Comparative Study of Different Techniques in Speaker Recognition: Review

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-3, Issue-3, Mar- 2017]
https://dx.doi.org/10.24001/ijaems.3.3.25 ISSN: 2454-1311
www.ijaems.com Page | 284
Comparative Study of Different Techniques in
Speaker Recognition: Review
Sonali T. Saste1
, Prof. S. M. Jagdale2
1
Department of Electronics and Tele-Communication Engineering, Bharati Vidyapeeth College of Engg. For Women, Pune,
Maharashtra, India
2
Department of Electronics and Tele-Communication Engineering, Bharati Vidyapeeth College of Engg. For Women, Pune,
Maharashtra, India
Abstract— The speech is most basic and essential method
of communication used by person.On the basis of
individual information included in speech signals the
speaker is recognized. Speaker recognition (SR) is useful
to identify the person who is speaking. In recent years
speaker recognition is used for security system. In this
paper we have discussed the feature extraction techniques
like Mel frequency cepstral coefficient (MFCC), Linear
predictive coding (LPC), Dynamic time wrapping (DTW),
and for classification Gaussian Mixture Models (GMM),
Artificial neural network (ANN)& Support vector
machine (SVM).
Keywords— ANN,DTW,GMM, LPC, MFCC.
I. INTRODUCTION
Speech is one of the most important ways to
communication. It gives many levels of information to the
listeners. It conveys the message; also give information
about gender, emotion and identity of speaker. There are
many situations where the correct recognition of speaker
is required. In biometric there are many ways to identify
the person like finger print, palm, iris, face, voice
recognition. The objective of the speaker
acknowledgment is to describe and separate the data from
discourse or speaker voice. Text independent and text
dependent speaker recognition are the types of speaker
recognition. [2]The behavioral aspect of human voice is
utilized for distinguishing proof by changing over a talked
expression from simple to computerized design, and
extracting unique vocal characteristics, such as pitch,
frequency, tone and cadenceto set up a speaker model or
voice test. In voice acknowledgment, enlistment and
confirmation procedures are included.Enrollment process
describes the registration of speaker by training his voice
features.[3]
II. RELATED WORK
There are many methodologies have been proposed for
speaker recognition. This is a system that can recognize a
person based on his/her voice. This is accomplished by
actualizing complex flag handling calculations that keep
running on an advanced PC or a processor.The speaker
recognition system can be classified as speaker
identification or speaker verification. [7]
1. Mel frequency cepstral coefficient (MFCC): This is
the most powerful feature extraction technique, used in
speaker recognition which works on human auditory
system. Seiichi Nakagawa used the MFCC for speaker
identification and verification.[1] Abdelmajid H. Mansour
also used MFCC for voice recognition.[3]
2. Linear predictive coding (LPC): LPC is the
simplified model for the speech production. It is a
technique for determining the basic parameters of speech
and provides precise estimation of speech parameters and
computational model of speech.Speech test can be
approximated as a straight mix of past speech tests is the
fundamental thought behind LPC. [11] KinnalDhameliya
used the LPC method for feature extraction in speaker
recognition.[2]
3. Dynamic time wrapping (DTW): DTW mainly
focuses on matching of two sequences of feature vectors.
It is used to calculate the distance between two time series
that vary in time.[3] Abdelmajid H. Mansour uses DTW
for feature extraction in voice recognition system.
III. FEATURE EXTRACTION
Feature extraction is the way toward distinguishing
distinctive elements from the information flag. After the
pre-processing this feature extraction is finished. For
speech there are many elements like MFCC, pitch,
energy, formant and so forth are extracted.
Fig.1: Basic speech recognition system
1. MFCC: MFCC is classical approach to analyze speech
signal, which represents the short-term power spectrum of
sound, based on linear cosine transform of a log power
spectrum on a nonlinear Mel scale of frequency. It is most

