Saurabh Bhardwaj, Ph.D.

Thapar Institute of Engineering & Technology, Patiala Mobile: +91-7528981415, 7009691106

Electrical and Instrumentation Engineering Department Office: +91- 175 - 2393135
Punjab- 147004, India Email: [email protected]
Room Number: BC107 [email protected]

EDUCATION AND RESEARCH

July 2013 DELHI UNIVERSITY, NEW DELHI

NETAJI SUBHAS INSTITUTE OF TECHNOLOGY (NSIT)
PhD (Instrumentation and Control Engineering)
• Thesis: System Identification and Control using Hybrid Combination of Statistical and Soft-Computing
Techniques
• Problems:
▪ Text Independent Speaker Identification
▪ Pattern Similarity Based Clustering
▪ Time Series Prediction
▪ Solar Radiation Estimation
March 2008 PANJAB UNIVERSITY, CHANDIGARH
UNIVERSITY CENTRE OF INSTRUMENTATION & MICROELECTRONICS
M.Tech (Instrumentation)
• Thesis: Ultrasonic Parking Aid Device
• Secured 81% marks (2nd position)
• Achieved 1st position in oral presentation of research paper entitled:
i. Monitoring of Environment Using Fuzzy Logic in National Level Symposium on Electronics
Technology – 2008
ii. Fuzzy Logic based Virtual Instrument for Environmental Monitoring in 2nd Chandigarh
Science Congress-2008
• Achieved 2nd position in oral presentation of research paper entitled, “Ultrasonic Parking aid device,” in Ist
phase of Student Research Convention, ANVESHAN-2008
June 2001 V.B.S. PURVANCHAL UNIVERSITY, JAUNPUR
MEERUT INSTITUTE OF ENGINEERING & TECHNOLOGY
B. Tech (Electronics and Instrumentation)
• Thesis: Micro-Controller Based Moving Message Display

SPONSORED PROJECT

TITLE : Brain Fingerprinting: Detection of Physiological Based Concealed Information in the Brain of
Culprit using Machine Learning Techniques
FUNDING AGENCY : DRDO-INMAS (2016)
AMOUNT OF GRANT : Rs. 9.89 Lacs
ROLE : Co- Investigator

TITLE : Development of Speaker and Dialect Recognition models for Forensic Applications*
FUNDING AGENCY : DeitY
AMOUNT OF GRANT : Rs. 45,00,000
ROLE : Principal Investigator
• Working Group Recommended; Final sanction letter awaited
pg. 1
Saurabh Bhardwaj, Ph.D.
PUBLICATIONS

SCI JOURNALS
[1] S. Srivastava, Gopal, S. Bhardwaj. "Multi-scenario dataset for speaker recognition." Journal of
Intelligent & Fuzzy Systems 34, no. 3 (2018): 1385-1392.
(IMPACT FACTOR: 1.261)
[2] Gopal, S. Srivastava, S. Bhardwaj. "Feature Extraction Methods for Speaker Recognition: A Review."
International Journal of Pattern Recognition and Artificial Intelligence 31, no. 12 (2017): 1750041.
(IMPACT FACTOR: 0.660)
[3] G. Chaudhary, S. Srivastava, S. Bhardwaj and S. Bhargava, “Fusion of palm-phalanges print with
Palmprint and dorsal hand vein,” Applied Soft Computing, Elsevier, vol. 47, pp. 12-20, 2016.
(IMPACT FACTOR: 2.857)
[4] S. Bhardwaj, S. Srivastava, M. Hanmandlu, J.R.P. Gupta, “GFM Based Methods for Speaker
Identification,” IEEE Transactions on Cybernetics, vol.43, pp. 1047-1058, 2013.
(IMPACT FACTOR: 4.943)
[5] S. Bhardwaj, V. Sharma, S. Srivastava, O.S. Sastri, J.R.P. Gupta, S.S. Chandel, B. Bandyopadhyay,
“Estimation of solar radiation using a combination of Hidden Markov Model and Generalized Fuzzy
Model,” Solar Energy, Elsevier, vol. 93, pp. 43-54, 2013.
(IMPACT FACTOR: 3.685)
[6] S. Bhardwaj, S. Srivastava, J.R.P Gupta, “Pattern Similarity based model for Time Series Prediction,”
Computational Intelligence, Wiley, vol. 31, pp. 106-131, 2013.
(IMPACT FACTOR: 0.722)

