Towards an Optimal Speaker Modeling in Speaker Verification Systems using Personalized Background Models

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 7, No. 6, December 2017, pp. 3655~3663
ISSN: 2088-8708, DOI: 10.11591/ijece.v7i6.pp3655-3663  3655
Journal homepage: http://iaesjournal.com/online/index.php/IJECE
Towards an Optimal Speaker Modeling in Speaker Verification
Systems using Personalized Background Models
Ayoub Bouziane, Jamal Kharroubi, Arsalane Zarghili
Laboratory of Intelligent Systems and Applications, University Sidi Mohamed Ben Abdellah
Fez, Morocco.
Article Info ABSTRACT
Article history:
Received Mar 19, 2017
Revised Jun 29, 2017
Accepted Jul 14, 2017
This paper presents a novel speaker modeling approachfor speaker
recognition systems. The basic idea of this approach consists of deriving the
target speaker model from a personalized background model, composed only
of the UBM Gaussian components which are really present in the speech of
the target speaker. The motivation behind the derivation of speakers’ models
from personalized background models is to exploit the observeddifference
insome acoustic-classes between speakers, in order to improve the
performance of speaker recognition systems.
The proposed approach was evaluatedfor speaker verification task using
various amounts of training and testing speech data. The experimental results
showed that the proposed approach is efficientin termsof both verification
performance and computational cost during the testing phase of the system,
compared to the traditional UBM based speaker recognition systems.
Keyword:
Speaker modeling
Speaker verification
Gaussian mixture models
Universal background model
Maximum a posteriori (MAP)
Personalized background
models
Copyright © 2017Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Ayoub Bouziane,
Laboratory of Intelligent Systems and Applications,
Sidi Mohamed Ben Abdellah University,
P.B: 2202 Immouzer Road, Fez, Morocco.
Email: ayoub.bouziane@usmba.ac.ma
1. INTRODUCTION
The GMM-UBM based speaker modeling approach was firstly introduced to the speaker recognition
community by Reynolds, in 2000 [1]–[3]. Since then, it have become the predominant approach for speaker
modeling in text-independent speaker recognition systems, and the basis of the most successful approaches
that have been emerged in the last decade: the hybrid GMM-SVM approach [4], [5], the joint factor analysis
approach [6]–[8], and the recently introduced i-vectors approach [9], [10]. The motivation behind the use of
adapted Gaussian mixture models for speaker modeling is generally based on the assumption that Gaussian
densities may model a set of hidden acoustical classes that reflect some general speaker dependent vocal tract
characteristics.
The main idea of the UBM-based speaker recognition systems consist of deriving the target speaker
model from a universal background model that represents the set of all human acoustic classes (e.g., gender
dependent acoustic classes, age-dependent acoustic classes, accent dependent acoustic classes…). The target
speaker modelwill therefore be composed of an adapted version of the various acoustic classes defined in the
universal background model. However, as it sounds to the human ear, speakers don’t share the same acoustic
classes. For example, the acoustic classes of male speakers are different from those of female speakers and
the acoustic classes of adult speakers are different from those of minor speakers, as well as, the acoustic
classes of timid speakers are different from those of bold speakers, etc. Thereby, it seems that it is not logical
to model a speaker by an adapted version of anacoustic class which may not exist in its voice. Furthermore,
the difference of acoustic classes between speakers can be exploited to discriminate between them in speaker
recognition systems.

