Voice Verification System for Mobile Devices based on
ALIZE/LIA_RAL
Hussein Sharafeddin
1
, Mageda Sharafeddin
2
and Haitham Akkary
2
1
Faculty of Sciences, Lebanese University, Hadat, Beirut, Lebanon
2
Department of Electrical and Computer Engineering, American University of Beirut, Beirut, Lebanon
Keywords: ALIZE, Biometric Authentication, LIA_RAL, Mobile Security, Speaker Verification, Text-independent
Identification, Voice Recognition.
Abstract: The main contribution of this paper is providing an architecture for mobile users to authenticate user
identity through short text phrases using robust open source voice recognition library ALIZE and speaker
recognition tool LIA_RAL. Our architecture consists of a server connected to a group of subscribed mobile
devices. The server is mainly needed for training the world model while user training and verification run
on the individual mobile devices. The server uses a number of public random speaker text independent
voice files to generate data, including the world model, used in training and calculating scores. The server
data are shipped with the initial install of our package and with every subsequent package update to all
subscribed mobile devices. For security purposes, training data consisting of raw voice and processed files
of each user reside on the user’s device only. Verification is based on a short text-independent as well as
text-dependent phrases, for ease of use and enhanced performance that gets processed and scored against the
user trained model. While we implemented our voice verification in Android, the system will perform as
efficiently in iOS. It will in fact be easier to implement since the base libraries are all written in C/C++. We
show that the verification success rate of our system is 82%. Our system provides a free robust alternative to
replace commercial voice identification and verification tools and extensible to implement more advanced
mathematical models available in ALIZE and shown to improve voice recognition.
1 INTRODUCTION
Security is a major barrier to using mobile devices in
business according to a study by Boa et al. (Bao et
al., 2011). A recent study by Communication Fraud
Control Association (Aronoff, 2013) shows that
global fraud loss in 2013 is up by 15% compared to
2011 and reports $8.84 Billion USD losses from
subscription fraud and identity theft alone.
There has been a lot of interest in the area of
biometric information for identification to leverage
security. This is mainly due to the fact that
authentication with biometric data is based on who
the person is and not on the passwords they are
supposed to remember. While a lot of work in the
literature examines multi modal biometrics where
several biometric measures are combined for
identification, in this work we focus on one
biometric parameter, human voice. In voice or
speaker identification the user does not claim an
identity, the tool rather determines the user’s identity
through a trained classifier. Speaker verification on
the other hand is when a tool determines whether the
user is who he/she claims to be. Speaker verification
and identification are generally two types, text
dependent and text independent. Current text
dependent voice verification tools are awkward to
use and perform poorly (Trewin et al. 2012). In
some applications the user is expected to remember
the phrase used and to pronounce it the way he/she
did during training. Hence, it is crucial to improve
robustness of voice verification and have the tool
automatically identify the user without introducing
additional restrictions or costs.
The main contributions of this work are:
1. Introducing the robust open source speaker
recognition platform ALIZE/LIA_RAL into
mobile development and research. This platform
is extensible and has shown potential for
improved success rate in Automatic Speaker
Recognition (ASR) (Larcher et al. 2013).
2. Our code modifications are contributed to the
248
Sharafeddin H., Sharafeddin M. and Akkary H..
Voice Verification System for Mobile Devices based on ALIZE/LIA_RAL.
DOI: 10.5220/0005218902480255
In Proceedings of the International Conference on Pattern Recognition Applications and Methods (ICPRAM-2015), pages 248-255
ISBN: 978-989-758-077-2
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
open source community to facilitate
enhancements to state of the art voice
identification on mobile devices (Sharafeddin et
al. 2014).
One voice identification Android application is
based on open source code reported in the literature
by Chen et al. (Chen et al. 2013). They use Modular
Audio Recognition Framework (MARF), an open
source Java tool. Linear Prediction Coefficients
(LPC) model is used in MARF. LPC is based on
linear combinations of past values of the voice
signal with scaled present values. LPCs greatly
simplify the voice recognition problem by ignoring
the complete signal modeling of human vocal
excitation (Campbel 1997). The Android application
in (Chen et al. 2013) reduces errors in voice
detection by generating distance information and
optimizing results using personal thresholds. Their
text-dependent speaker identification tool achieves
an 81% success rate. Using thresholds is common in
voice verification to adjust two error types: False
Acceptance Rate (FAR) when an imposter is
authenticated and False Rejection Rate (FRR) when
an authentic user is denied access (Faundez-Zanuy
2006).