INTERNATIONAL CONFERENCES:
[1] P. Bhola, S. Bhardwaj. "Solar energy estimation techniques: A review." In Power Electronics
(IICPE), 2016 7th India International Conference on, pp. 1-5. IEEE, 2016.
[2] Gopal, S. Srivastava, S. Bhardwaj, S. Srivastava. "Information Fusion in Animal Biometric
Identification." In Proceedings of the 5th International Conference on Frontiers in Intelligent
Computing: Theory and Applications, pp. 609-617. Springer, Singapore, 2017.
[3] G. Chaudhary, S. Srivastava, S. Bhardwaj, “Multi-level Fusion of Palmprint and Dorsal Hand Vein,”
in Proceedings of 3rd International Conference on Information Systems Design and Intelligent
Applications, Springer, pp. 321-330, 2016. (BEST PAPER AWARD)
[4] G. Chaudhary, S. Srivastava, S. Bhardwaj, P. Kiran, “Gaussian Membership Function-Based
Speaker Identification Using Score Level Fusion of MFCC and GFCC,” in Proceedings of the
International Congress on Information and Communication Technology: ICICT 2015, Springer, pp.
283—291, 2016
[5] S. Bhardwaj, S. Srivastava, “Choquet Fuzzy Integral Based Controller,” IEEE Conference on
Intelligent Systems and Control, Coimbatore, 2013
[6] S. Srivastava, S. Bhardwaj, J.R.P Gupta, “A novel clustering approach using shape based similarity,”
Intelligent Informatics, springer, pp. 17–27, 2013.
[7] S. Srivastava, S. Bhardwaj, A. Bhandari, K. Gupta, H. Bahl, J.R.P. Gupta, “Wavelet packet based mel
frequency cepstral features for text independent speaker identification,” Intelligent Informatics,
springer, pp. 237–247 2013.
[8] S. Srivastava, S. Bhardwaj, “A Novel Hybrid Model for Solar Radiation Prediction,” IEEE
Conference on Emerging Trends in Electrical Engineering and Energy Management, Chennai, 2012.

pg. 2
Saurabh Bhardwaj, Ph.D.
[9] S. Bhardwaj, S. Srivastava, J.R.P Gupta, A. Madhvan, “A novel shape based batching and prediction
approach for sunspot data using HMMs and ANNs,” IEEE Conference on Power Electronics, NSIT,
New Delhi, pp. 1-5, 28-30 Jan., 2011
[10] S. Bhardwaj, S. Srivastava, S. Vaishnavi, J.R.P Gupta, “Chaotic time series prediction using
combination of Hidden Markov Model and Neural Nets,” IEEE Conference on Computer Information
Systems and Industrial Management Applications, Gwalior, pp. 585-589, 8-10 Oct., 2010.
[11] S. Srivastava, S. Bhardwaj, A. Madhavan, J.R.P Gupta, “A novel shape based batching and
prediction approach for time series using HMMs and FISs,” IEEE Conference on Intelligent Systems
Design and Applications, Cairo, Egypt, pp. 929-934, Nov. 29- Dec 1, 2010

EXPERIENCE

2002-2016 TEACHING
June 2014 – THAPAR UNIVERSITY, PATIALA
Till Now ELECTRICAL AND INSTRUMENTATION ENGINEERING DEPARTMENT
Associate Professor
• Coordinated the Self-Assessment Report of NBA (Washington Accord Signatory) for UG EIC
• Member Self-Study Report ABET for UG EE
• UG Coordinator (EIC)
Feb. 2009 – GALGOTIAS UNIVERSITY, GREATER NOIDA
June 2014 SCHOOL OF ELECTRICAL, ELECTRONICS AND COMMUNICATION ENGINEERING
Assistant Professor
• Program Chair (Electronics & Communication Engineering) for B. Tech & M. Tech Programme
• Research Coordinator (School of Electrical, Electronics & Communication Engineering)
July 2009– NETAJI SUBHAS INSTITUTE OF ENGINEERING AND TECHNOLOGY (NSIT), NEW DELHI
Jan. 2013 DEPARTMENT OF INSTRUMENTATION & CONTROL ENGINEERING
Teaching Cum Research Fellow
June 2008 – DEWAN V.S INSTITUTE OF ENGINEERING & TECHNOLOGY, MEERUT
July2009 ELECTRONICS & INSTRUMENTATION ENGINEERING DEPARTMENT
Assistant Professor
• Head (Electronics & Instrumentation Engineering Department)
• Assistant Centre Superintendent (UPTU-SEE)
Aug. 2003 – RADHA GOVIND ENGINEERING COLLEGE, MEERUT
July 2006 ELECTRONICS DEPARTMENT (ECE & EIE)
Senior Lecturer
• Department Coordinator (ECE & EIC)
• Coordinator: Embedded System Training Program
Aug. 2002– IIMT ENGINEERING COLLEGE , MEERUT
Jul. 2003 ELECTRONICS AND COMMUNICATION ENGINEERING DEPARTMENT
Lecturer
2001– 2002 INDUSTRY
June 2001– HOTLINE SWITCHGEAR & CONTROLS, DELHI
July 2002 R&D Engineer
• Design and Development of Microcontroller based Single Phase Energy Meter
• Best performance award in R & D by the Managing Director of Hotline Switchgear & Controls
pg. 3
Saurabh Bhardwaj, Ph.D.
AWARDS

