp. 731-737, 2014.
[18] S. H. Deshmukh and S. Bhirud, "A Novel Method to Identify Audio Descriptors, Useful in
Gender Identification from North Indian Classical Music Vocal," International Journal
of Computer Science and Information Technologies, vol. 5, no. 2, pp. 1139-1143, 2014.
[19] S. S. Shete and S. H. Deshmukh, "Analysis and Comparison of Timbral Audio Descriptors
with Traditional Audio Descriptors Used in Automatic Tabla Bol Identification of North
Indian Classical Music," in Proceeding of International Conference on Computational
Science and Applications, Algorithms for Intelligent Systems, Pune, 2019.
[20] S. H. Deshmukh, D. Bajaj, and S. Bhirud, "Audio Descriptive Analysis of Singer and
Musical Instrument Identification in North Indian Classical Music," IJRET:
International Journal of Research in Engineering and Technology, vol. 04, no. 06, pp.
505-508, June 2015.
[21] P. Chordia, "Segmentation and Recognition of Tabla strokes," in 6th International
Conference on Music Information Retrieval, London. UK, September 11-15 2005.
[22] P. A. Manoj Kumar, J. Sebastian, and H. A. Murthy, "Musical Onset Detection on Carnatic
Percussion," IEEE, 2015.
[23] J. P. Bello, L. Daudet, and S. Abdallah, "A Tutorial on Onset Detection in Music Signals,"
IEEE Transactions on Speech and Audio Processing, p. 13, 2005.
[24] P. R. Kshirsagar and D. G. Akojwar, "Classification and Prediction of Epilepsy using
FFBPNN with PSO," in 2015 IEEE International Conference on Communication
Networks (ICCN), Gwalior, 2015.
[25] L. G. Tampelini, C. Boscarioli, S. M. Peres and S. C. Sampaio, "An Application of Elman
Networks In Treatment and Prediction of Hydrologic Time Series," Learning and
Nonlinear Models (L&NLM) – Journal of the Brazilian Neural Network Society, vol. 9,
no. 3, pp. 148-156, 2011.
[26] W. Jia, D. Zhao, T. Shen, Y. Tang, and. Y. Zhao, "Study on Optimized Elman Neural
Network Classification Algorithm Based on PLS and CA," Computational Intelligence
and Neuroscience, vol. 2014, 2014.

View publication stats