Another ASR tool for mobile phones that has
been recently introduced is SpeakerSense (Lu et al.
2011). The tool focuses on speaker identification,
hence it recognizes who is speaking out of a group
of many trained speaker voices. As such their work
is an class pattern recognition problem, and
since it deals with audio data, it requires the training
time to be not less than 3 minutes per trainee. This
necessitates studying the power consumption
efficiency as a known restriction on small devices.
SpeakerSense utilizes similar pre-processing and
feature extraction approaches as our tool; however,
since the tools differ in the ultimate goals of
identification versus verification, the classification
models differ. SpeakerSense does utilize a Gaussian
Mixture Model (GMM) using only 20 Mel-
frequency Cepstral Coefficients (MFCCs) while our
tool uses a GMM with a 60-dimensional component
vector per user. The extended vector we use includes
19 MFCCs, 40 first and second order derivatives and
1 energy component. Additionally, while the GMM
training is done on a remote server running Matlab
for SpeakerSense, we run the C++ based user
training with the option to perform the Universal
Background Model (UBM) training natively
onboard the mobile device. The reason why we
leave the UBM as an option is that usually we can
have one UBM, pre-trained using randomly selected
speech segments, work with any individual user
GMM. Hence depending on the particular
application and deployment policy, we can have the
UBM data incorporated within the installation
package transparently to the user. This not only
simplifies but also lends flexibility to our
architecture.
Using a cloud based system for speech
recognition using a mobile phone was introduced by
Alumae et al. (Alumae and Kaljurand 2012). The
system uses open source CMU Sphinx system
(CMUSphinx 2014) and targets Estonian language
recognition where the authors try to provide a user
experience comparable to the Google voice search.
Unlike our system, their system sends raw voice data
from the mobile device to the server for recognition.
This is fundamentally different from the role of
servers in our system which is mere calculation of
the UBM model and processing public speaker data.
In this work we use the Mel-Wrapped Cepstrum
analysis, which has been shown to work well in
speaker recognition (Gish and Schmidt 1994),
compared to less efficient techniques such as LPC
clustering distances and neural networks.
The rest of this paper is organized as follows. In
section 2 we provide background information on
main algorithms used in our system. In section 3 we
describe our system and in section 4 we discuss our
experimental setup. We show performance results
and success rates in section 5. We finally conclude
in section 6 and discuss future work in section 7.
2 BACKGROUND
Two main components are relevant in speaker
verification, features that discriminate among human
voices and suitable classification approaches.
Among the most common models that examine both
components are LPCs and MFCCs. Ideally we were
looking for open source packages that implement the
full chain of voice pre-processing, feature extraction,
classification model creation, user training and
classification. Our job would be to evaluate such
tools on short speech segments from a limited set of
users so that we can port it to a portable device
running Android for example.
For speaker verification or identification two
open research tools appear to have received research
attention in the past decade. The first is MARF and
the second is the combination of ALIZE/LIA_RAL
packages (D'Avignon Laboratoire Informatique
2011). MARF is an open source (Mokhov et al.
2006) research tool mainly for speaker identification
written in Java. Among several functions, it extracts
VoiceVerificationSystemforMobileDevicesbasedonALIZE/LIA_RAL
249
LPCs and implements several pattern recognition
techniques for classification which can be used for
speaker identification. Being in Java, and thus
platform independent, it is easier to deploy and
enhance than other similar tools. We tested MARF
on our data set using the LPC features. The best
results among the different classifiers it provides did
not exceed 80% for speaker identification.
ALIZE (D'Avignon Laboratoire Informatique
2011) is a core toolkit that encapsulates several
base-level functionalities for managing speech files
and the subsequent data extracted thereof. Utilizing
ALIZE functionalities is LIA_RAL, a set of modules
that implement different speech processing and
recognition algorithms which can be cascaded into a
customized tool chain for testing different
mathematical models. Both ALIZE and LIA_RAL,
developed at the Laboratoire d'Informatique
d'Avignon (LIA) –France, are open source platforms
for research in speaker verification and speech
recognition. Several methods have been investigated
for the classification model including GMM, Joint
Factor Analysis (JFA), and Support Vector
Machines (SVM). In this work we focus on the
GMM implementation leaving the others for future
work. It is worth mentioning here that LIA_RAL
uses speech features extracted by other tools. One
such tool is the open source toolkit for Signal
Processing (SPro) (Spro 2004), a speech processing
and feature extraction library that implements
commonly used pre-processing steps and feature
extraction algorithms. In this work we utilize it for
pre-emphasis, windowing and MFCC extraction.