• Best Paper Award in Springer Conference: G. Chaudhary, S. Srivastava, S. Bhardwaj, “Multi-level fusion of palm
print and dorsal hand vain,” 3rd International Conference on Information Systems Design and Intelligent Applications,”
on Jan 9, 2016, Visakhapatnam, India.
• Best Presentation Award in IEEE workshop: S. Bhardwaj, S. Srivastava, Sanidhya, Mani, “Multi-Environment
Dataset for Speaker Identification,” Computational Intelligence: Theories, Applications and Future Directions on July
14, 2013 at IIT Kanpur, India.

SHORT TERM COURSES / WORKSHOPS

[1] Participated in five days Summer School on, “Speaker and Language Recognition,” from 8th - 12th July, 2017
at Dhirubhai Ambani Institute of Information and Communication Technology, Gandhinagar, India
[2] Participated in four days workshop on, “Fuzzy Sets, Fuzzy Logic and its Applications in Big Data Analytics" from
12-15 April, 2016 at Thapar University Patiala.
[3] Participated in , “Three Days Workshop On Proven Strategies for Effective Teaching ,” held on August 29 - August
31, 2013 at Stellar Gymkhana, Greater Noida, India
[4] Participated in , “Seven Days Workshop On Universal Human Values and Ethics,” held on July 29 - August 5, 2013
at Surya Group of Institutions, Lucknow, India
[5] Participated in , “Three Days Workshop On New Outcome-Based Education And Outcome Based Accreditation,”
held on May 2 - May 11, 2013 at GCET Auditorium, Galgotias University Campus One, Greater Noida, India
[6] Participated in , “One Day NBA webinar workshop on outcomes based Accreditation process,” held on April 29,
2013 at GCET Auditorium, Galgotias University Campus One, Greater Noida, India
[7] Participated in the workshop, “Conducting Design Science Research in information Technology,” held on
November 9, 2011 at Netaji Subhas Institute of Technology, New Delhi, India
[8] Worked as a rapporteur in, “India International Conference on Power Electronics (IICPE 2010)” held in Netaji
Subhas Institute of technology, New Delhi on January 28 - 30, 2011.
[9] Participated in the 54th congress of, “Indian Society of Theoretical and Applied Mechanics (ISTAM),” December
18-21, 2009 Netaji Subhas Institute of Technology, New Delhi as a volunteer for organizing the event
[10] Participated in the refresher course on, “LATEX and MATLAB,” held in Netaji Subhas Institute of Technology,
New Delhi on July 12 – 24, 2010
REVIEWER/ PROGRAM COMMITTEE MEMBER/SESSION CHAIR
[1] Session Chair in, “7th IEEE India International Conference on Power Electronics,” held at TIET, Patiala, India from
17-19 Nov. 2016.
[2] Reviewer: IEEE Transactions on system, man and cybernetics
[3] Reviewer: Solar Energy, Elsevier
[4] Programme committee member and reviewer of, “IEEE International Conference on World Congress on
information and communication technologies,” held on Mumbai, India from 11-14 December – 2011.
[5] Programme committee member and reviewer of, “IEEE International Conference on Computational Intelligence
and communication Networks,” held on Gwalior, India from 07-09 October – 2011.

SAURABH BHARDWAJ

pg. 4
Saurabh Bhardwaj, Ph.D.
 Research Statement

My research interests are focused on the development and improvement of statistical and machine
learning based solutions for the problems that involve recognition, classification, clustering, modelling
and information retrieval.
As part of my Ph.D. thesis, I have investigated different statistical and soft-computing models
and their combinations for different applications including speaker recognition, clustering, solar
radiation estimation, control and for time series forecasting. The idea of combining different models is
that their individual weaknesses can be compensated by their individual merits while solving complex
problems. During this research, I have used the statistical techniques such as Hidden Markov Model
(HMM) and Gaussian Mixture Model (GMM) and soft-computing techniques such as Artificial Neural
Network (ANN), Fuzzy Logic (FL) and Choquet Fuzzy Integral(CFI).
In the paragraphs that follow, I will describe my contributions to each of the areas stated above
and give the related and other research directions that I am planning to pursue in the future.
Early Research