 ISSN:2088-8708
IJECE Vol. 7, No. 6, December 2017 : 3655–3663
3656
In view of these observations, the present study attempts, firstly,to experimentally investigate the
degree of acoustic-classes’difference between speakers,and secondly, to propose a speaker modeling
approach that takes into account these observations and exploits the differences in acoustic classes between
speakers to improve the performance of speaker recognition systems.
The remainder of this paper is organized as follows. The first section gives a brief overview of the
traditional GMM-UBM based Speaker verification systems. Next, the proposed speaker modeling approach,
based on personalized background models, is introduced. Afterward, experimental resultsand discussions are
presented in the third section. Finally, conclusions and future research directions are drawn in the last section.
2. TRADITIONAL GMM-UBM BASED SPEAKER VERIFICATION SYSTEMS
The GMM-UBM based speaker verification system was originally proposed by Reynolds in 2000
[1], [11]. Since then, it has become the predominant approach for speaker modeling in text-independent
speaker recognition systems and the basis of the most successful approaches that have been emerged in the
last decade. The main idea of the GMM-UBM approach consists, as shown in Figure 1, of deriving speakers’
models from a universal background model using maximum a posteriori (MAP) adaptation [1]. The
Universal Background Model (UBM) is typically a Gaussian mixture model (GMM) that represents the
distribution of the entire acoustic space of speech [2], [3].
Figure 1. A MAP adaptation of a GMM comprises 5 Gaussian densities. Original Gaussian densities of the
UBM are depicted as unfilled ellipses (dotted line), whereas the adapted Gaussian densities are denoted by
filled ellipses, and the observed feature vectors are depicted as small circles
The MAP adaptation process is generally composed of two steps. First, the sufficient statistics
estimates of the speaker’s training data are computed for each mixture in the UBM. Next, the computed
sufficient statistics estimates are combined with the old sufficient statistics from the UBM mixture
parameters using a data-dependent mixing coefficient. The specifics of the adaptation are as follows. Given
that the universal background model is composed of M Gaussian components, each of which is
parameterized by a mean vectorμi, variance matrix σ2
i and its weightwiin the mixture model. Initially, the
posteriori probability of each UBM component {μi, σi, wi} given the feature vector xt is computed as follows:
𝑃𝑟(𝑖|𝑥𝑡) =
𝑤𝑖 𝑝𝑖(𝑥𝑡)
∑ 𝑤𝑖 𝑝𝑗(𝑥𝑡)𝐿
𝑗=1
(1)
where pi(xt) is the density of the probability, given by:
𝑝𝑖(𝑥𝑡) =
1
(2𝜋)
𝐷
2⁄
|𝜎𝑖|
1
2⁄
𝑒𝑥𝑝 {−
1
2
(𝑥𝑡 − 𝜇𝑖) 𝑇
𝜎𝑖
−1
(𝑥𝑡 − 𝜇𝑖)} (2)
The sufficient statistics for the weight, mean, and variance parameters are then computed as follows:
Universal Background Model Universal Background
Model Specific Speaker ModelTesting phase
MAP Adaptation