These are extracted directly from the raw voice files
and stored into corresponding binary files.
Reynolds et al. (Reynolds and Rose 1995)
detailed the GMM model and justified its veracity
for speaker identification. Since the main interest
here is to validate a speaker by processing a short
speech segment, the model is desired to utilize
features that effectively represent the speaker’s
vocal tract physiological and acoustic properties.
Several feature types have been studied in the
literature including LPC that model the vocal tract as
a linear system. Although those features offer good
representation of a person’s vocal tract properties,
they are sensitive to additive noise such as
microphone background.
Spectral analysis based features offer frequency
range selectivity that might help in applications
where noise is inevitable (Reynolds and Rose 1995).
In particular, MFCCs have been widely used in
speech processing and recognition applications. The
extraction process can be summarized by the
following steps: voice signal pre-processing which
includes pre-emphasis and windowing; Fourier
transform and coefficient modulus calculation;
filtering through a Mel-frequency filterbank for
smoothing and envelope extraction and therefore
vector size reduction.
1000log
1
1000
(1)
Finally, performing a discrete cosine transform
(DCT) on the log scaled values (Rabiner and Schafer
2010) (Petrovska-Delacrétaz, et al. 2009):
cos
1
2
;
1,2,…,,
(2)
where
1,…, is the log-absolute value of
the
Fast Fourier Transform (FFT) coefficient;
is the number of FFT coefficients;
is the
number of cepstral coefficients
to keep. Each
window will then be represented by its own cepstral
vectors. The logarithm calculation in the final stage
makes the model closer to the human hearing
system. The DCT decorrelates the log filterbank
energies because of their overlap. This simplifies
downstream model calculations by using diagonal
covariance matrices.
The MFCC features calculated using Spro (Spro
2004) are further normalized to have a zero-mean
and unit variances by the feature normalization step
shown in figure 1a-c for noise reduction. Energy
extraction (Spro 2004) also shown in figure 1a-c is
applied next to the voice frames in order to minimize
error contribution from unvoiced speech, silence or
background noise; their corresponding feature
vectors, linked by the frame time labels, are then
easily eliminated.
The remaining features are well suited for use in
speaker verification, and hence, make up the feature
vector components of the GMM model that we also
use in this work. Moreover, the smoothing property
of the Gaussian model yields robustness to the
stochastic parameter estimation of a speaker’s voice
underlying component distributions. The GMM
represents a multi-modal distribution that provides a
better and smoother fit when compared with the
simpler uni-modal Gaussian or the vector
quantization codebook model.
For a single speaker -dimensional feature
vector by , a weighted sum of Gaussian densities
;
,Σ
with weights
constitutes a GMM for
to be utilized for the likelihood ratio test:
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
250
|
≜
|
,
,Σ
;
1,…,and
1;
(3)
1
2
|
Σ
|
(4)
The current approach in the field for estimating the
individual speaker GMMs is to first estimate a
UBM, also called GMM world model, from random
features taken from numerous speakers. In our work
we demonstrated this approach by recording several
short speech segments (10 seconds) from 7
randomly selected speakers. We subsequently pass
their features to the training step shown in figure 1a.
This UBM represents speaker-independent feature
probability density function (pdf) mixture and is
used in adapting the individual speaker GMM
parameters to better represent the acoustic classes
present in the speech segments.
Verifying whether a given speech vector would
belong to speaker is done by a hypothesis test. The
pdf for the null hypothesis
is
|
; the pdf
for the alternative hypothesis
is
|
where
represents all models not
(
. In fact
this is the dominant approach in the literature as
done by LIA_RAL. The likelihood ratio test
becomes:
|
|
,authenticateas
,rejectasimpostor
(5)
The decision threshold is set empirically. In our
implementation, we introduced a tool sensitivity
selector to be set by the user. Fauve et al. (Fauve et
al. 2007) studied limitations of the basic GMM
model and of a modified model that uses support
vector machines; they also demonstrated
improvements on a modified eigenvoice model.