popular system because it approximates the human
auditory system response more closely than other
system.[1][8]
Fig.2: Steps involved in MFCC algorithm.
MFCC can be computed by using the formula,F(Mel) =
[2595* log 10 [1+F] /700]………….(1)The recognition
accuracy is high. That means the performance rate of
MFCC is high. MFCC captures main characteristics of
phones in speech. Low Complexity when using
MFCC.[18]
2. LPC:To decide the fundamental parameters of speech
and gives exact estimation of speech parameters and
computational model of discourse this LPC system is
utilized. Speech test can be approximated as a direct
blend of past speech tests is the fundamental thought
behind LPC.[11]The following fig. shows the steps
involved in the LPC feature extraction.
Fig.3: Steps involve in LPC feature extraction
Is a reliable, accurate and robust technique for providing
parameters which describe the time varying linear system
which represents the vocal tract.Computation speed of
LPC is good and provides with accurate parameters of
speech. LPC is useful for encoding speech at low bit
rate.[18]
3. DTW:This algorithm depends on element
programming which utilized for measuring similarity
between two time arrangement which may differ in time
or speed.This warping between two time series can then
be used to find corresponding regions between the two
time series or to determine the similarity between the two
time series.[6]The groupings are "warped" non-directly in
the time measurement to decide a measure of their
likeness free of certain non-straight varieties in the time
measurement. This succession arrangement technique is
frequently utilized as a part of time arrangement
grouping. In spite of the fact that DTW measures a
separation like amount between two given groupings, it
doesn't ensure the triangle disparity to hold.
Notwithstanding a closeness measure between the two
groupings, a purported "warping path" is delivered,by
warping according to this path the two signals may be
aligned in time.
DTW that is ordered into highlight based example
acknowledgment procedures does not have to set up (or
prepare) an arrangement display in progress; thusly, it is
for the most part seen as an adroitly straightforward and
coordinate acknowledgment strategy. Ordinary DTW is
by and large utilized for performing discourse
acknowledgment, and few reviews are believed to utilize
the DTW method for execute speaker acknowledgment.
[19]
4. Other feature extraction techniques: Various other
feature extraction strategies are accessible by basically
altering the above element extraction procedures,
including the following:
 Mean and Variance of the residual phase
 Delta and double delta of MFCC features
Along these many feature extraction strategies are
accessible. One can utilize any method as suitable by
application and can show signs of improvement
acknowledgment exactness. Presently these elements we
need to store in a database for N distinctive speakers and
this database will be utilized for classification.[2]
IV. CLASSIFICATION
1. GMM: A Gaussian mixture model is a probabilistic
model that accepts every one of the information focuses
are produced from a blend of a limited number of
Gaussian circulations with obscure parameters. One can
consider blend models as summing up k-cluster grouping
to join data about the covariance structure of the
information and additionally the focuses of the inactive
Gaussians.[10]
The Gaussian Mixture Model executes the desire
expansion (EM) calculation for fitting blend of-Gaussian
models. It can likewise draw certainty ellipsoids for
multivariate models, and register the Bayesian
Information Criterion to evaluate the quantity of bunches
in the information. A Gaussian Mixture. fit strategy is
given that takes in a Gaussian Mixture Model from
prepare information. Given test information, it can dole
out to each example the Gaussian it generally most likely
has a place with utilizing the Gaussian Mixture. predict
technique.[1][10]
GMM require less preparing and test data. It performs
better as it requires less measures of information to
prepare the classifier consequently memory prerequisite is
less.[19]

2. ANN: An artificial neural system (ANN) is a
computational model in light of the structure and
elements of organic neural systems. Data that moves
through the system influences the structure of the ANN in
light of the fact that a neural system changes - or learns, it
could be said - in view of that info and yield. ANNs are
viewed as nonlinear factual information demonstrating
apparatuses where the mind boggling connections
amongst data sources and yields are displayed or
examples are found.[5][8]ANNs is that element extraction
and speaker demonstrating can be consolidated into a
solitary system, empowering joint improvement of the
(speaker-dependent) include extractor and the speaker
model.[20]
3. SVM:The support vector machine is mainly used for
the speaker verification or speaker identification.SVM is
a twofold classifier which models the choice limit
between two classes as an isolating hyperplane. In
speaker check, one class comprises of the objective
speaker preparing vectors, and alternate class comprises
of the preparation vectors from an background data.
Utilizing the marked preparing vectors, SVM
streamlining agent finds an isolating hyperplane that
amplifies the edge of division between these two
classes.[19] In a high-dimensional space, the two classes
are less demanding to isolate with a hyperplane.
Instinctively, linear hyperplane in the high dimensional
portion include space compares to a nonlinear choice
limit in the first information space.[20]
V. CONCLUSION
In this paper we discuss feature extraction techniques and
classification techniques forspeaker recognition. For high
performance and low complexity MFCC is preferred.
With low bit rate for encoding LPC is useful.In DTW not
need to establish (or train) a classification model in
advance. In classification for less memory and less
preparing and test data GMM is useful.When there is
combination of feature extraction and speaker modeling
in to a single network, ANN is useful. In binary
classification SVM is simple and effective algorithm.
Depending upon the application one maychoose any
technique or it is also possible to use combination of one
or more techniques out of this described techniques for
better speaker recognition system.
ACKNOWLEDGMENT
I would like to express my gratitude towards Prof. Dr. H.
V. Vankudre, Principal, for encouraging us, also my
gratitude to all those who supported me for this paper
work.
REFERENCES
[1] Seiichi Nakagawa, Member, IEEE, Longbiao Wang,
Member, IEEE, and Shinji Ohtsuka “Speaker
Identification and Verification by Combining MFCC
and Phase Information”IEEE transactions on audio,
speech, and language processing, vol. 20, no. 4, may
2012.
[2] KinnalDhameliya, Ninad Bhatt “Feature Extraction
and Classification Techniques for Speaker
Recognition: A Review” 978-1-4799-7678-
2/15/$31.00 ©2015 IEEE
[3] Abdelmajid H. Mansour , Gafar Zen AlabdeenSalh,
Khalid A. Mohammed “Voice Recognition using
Dynamic Time Warping and Mel-Frequency
Cepstral Coefficients Algorithms”International
Journal of Computer Applications (0975 – 8887)
Volume 116 – No. 2, April 2015
[4] B. P. Bogert, J. R. Healy and J. W Tukey, “The
Quefrency Analysis of Time Series for Echoes:
Cepstrum, Pseudo-Autocovariance, Cross Cepstrum,
and Saphe Cracking,” Proceedings of the
Symposium on Time Series Analysis, 1963, pp. 209-
243.
[5] Alfredo Maesa, Fabio Garzia, Michele Scarpiniti
and Roberto Cusani, “Text Independent Automatic
Speaker Recognition System Using Mel- Frequency
Cepstrum Coefficient and Gaussian Mixture
Models,” Journal of Information Security, Vol. 3,
No. 4, October 2012, pp.335-3
[6] Pratik K. Kurzekar , Ratnadeep R. Deshmukh ,
Vishal B. Waghmare , Pukhraj P. Shrishrimal “A
Comparative Study of Feature Extraction
Techniques for Speech Recognition
System”International Journal of Innovative
Research in Science, Engineering and Technology
(An ISO 3297: 2007 Certified Organization) Vol. 3,
Issue 12, December 2014
[7] Miss. Sarika S. Admuthe, Mrs.
ShubhadaGhugardare “Survey Paper on Automatic
Speaker Recognition System” IJECS Volume 4 Issue
3 March, 2015 Page No.10895-10898
[8] Sundeep Sivan and Gopakumar C, “An MFCC
Based Speaker Recognition using ANN with
Improved Recognition Rate,” International Journal
of Emerging Technologies in Computational and
Applied Sciences (IJETCAS), Vol. 8, Issue 4, March-
May 2014, pp. 365-369.
[9] Jianglin Wang, AnJi and Michael T. Johnson,
“Features for Phoneme Independent Speaker
Identification,” IEEEInternational Conference on
Audio Language and Image Processing (ICALIP),
Shanghai, July 2012, pp. 1141-1145.