IJECE ISSN: 2088-8708 
Towards an Optimal Speaker Modeling in Speaker Verification Systems … (Ayoub Bouziane)
3657
𝑛𝑖 = ∑ 𝑃𝑟(𝑖|𝑥𝑡)
𝑇
𝑡=1
; 𝐸𝑖(𝑥) =
1
𝑛𝑖
∑ 𝑃𝑟(𝑖|𝑥𝑡)
𝑇
𝑡=1
𝑥𝑡 ; 𝐸𝑖(𝑥2
) =
1
𝑛𝑖
∑ 𝑃𝑟(𝑖|𝑥𝑡)
𝑇
𝑡=1
𝑥𝑡
2 (3)
Thereafter, the computed sufficient statistics are used for estimating the adapted mixture weights wî,
means μî and variances σî of the given speaker:
𝜇𝑖̂ = [𝛽𝑖
𝜇
𝐸𝑖(𝑥) + (1 − 𝛽𝑖
𝜇
)𝜇𝑖] (4)
𝑤𝑖̂ = [𝛽𝑖
𝑤
𝑛𝑖)/𝑇 + (1 − 𝛽𝑖
𝑤
)𝑤𝑖]𝛾 (4)
𝜎2̂ 𝑖 = [𝛽𝑖
𝜎2
𝐸𝑖(𝑥2) + (1 − 𝛽𝑖
𝜎2
)(𝜎2
𝑖 − 𝛽𝑖
𝜎2
) − 𝜇𝑖̂2
] (5)
with, 𝛽𝑖
𝜌
= ni/(ni + 𝑟 𝜌
), ρ ∈ {w, 𝜇, 𝜎2
} (6)
Here, γ is a scale factor computed over all adapted mixture weights to ensure that they sum to unity,
βi
ρ
, ρ ∈ {w, μ, σ2} are the adaptation coefficients that control how the adapted GMM parameters will be
affected by the observed speaker data, and rρ
is a fixed relevance factor for parameter ρ.
An overall diagram of the GMM-UBM based speaker verification system is shown in Figure 2. The
basic operating structure of the system, as shown in Figure 2, is composed of three phases: the training phase,
the enrollment phase and the testing phase. During the first phase, i.e. the training phase, a large collection of
speech utterances is collected from a background population of speakers, their corresponding feature vectors
are extracted and used to train the universal background model. The training process of the UBM is done
generally using maximum likelihood estimation via the EM algorithm. In the second phase, i.e. the
enrollment phase, speaker models of new client speakers are derived from the universal background model
through MAP adaptation using the speakers’ training feature vectors [13]. While in the testing phase, the
extracted feature vectors of the unknown speaker’s utterance Xu
= {xu
1, xu
2, . . ., xu
N} are compared against
both the claimed target speaker model and the background model. The log likelihood ratio
LLR(Xu
; λspk, λUBM) between the claimed speaker model and the universal background model is then
calculated and used to make a decision about the acceptance/rejection of the claimed identity. The log
likelihood ratio (LLR) of the test utterance Xu
between the speaker model λj and the UBM model λUBM:
𝐿𝑅(𝑋 𝑢
; 𝜆 𝑠𝑝𝑘, 𝜆 𝑈𝐵𝑀) =
1
𝑁
[∑ 𝑙𝑜𝑔 𝑝(𝑥 𝑢
𝑖|𝜆 𝑠𝑝𝑘)
𝑁
𝑖=1
− ∑ 𝑙𝑜𝑔 𝑃(𝑥 𝑢
𝑖|𝜆 𝑈𝐵𝑀)
𝑁
𝑖=1
] (7)
With, 𝑝(𝑥𝑡|𝜆) = ∑ 𝑤𝑖 𝑏𝑖(𝑥𝑡)𝑀
𝑖=1 (8)

 ISSN:2088-8708
3658
Figure 2. Block diagram of the GMM-UBM based speaker verification system
3. SPEAKER RECOGNITION USING PERSONALIZED BACKGROUND MODELS (PBMS)
The main idea of the proposed PBM-based speaker modeling approach consistsof adapting the
target speaker model from a personalized background model (PBM), composed only of the UBM Gaussian
components which are actually present in the speaker’s speech. The MAP adaptation step of traditional UBM
based systems will, therefore, be preceded by a selection step that selectsthe background Gaussian
component swhich reflect the general form of the acoustic classes characterizing the speaker, see Figure 3.
Figure 3. A MAP adaptation of a GMM comprises 5 Gaussian densities. Original Gaussian densities of the
UBM are depicted as unfilled ellipses (dotted line), whereas the adapted Gaussian densities are denoted by
filled ellipses, and the observed feature vectors are depicted as small circles.
Given a UBM of M Gaussian component λUBM = {μUBMi
, σUBMi
, wUBMi
}/ i ∈ {1,2, … , N}, and M
training feature vectors X = {x1, x2, … , xN}, extracted from the target speaker’s speech.The PBMGaussian
components are generally chosen from the UBM using a winner-take-all based strategy. A UBM Gaussian