Bousquet et al. (Bousquet et al. 2011) proposed
an improvement to the i-vector approach by
performing simple linear and nonlinear
transformations to remove the session effects. They
also proposed an improvement on the scoring
technique by utilizing the Mahalanobis distance in
the classifier. In this work we do not include Fauve
et al.’s (Fauve et al. 2007) and Bousquet et al.’s
(Bousquet et al. 2011) enhancement and leave this
for future work.
3 PROPOSED SYSTEM
In this work we advocate using high quality open
source speech recognition techniques to build a
speaker verification application that can be used
stand alone or integrated with other authentication
methods. Most of the available code suitable for this
task is written in C and C++. While this can be
integrated into iOS applications easily, it is not
straight forward to implement in Android. The Java
Native Interface (JNI) environment is needed for
building dynamic libraries to be loaded at run-time.
Specialized Makefiles and coding rules for function
names and parameters need to be followed in order
for the library to properly interface with the Java
Dalvik Virtual Machine inside Android. This multi-
layered hybrid system makes it challenging to debug
the application.
3.1 System Overview
Figure 1 shows a general overview of our proposed
system. The figure shows the three separate steps in
our system. Figure 1a shows the server training step,
a process that is performed offline and involves
public speaker voice signals or system speaker voice
signals who allow the use of these signals. Spro
(Spro 2004) is an open source ANSI C library with
various speech analysis techniques. In our work we
use Spro version 4 which is released under the GPL
license. The default settings released with the
ALIZE/LIA_RAL code assume the audio file format
is SPHERE (.sph) which is an older format used
initially by the National Institute of Standards and
Technology (NIST) (NIST 2014). In this work we
implement an application running on current mobile
devices and hence we use the WAVE format (.wav)
which is also supported by the Spro file reader. Our
main use for Spro is to perform the first
preprocessing steps which are pre-emphasis and
windowing and to extract the MFCC features. The
main problem we faced with Spro was a portability
to Android. Some memory management system calls
ended up using uninitialized memory blocks and
causing core dumps. We wrote macros to
circumvent the problem in a way not to violate the
license. We contacted the author of Spro with this
fix which will be added to the new release. The
system was developed on an Intel machine running
Ubuntu linux using eclipse IDE; the native code was
cross-compiled to ARM-based dlls using JNI.
VoiceVerificationSystemforMobileDevicesbasedonALIZE/LIA_RAL
251
Figure 1: Our System's Main Activities- (a) server
training, (b) user training, and (c) user identification.
3.2 Server Training
The voice signals of the random users undergo four
Spro steps, MFCC extraction and normalization,
energy extraction and normalization shown in figure
1a. This will generate PRM files which include
normalized feature vectors and LBL files denoting
durations of speech periods in a given speaker voice
file. The last two steps involve server training the
input WAV files to create the UBM model. All of
these files are needed in the mobile device user
training and verification and are hence packaged and
shipped with the application to the various users.
Periodically, the server can resend updated data in
the form of application upgrade to all subscribed
users. Note that the design of server training
carefully separates user data from public data for
two main reasons, first to remove any security
concerns by not sending user voice files over the
network and second to reduce the mobile user
training time.
3.3 Mobile User Training
Figure 1b shows the mobile user training steps. Note
that the same first four Spro steps in server training
are applied here. Various PRM files are generated as
a result. Additionally, the user is asked to train
his/her device. This is a timed training session which
lasts 8s to 30s. During this time, the user is supposed
to utter a sentence of their choice. The generated
user PRM files with the LBL files denoting periods
of utterance, as opposed to silence, are saved in the
package file space to avoid exposing the files when
the device is compromised temporarily. The
imposter will need to have the root id to be able to
extract these files.
3.4 Mobile User Verification
Finally, figure 1c shows the user verification step. In
this step, the user is asked to utter sentences of
his/her choice. This is also a timed session that lasts
8s to 15s. After feature and energy extraction and
normalization, the new data is compared against the
world UBM plus owner model to create verification
scores. The scores are finally normalized. Score
normalization is based on the following three
normalization techniques:
1. Zero normalization technique to compensate
for inter-speaker score variations.
2. Test normalization technique to compensate
for inter-session score variation.
3. Confused Zero and Text normalization to do
both and can perform better that 1 & 2
(Srikanth and Hegde 2010).
We found that using scores from either 1 or 2 are
sufficient in addition to scores from 3. The
verification rates reported in section 5 are based on
scores from 1 & 3. Normalization techniques allow
setting a global threshold independent of speakers
using our system.