[10]Nidhi Desai, KinnalDhameliyaand Vijayendra
Desai, " Feature Extraction and Classification
Techniques for Speech Recognition: A Review,"
International Journal of Emerging Technology and
Advanced Engineering, Volume 3, Issue 12,
December 2013, pp. 367-371.
[11]Seiichi Nakagawa, Longbiao Wang and Shinji
Ohtsuka, “Speaker Identification and Verification by
Combining MFCC and Phase Information,” IEEE
transaction on audio, speech and language
processing, Vol. 20, No. 4, May 2012, pp. 1085-
1095. 13.
[12]Jianglin Wang, AnJi and Michael T. Johnson,
“Features for PhonemeIndependent Speaker
Identification,” IEEE International Conference on
Audio Language and Image Processing (ICALIP),
Shanghai, July 2012, pp. 1141-1145.
[13]Srikanth R Madikeri and Hema A Murthy, “Mel
Filter Bank Energy- Based Slope Feature and Its
Application to Speaker Recognition,” IEEE National
Conference on communication (NCC), Bangalore,
January 2011, pp. 1-4.
[14]Hemant A. Patil, Purushotam G. Radadia and T. K.
Basu, “Combining Evidences from Mel Cepstral
Features and Cepstral Mean Subtracted Features for
Singer Identification,” IEEE International
Conference on Asian Language Processing, Hanoi,
November 2012, pp. 145-148.
[15]S. Rajasekaran and G.A. VijayalakshmiPai, “Neural
Networks, Fuzzy Logic, and Genetic Algorithms
Synthesis and Applications”, PHI, 2003.
[16]Shahzadi Farah and AzraShamim, “Speaker
Recognition System Using Mel-Frequency
Cepstrum Coefficients, Linear Prediction Coding
and Vector Quantization,” 3rd IEEE International
Conference on Computer, Control &Communication
(IC4), Karachi, September 2013, pp. 1-5.
[17]Nidhi Desai, KinnalDhameliya and Vijayendra
Desai, “Recognizing voice commands for robot
using MFCC and DTW,” International Journal of
Advanced Research in Computer and
Communication Engineering, Vol. 3, Issue 5, May
2014.
[18]ShreyaNarang, Ms. Divya Gupta “Speech Feature
Extraction Techniques: A Review”IJCSMC, Vol. 4,
Issue. 3, March 2015, pg.107 – 114
[19]Ing-Jr Ding, Chih-Ta Yen and Da-Cheng Ou “A
Method to Integrate GMM, SVM and DTW for
Speaker Recognition”International Journal of
Engineering and Technology Innovation, vol. 4, no.
1, 2014, pp. 38-47
[20]TomiKinnunen, Haizhou Li “An overview of text-
independent speaker recognition:From features to
supervectors”T. Kinnunen, H. Li / Speech
Communication 52 (2010) 12–40

Comparative Study of Different Techniques in Speaker Recognition: Review

More Related Content

What's hot (20)

Similar to Comparative Study of Different Techniques in Speaker Recognition: Review (20)

Recently uploaded (20)

Comparative Study of Different Techniques in Speaker Recognition: Review