IJECE ISSN: 2088-8708 
3659
component θUBM 𝑖
= (μUBM 𝑖
, σUBM 𝑖
, wUBM 𝑖
) isselected to belong to the personalized background model λPBM
of the target speaker, if there is at least one feature vector xn ∈ X, where the UBM Gaussian component
θUBMi
achieves the maximum posterior probability of xn belongingness:
𝑃𝑟(θ 𝑈𝐵𝑀 𝑖
|𝑥n) ≥ 𝑃𝑟(θ 𝑈𝐵𝑀 𝑙
|𝑥n) , ∀𝑙 ∈ {1,2, … , N} (9)
Once the PBM Gaussian components are selected, the weights w 𝑈𝐵𝑀j
of the selected components are
divided by their sum so that the total weight is equal to unity.
A block diagram of the speaker modeling process using personalized background models is shown
in Figure 4. Firstly, the training feature vectors of the target speaker are extracted from itsenrollment
utterances. Next, the extracted feature vectors are used to select the UBM Gaussian components which will
compose the speaker’s personalized background model.Afterwards, the composed personalized background
model is utilized do derive the speaker model using the MAP adaptation procedure. Finally, the adapted
model is stored together with the corresponding indices of the PBM Gaussian components in the UBM.
An example of a two-dimensional projectionof two speakers’ features and the means of their
corresponding UBM and PBM adapted modelsis shown in Figure 5. As it can be seen from this figure, the
means of the PBM adapted models fit the speakers’ features better than the means of the traditional UBM
adapted models. Moreover, it seems that the adaptation of the UBM Gaussian components which haven't any
relationship with the target speakerinfluence on the feature vectors belongingness to the appropriate Gaussian
components, which therefore affect negatively the adapted model.
Figure 4. Block diagram of speaker modeling process using personalized background models
The speaker’s feature vectors The Universal Background Model
The UBM adapted model The PBM adapted model

 ISSN:2088-8708
3660
Figure 5. A two-dimensional projection of two speakers’ features (blue points), the means of their
corresponding UBM and PBM adapted models (in green and yellow points, respectively) and the means of
the UBM model (red points)
During the test phase, the log-likelihood ratio LLR(Xu
; λspk, λPBM) between the claimed speaker
model and the personalized background model is used to make a decision about the acceptance or the
rejection of the claimed identity:
The motivation behind the use of the personalized background modelinstead of the universal
background model is to penalize the decision score of impostor speakers who don’t share the same acoustic
classes with the target speaker.
4. EXPERIMENTS, RESULTS AND DISCUSSION
4.1. The Experimental Protocol
The performed experiments in this study were conducted on the THUYG-20 SRE database [14].
This database isgenerally composed of 353 speakers, collected in a controlled environment (silent office by
the samecarbon Microphone). The entire speech corpuswas divided into three data sets: the first dataset
consists of 200 genderbalanced speakers (100 Male and 100 Female) and devoted totrain the Universal
Background Model (UBM), the second and the third data sets are composed of the same set of 153 client
speakers.The first dataset comprises the training speech data, whereas the second comprises the testing
speech data of the 153 speakers.During the testing phase of the system, the client speakers were tested
against each other,resulting in total of 896,886 trials of 4 seconds and 441,558 trials of 8 seconds.
The feature vectors of the overall speech utterances were extracted using the MFCC approach: the
digitized speech is firstly emphasized using a simple first order digital filter with transfer function H (z) = 1 –
0.95z. Next, the emphasized speech signal is blocked into Hamming-windowed frames of 25 ms (400
samples) in length with 10 ms (160 samples) overlap between any two adjacent frames.Finally, 19 Mel-
Frequency Cepstral Coefficients were extracted from each frame [15]. During the training phase, a universal
background model (UBM) of 1024 Gaussian components was trained on the overall training data (7.5 hours
of speech) using the EM algorithm.
4.2. Investigation on the Degree of Difference in Acoustic Classes between Speakers
Theaim of the performed experiments in this sectionisto investigate the degree of difference in
acoustic classesbetween speakers. For this purpose, we have carried out several speaker verification
experiments in which we have basedonly on the difference between the acoustic classes present in the target
speaker’s speech and those present in the claimed speaker’s speech. To proceed, we have represented the
training speech utterance(s) of the target speaker and the test speech utterance of the claimed speaker by
histograms. Each histogramis composed of 1024 bins, whereeach bin represents aUBM Gaussian component.
𝐿𝑅(𝑋 𝑢
; 𝜆 𝑠𝑝𝑘, 𝜆 𝑃𝐵𝑀) =
1
𝑁
[∑ 𝑙𝑜𝑔 𝑝(𝑥 𝑢
𝑖|𝜆 𝑠𝑝𝑘)
𝑁
𝑖=1
− ∑ 𝑙𝑜𝑔 𝑃(𝑥 𝑢
𝑖|𝜆 𝑃𝐵𝑀)
𝑁
𝑖=1
] (10)
With, 𝑝(𝑥𝑡|𝜆) = ∑ 𝑤𝑖 𝑏𝑖(𝑥𝑡)𝑀
𝑖=1 (11)