Speaker Recognition.
There are a number of situations in which correct recognition of persons is required. The use of biometric-
based recognition is the most “natural” way of recognizing a person. This is also very safe as these
characteristics cannot be stolen or forgotten. One of the benefits of using speech over other biometric
traits is that it can be used remotely and it contains both the physiological and behavioural
characteristics of humans.
I have developed three approaches for the text independent speaker identification of which two
methods utilize both the continuous density hidden Markov model (HMM) and the generalized fuzzy
model (GFM), which has the advantages of both Mamdani and Takagi–Sugeno models.
In the first approach, the HMM is utilized for the extraction of pattern similarity-based batch
feature vector that is fitted with the GFM to identify the speaker. Next HMM is applied on the feature
data of all users to yield the HMM parameters assuming a certain number of states and Gaussian
mixtures. The feature data are clustered into batches by sorting the Log Likelihood (LL) values obtained
from the forward algorithm. The sorted LL values are thresholded into batches by finding the correlation
coefficients. When this is done, the features of different users fall into different batches. Hence, each
feature is identified with the associated LL and the user number. These batches are then fitted with
GFM. When the test feature vector arrives, its LL is found by the learned HMM parameters. This
indicates which model it belongs to. The output of the GFM gives the identity of the speaker.
In the second approach, the equivalence between the defuzzified output of the GFM and the
conditional mean of the GMM under certain conditions is used for the identification of speakers. In this
the parameters of the GFM are calculated with the help of GMM. Here, GMM–GFMs as many as the
number of users are found using the feature data of all users with the difference that the output values
corresponding to the user are taken as unity with all other output values remaining zero. When the test
feature arrives, the outputs of all the speaker models are evaluated, and the model with the highest
output gives the user identity.
Finally, the third method has been inspired by the way humans cash in on the mutual
acquaintances while identifying a speaker. In this, several sets of HMM models are created with different
initial parameters such as the number of states and the number of Gaussian mixtures. Out of these, any
two sets of HMM models that identify most of the speakers correctly but give different misidentified
speakers are selected. If all the misidentified speakers are different, then we term it as zero similarity in
the misidentification. If one of the misidentified speakers is the same, then we have one similarity in the
misidentification and so on. When the test feature vector arrives, the outputs are evaluated with the two
sets of HMMs. If both the sets of models provide the same output for a particular speaker, then we get
the speaker’s identity. If not, we calculate the output from the GFM to distinguish between the two.

pg. 5
Saurabh Bhardwaj, Ph.D.
Solar Radiation Estimation.
Owing to considerable advancement during last two decades, solar photovoltaic has become a reliable
technology for power production worldwide. For solar radiation estimation, a model is developed which
uses HMM with Pearson R model for the extraction of shape-based clusters from input meteorological
parameters and it is then processed by GFM to accurately estimate the solar radiation. The estimation
method used in this work exploits the pattern identification prowess of HMM for cluster selection and
generalization and nonlinear modelling capabilities of GFM to predict the solar radiation.

Time Series Prediction.

Time series prediction is a topic of widespread research interest in various fields such as econometrics,
business, biology, physics, and meteorology. A pattern similarity-based clustering approach is developed
which is used for time series prediction. In this approach instead of using distance function, shape of the
sequence is used as similarity index for clustering, which overcomes few of the shortcomings associated
with distance-based clustering approaches.

Choquet Fuzzy Integral Based Controller.

Fuzzy measures and fuzzy integrals are derived from fuzzy set theory and classical measure theory. The
main characteristic of measure theory is additivity. This property is very influential but suffers from the
difficulty of rigidness. To mitigate this rigidness problem fuzzy measure is proposed.
A Fuzzy measure is needed on a finite Set-X on which a fuzzy integral can be ex-pressed as a
computational scheme to integrate all values from the individual subsets non-linearly. For the design of
controller, a fuzzy integral is incorporated in the consequent part of a fuzzy rule. Unlike simple Takagi–
Sugeno (T-S) Fuzzy modelling, where the antecedent part consists of input fuzzy sets and the consequent
part is linear addition of inputs, the consequent part of rule in the approach makes use of Choquet
Integral and fuzzy measure to combine the input information non-linearly. Here, fuzzy integral was used
as a neuro–computation in the context of fuzzy model-based control. For non-additive fuzzy systems, the
consequent part of fuzzy rule in T-S model is replaced with Choquet Integral.