Train &
Feature
Files
To
System
Users
Random
Users
WAVE
Files
(a) Server Training Steps
MFCC
Extrac
tion
MFCC
Normalizati
on
Energy
Extracti
on
Energy
Normaliz
ation
Train
With
Input
WAVE
Files
Train
With
Input
WAVE
Files
Training Success
Acknowled
g
ement
Mobile
User
WAVE
Files
(b) Mobile User Training Steps
Feature
Extract
ion
Feature
Normalization
Energy
Extraction
Energy
Normalization
Train
with
User
Decision
Mobile User
Text Inde
p
endent WAVE File
(c) Mobile User
Feature
Extraction
Feature
Normalization
Energy
Extraction
Energy Normalization
Pattern Matching
and Score
Normalization
Pattern Matching
and Score
Normalization
Compute Gaussian
Mixture Model
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
252
4 EXPERIMENTAL SETUP
We carried several build tests and application
experiments in order to achieve the following goals:
1. Concept: A proof of concept assurance that all the
required code gets ported and works as expected.
This was achieved through a set of C/C++ macro
definitions and some recoding needed for memory
stability and exception handling. We ran several
tests on a 64-bit Intel-based linux machine as well as
on an ARM-based handheld Android device using
the same NIST04 data and configuration sets
released with the ALIZE_GMM (D'Avignon
Laboratoire Informatique 2011) tutorial which
utilizes the baseline GMM MAP approach. Figure 2
below demonstrates that the final normalization
steps on either machine were verified to be very
similar, the minor differences being due to floating
point precision. 2. Threshold: Since making the
authentication decision requires that the
normalization scores be thresholded, finding a global
working threshold is a known challenge in the field
especially that this depends on training and on
testing conditions. Being a first study in
implementing the ALIZE/LIA_RAL system on
mobile devices, the main focus here is not
methodical especially that there are further
published improvements to the basic GMM MAP
model that we did not include yet. That said, we
automated anonymous voice, score, and
performance data collection on the device for further
analysis. 3. Data source variance: Since this
application is intended for everyday mobile device
users, practical considerations in terms of noise,
device variability and text similarity should be
reflected in fresh recorded speech data. Device
variability in terms of performance is due to the long
chain of processing and calculations involved and
microphone operating characteristics.
We conducted simultaneous tests on two
different ARM-based Android devices: a single core
LG Optimus L5 phone running Android 4.0.3 on an
ARM Cortex-A5 800MHz processor and a quad-
core Samsung SIII phone running Android 4.3 on an
ARM Cortex-A9 1.4GHz processor. The audio
recorders were configured at 8 KHz sampling rate
with16 bits per sample using a single audio channel.
Eight volunteers enrolled their voices for UBM
training with short duration speech data (8-15
seconds) recorded. Another set of eight volunteers
enrolled as targets in a quiet environment for 8
seconds in one experiment and for 30 seconds in
another. Each target then performed three true
positive tests against their own trained model, and
similarly three true negative tests against other
targets. The processing times and normalized scores
were recorded for each test. All true positive tests
were repeated using the same trained text as well as
using different text with variable recording times.
The test sets were conducted first in a quiet setting
then repeated with a noisy background.
5 RESULTS
Goal 1: Initial tests on public NIST04 data proved
the concept and produced very close results with
minor differences due to floating point precision.
Using our recorded data also confirmed feasibility;
however, the results were suboptimal due to the
limited number of UBM training and recording
durations. Figure 3 shows the overall Equal Error
Rate (EER) achieved with this limited data set.
Goal 2: Since our tests were repeated in different
conditions, we separated the results first based on
background noise then based on text similarity.
Since our goal is text-independence, no discrete time
warping (DTW, done for voice segment alignment)
was implemented; yet and as expected, the text-
similar tests yielded 24% better results than random
speech test cases. The superiority in the former is
mainly attributed to more energy similarities in the
same frequency bands uttered. Noisy backgrounds
did contribute to degraded performance compared to
quiet test setting by 21% EER increase. The 30s
enrollment duration also yielded better separation
distance as opposed to the 8s enrollment.
Figure 2: Sample result on mobile phone using NIST04
sample confirming previous results.
Goal 3: Our tests did not show significant impact
of microphone variability between the two devices;
however, performance was noticeably faster on the
False Negative Rate (%)
DET Curve for sample from NIST04 data EER = 2.5%
False positive rate %
VoiceVerificationSystemforMobileDevicesbasedonALIZE/LIA_RAL
253
multicore device, as expected. Data processing itself
was just 35% faster since our application runs on a
single thread on either core, the improvement being
mostly due to the clock speed; however, the overall
application user experience was much better due to
different threads handling context switching on the
multicore device.