IJECE ISSN: 2088-8708 
3661
The value of each bin is defined as the number of times that the corresponding UBM Gaussian component
has the maximum posterior probability over the feature vectors of the speaker. The specifics of the histogram
construction process are shown in Algorithm 1.
Algorithm 1 The histogram construction process
Input:The feature vectorsX = {x1, x2, … , xN} of the speech utterance,
The universal background model λUBM = {μUBMi
, σUBMi
, wUBMi
}/ i ∈ {1,2, … , N}.
Output: The corresponding histogram ℋof the speech utterance.
ℋ = 𝑧𝑒𝑟𝑜𝑠(1,1024)
𝐅𝐎𝐑 𝐄𝐀𝐂𝐇 𝒙𝒊 𝐈𝐍 𝐗
𝑗 = 𝑎𝑟𝑔 𝑚𝑎𝑥
𝑖
𝑃𝑟(𝑖|𝑥𝑡)
ℋ(𝑗) = ℋ(𝑗) + 1
END
Once the histogram of the target and the claimed speakers are constructed, ℋT and ℋC respectively, the
comparison between themis done using the Bhattacharyya distance:
D(ℋT, ℋC) = − log ∑ √ℋT(𝑖). ℋC(𝑖)
1024
𝑖=1
(12)
The computed distance D(ℋ 𝑇, ℋC) is then used to make a decision about the acceptance or the
rejection of the claimed speaker.The obtained results of the performed experiments, while varying the
amount of training and testing speech data, are shown in Table1.
Table 1. The obtained EERs using several amount of training and testing speech data.
Amount of Enrollment Speech Data
20 seconds 40 seconds 60 seconds
Amount of
Testing
Speech Data
4 seconds 6.38 5.90 4.64
8 seconds 5.86 5.31 3.98
First and foremost, as it can be seen from Table 1, the obtained results are highly encouraging. The
lower obtainedequal error rates, based only on the difference in hidden acoustic classes between speakers,
reflect a great difference in those hidden acoustic classes between speakers. Additionally, it appears that each
increase in the amount of training or testing speech data is translated into better verification performance.This
proportional relationshipbetween the amounts of speech data and the performance of the system reflects the
fact that the overall acoustic classes of a speaker cannot be assembled in its pronunciation of one or two
utterances.
4.3. Assessment of the Proposed GMM-PBM Approach Compared to the Traditional GMM-UBM
Approach
The performed experiments in this section attemptto assess the performance of the proposed GMM-
PBM based speaker modeling approach, compared to the traditional GMM-UBM based approach. Hence,
various experiments were carried out using the two approaches while varying the amount of training and
testing speech data. The obtained results are illustrated in Figure 6.
The experimental results show, across the various amounts of training and testing speech data that
our proposed approach has achieved a better verification performance compared to the traditional UBM
based approach. Even in little amounts of training speech data, where the speaker’s acoustic classes may not
be all present, our proposed approach demonstrates its enhanced verification performance as compared with
the UBM based approach. Furthermore, we can see that the obtained performance under the UBM based
system using test utterances of 8 seconds was obtained under our proposed PBM based system using test
utterances of only 4 seconds. Moreover, it can be seen that the relative error reduction was doubled when we
have doubled the amount of testing utterances.
In addition to its performance advantage, the proposed approach can significantly reduce the CPU
time required for speaker verification during the testing phase of the system compared to the traditional
system, see Figure 7. In fact, the derivation of speakers’ models from personalized background models
reduces the order of their adapted models, which consequently, reduces theirstorage and computational costs.