Present Research:

An Information Set-based Text-Independent Speaker Authentication.

At present, I am working on the development of Two-fold Information Set features (TFIS) for robust text-
independent speaker identification. These features are based on the Information Set theory. In the first
phase the audio signal is partitioned into frames to obtain MFCC from each frame. From these
coefficients, MFCC matrix is formed such that each row of this matrix corresponds to a dimension and
each column corresponds to a frame. This matrix representation facilitates the derivation of TFIS
features both from frames that yield the spatial information and dimensions that yield the temporal
information. Thus, at each position in the matrix there are two types of information components adding
which TFIS features are obtained. The TFIS features comprising their combination of two components
are less in number thus reducing the computational time and complexity.
The proposed approach will be validated on LDC2017S06: 2010 NIST Speaker Recognition
Evaluation test set which we have recently obtained through LDC dataset scholarship program.

Multi-scenario dataset for speaker recognition.

I along with my undergraduate students have collected a speech database at the institute in different
scenario for text independent speaker identification in the Indian context. In order to get the Multi-
Scenario dataset, each speaker performed multiple sessions recording in reading style with English and
Hindi language with same passages but under different conditions. Here four different scenarios are
considered; sensor and environment, language, aging and health. To study the effect of sensor, language
and environment on the performance of ASR system a database of 200 speaker was created. Under
different environmental conditions, four different types of sensors in parallel configuration were used to
study the sensor mismatch conditions over testing and training phase. The database contains speech

pg. 6
Saurabh Bhardwaj, Ph.D.
samples of the individual in English and Hindi in read speech styles under two environments i.e. a
controlled recording chamber and library. To study the aging effect, an aging speaker database of 53
famous personalities was collected from online source varying over a period of 10–20 years.

Future Research Directions:

Some of the research directions which I will pursue are as follows:

Speaker Recognition.
As part of my early work in speaker recognition I have worked on the development of models for biometric
speaker recognition but as a part of my future research plan I would like to extend this work for forensic
speaker recognition. The models developed for biometric speaker recognition are not suitable for forensic
case. For the law enforcement agencies, a ‘confidence measure’ is required for speaker recognition. In
this, probabilistic certainty level of correctness is addressed for every decision based on statistics with
known error rates generated from large sample populations.
Normally, the speech samples which are received for forensic examination and comparison are in
conversational form so it is my plan to develop models for identification and verification in forensic
scenario for the speaker present in a conversational speech along with the conventional task of speaker
detection where speech by only a single speaker is present. For the development of models in forensic
scenario my plan is to apply the following techniques:

a. To model the feature vectors with the help of non-additive fuzzy systems in which the consequent
part of fuzzy rule in T-S model will be replaced by Choquet Integral.
b. To use the information set theory for the extraction of features.
c. To use deep learning for the fuzzy and neural networks.

Dialect Identification.
My further plan is to develop models for dialect recognition for forensic applications. As the speech of a
person belonging to a dialectal group always has its specific dialectal accent and this shows the potential
in the investigation of crime as accented speech carries linguistic information regarding the regional
dialect of an individual. Classification of speakers by their dialects is of particular interest if a suspect is
classified according to his/her dialect then this would assist in police investigation in narrowing down
the search region and can help in the person identification tasks. Normally, the speech samples which
are received for forensic examination and comparison are in conversational form in which multiple
speakers are present with channel variability. Therefore, a model for the identification of dialects is
proposed for single and conversational speech. Approximately, 600 million people across the globe speak
Hindi with about 200 different regional dialects. It is impossible to collect a single database having all
possible speaking styles, channel variation, noises and dialectal information. So, keeping in view the
crime investigation and law enforcement cases in the Indian scenario it is also in my plan to develop a
database of dialects of Hindi language in various forensic situations. The work will be helpful at any
phase of justice, for police investigation, for investigating agencies, and for researchers working in the
area of dialect recognition.

Emotion Recognition through speech.

Finally, one of my future plans is to work on speaker-independent emotion recognition system. Humans
can easily recognize the underlying emotion in the speech spoken by the other person. Recognition of
emotions have applications in many fields like medicine, music therapy, E-learning, monitoring,
entertainment, law, marketing etc. Speech signals can be used for detecting emotions where a person is
not in direct contact; for example, in call centre applications and mobile communication. Particularly,
emotion recognition from speech signals has the utility in areas that require natural man–machine
interaction such as web movies and computer tutorial applications. For the application of automatic
translation systems, the knowledge of emotional state of the speaker also plays an important role.

pg. 7
Saurabh Bhardwaj, Ph.D.