 ISSN:2088-8708
3662
Figure 6. The obtained Equal Error Rates using test utterances of 4 seconds (left figure) and 8 seconds (right
figure)
Figure 7. CPU time required for speaker verification using the UBM based approach and the PBM based
approach
5. CONCLUSION AND FUTURE RESEARCH DIRECTIONS
The aim of the present study was two-fold. The first aim was to investigate the degree of difference
in hidden acoustic-classes between speakers. The second aim was to propose a novel speaker modeling
approach that exploits this difference to improve the performance of speaker recognition systems. The
findings of the study revealed that there is a great difference in hidden acoustic classes between speakers.
Additionally, the evaluation of the proposed approach demonstrates its efficiency in terms of both
verification performance and computational cost during the verification phase of the system, compared to the
traditional approach.Future researchwill concentrate on applying the proposed approach within the hybrid
GMM-SVM and the i-vectors based systems.
REFERENCES
[1] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, “Speaker Verification Using Adapted Gaussian Mixture Models”,
Digit. Signal Process, vol. 10, no. 1, pp. 19–41, Jan. 2000.
[2] D. Reynolds, “Universal Background Models”, in Encyclopedia of Biometrics, S. Z. Li and A. K. Jain, Eds.
Springer US, 2015, pp. 1547–1550.
[3] T. Kinnunen and H. Li, “An overview of text-independent speaker recognition: From features to supervectors”,
Speech Commun., vol. 52, no. 1, pp. 12–40, Jan. 2010.
[4] N. Dehak and G. Chollet, “Support vector GMMs for speaker verification”, in 2006 IEEE Odyssey-The Speaker
and Language Recognition Workshop, 2006, pp. 1–4.
2
2.3
2.6
2.9
3.2
3.5
3.8
Amount of Enrollment speech data
GMM-UBM system
GMM-PBM system
EqualErrorRate(EER)
2
2.3
2.6
2.9
3.2
3.5
3.8
GMM-UBM system
GMM-PBM system
EqualErrorRate(EER)
0
0.05
0.1
0.15
0.2
0.25
GMM-UBM based system GMM-PUBM based system
CPUtime(inseconds)

IJECE ISSN: 2088-8708 
3663
[5] W. M. Campbell, D. E. Sturim, and D. A. Reynolds, “Support vector machines using GMM supervectors for
speaker verification”, IEEE Signal Process. Lett., vol. 13, no. 5, pp. 308–311, mai 2006.
[6] N. Dehak et al., “Support vector machines and Joint Factor Analysis for speaker verification”, in 2009 IEEE
International Conference on Acoustics, Speech and Signal Processing, 2009, pp. 4237–4240.
[7] N. Dehak, R. Dehak, P. Kenny, and P. Dumouchel, “Comparison between factor analysis and GMM support vector
machines for speaker verification”, in Odyssey, 2008, p. 9.
[8] P. Kenny, “Joint factor analysis of speaker and session variability: Theory and algorithms”, CRIM MontrealReport
CRIM-0608-13, 2005.
[9] P. Kenny, “A small footprint i-vector extractor”, in Odyssey, 2012, pp. 1–6.
[10] D. Wu, J. Cao, and H. Wang, “Speaker Recognition Based on i-vector and Improved Local Preserving Projection”,
Indones. J. Electr. Eng. Comput. Sci., vol. 12, no. 6, pp. 4299–4305, Jun. 2014.
[11] H. Beigi, Fundamentals of Speaker Recognition. Springer, 2011.
[12] D. A. Reynolds, “Comparison of background normalization methods for text-independent speaker verification”, in
Eurospeech, 1997.
[13] D. Reynolds, “Gaussian Mixture Models”, in Encyclopedia of Biometrics, S. Z. Li and A. K. Jain, Eds. Boston,
MA: Springer US, 2015, pp. 827–832.
[14] A. Rozi, D. Wang, Z. Zhang, and T. F. Zheng, “An open/free database and Benchmark for Uyghur speaker
recognition”, in Oriental COCOSDA held jointly with 2015 Conference on Asian Spoken Language Research and
Evaluation (O-COCOSDA/CASLRE), 2015 International Conference, 2015, pp. 81–85.
[15] B. Ayoub, K. Jamal, and Z. Arsalane, “An analysis and comparative evaluation of MFCC variants for speaker
identification over VoIP networks”, in 2015 World Congress on Information Technology and Computer
Applications Congress (WCITCA), 2015, pp. 1–6.
BIOGRAPHIES OF AUTHORS
Ayoub Bouziane received the M.Sc. degree in intelligent systems and networks from the Faculty
of Sciences and Technologies, Fez, Morocco, in 2012. He is currently pursuing the Ph.D. degree
in the intelligent systems and applications laboratory, Sidi Mohammed ben Abdullah University,
Morocco. His research interests cover signal processing and machine learning, mostly with
applications in automatic speaker recognition. In parallel with his research activities, he teaches
undergraduate level courses in computer science, at the Faculty of Sciences and Technologies,
Fez. Additionally, He is a member of IEEE Signal Processing Society, IEEE Computational
Intelligence Society, IEEE Computer Society and the International Speech and Communication
Association (ISCA).
Jamal Kharroubi has his B.Sc. in Computer Sciencefrom Sidi Mohamed Ben Abdellah
University (Fez-Morocco) in 1996. Two years after, he got his postgraduate degree in the
domain of Artificial Intelligence from Galilee’s Institute - Paris XIII University. In 2002,
Hereceived his Ph.D. degree in automatic speaker recognition systems from Telecom Paris Tech
(Ecole Nationale Supérieure des Télécommunications de Paris-France)”. Since January 2003, he
isan associate professor in the Department of Computer Scienceat the Faculty of Science and
Technology. In 2008, he received his HDR diploma (Habilitation to conduct research).
Additionally, He iscurrently the coordinator of the Master of Intelligent Systems and Networks.
Moreover, he is the author of more than thirty publications in peer-reviewed scientific journals &
conference proceedings. His research interests are focused onsignal andimage processing, pattern
recognition, etc.
Arsalane Zarghili is a Doctor of Sciencefrom Sidi Mohamed Ben Abdellah University (Fez-
Morocco). He received his Ph.D. in 2001 and joined the same University in 2002 as Professor at
thecomputer science department of the Faculty of Science and Technology (FST). In 2007 he
was head of the computer sciences department andchair of the Software Quality Master in the
FST-Fez. Helectures Programming, Distributed, compilation and Information processing, for
both under graduate and masterlevels. In 2008 he obtained his HDR in information processing.
In 2011, he is the co-founder and the head of the Laboratory of Intelligent Systems and
Applications in the FST of Fez. He is amember of the steering committee of the department of
computer sciences and was a member of the faculty board. He is also IEEE member since 2011.
His main research is about pattern recognition, image indexing and retrieval systems incultural
heritage, biometric, etc.

Towards an Optimal Speaker Modeling in Speaker Verification Systems using Personalized Background Models

More Related Content

Similar to Towards an Optimal Speaker Modeling in Speaker Verification Systems using Personalized Background Models (20)

More from IJECEIAES (20)

Recently uploaded (20)

Towards an Optimal Speaker Modeling in Speaker Verification Systems using Personalized Background Models