Figure 3: Sample result from our limited data set recorded
on mobile device.
6 CONCLUSIONS
In this work we introduce a system for voice
verification consisting of a server and subscribed
mobile users. We use open source code SPro and
ALIZE library as well as LIA_RAL speaker tool for
user training and verification. Our system
preliminary results show an 82% identification
success rate.
The full android application with all
dependencies is available online (Sharafeddin et al.
2014). This includes a README file with all
modifications made to both Spro and LIARAL. It
also has a sample UBM model and user wave files.
The various modules used in our system, shown in
figure 1, are documented in the respective tools.
7 FUTURE WORK
Results can be significantly improved first by
collecting enrollment segments from many users
anonymously using different devices in regular daily
life settings (noisy or quiet) in order to build a more
reliable UBM to be distributed with the application.
Moreover, the enrollment time would be set to a
minimum of 30s. Further improvements to the
baseline GMM MAP system including Joint Factor
Analysis (Kenney et al. 2007), i-vectors (Dehak et
al. 2009) and SVM (Campbell et al. 2006) based
classification as included in ALIZE_3.0 (Larcher et
al. 2013) will be incorporated and compared. We
intend to evaluate the work of Chan et al. (Chan et
al. 2007) where wavelets for features are compared
with the MFCC features on a GMM model. Finally,
we would also like to understand if implementing
dynamic time warping to create distance information
for determining personal thresholds as reported by
(Chen et al. 2013) improves verification rates.
Finally, previous baseline GMM studies showed
suboptimal performance of the GMM MAP model
especially with short duration training and testing
(Fauve et al. 2008). We intend to use this work as a
start for evaluating ALIZE/LIA_RAL on different
mobile phones. Our application works in production
mode as well as in research mode where it keeps
speech data and collects score and performance
information. This will be leveraged as new
volunteers enroll to further enhance the UBM and
the methods tested.
REFERENCES
Alumae, T., Kaljurand, K., 2012. Open and extendable
speech recognition application architecture for mobile
environments. In SLTU'12, The 2nd International
Workshop on Spoken Language Technologies for
Under-resourced Languages.
Aronoff, R., 2013. Global fraud loss survey.
Communications Fraud Conrol Association. Roseland,
NJ.
Bao, P., Pierce, J., Wittkaer, S. , Zhai, S., 2011. Smart
Phone Use by Non-Mobile Business Users. In
MobileHCI.
Bousquet, P. M., Matrouf,, D., Bonastre, J. F., 2011.
Intersession Compensation and Scoring Methods in
the i-vectors Space for Speaker Recognition. In 12th
Annual Conference of the International Speech
Communication Association.
Campbel, J. Jr., 1997. Speaker recognition: a tutorial. In
Proceedings of the IEEE. Vol. 85, no. 9, Sept., pp.
1437-1462.
Campbell, W. M., J. P. Campbell, J. P., Reynolds, D. A.,
Singer, E., Torres-Carrasquillo, P. A., 2006. Support
Vector Machines for Speaker and Language
Recognition. In Computer Speech & Language,
Elsevier, Vol. 20.
Chan, W. N., Zheng, N., Lee, T., 2007. Discrimination
Power of Vocal Source and Vocal Tract Related
Features for Speaker Segmentation. In IEEE
TRANSACTIONS ON AUDIO, SPEECH, AND
LANGUAGE PROCESSING.
False positive rate %
False Negative Rate (%)
DET Curve for a sample from our limited dataset EER = 18%
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
254
Chen, Y., Heimark, E., Gligorski, D., 2013. Personal
threshold in a small scale text-dependent speaker
recognition. In International Symposium on
Biometrics and Security Technologies.
CMUSphinx, 2014. Carnegie Melon University.
http://cmusphinx.sourceforge.net.
D'Avignon, Laboratoire Informatique, 2011.
http://mistral.univ-avignon.fr/index_en.html.
Dehak, N., Dehak, R., Kenny, P., Brummer, N., Ouellet,
P., Dumouchel, P., 2009. Support Vector Machines
versus Fast Scoring in the Low-Dimensional Total
Variability Space for Speaker Verification. In
InterSpeech 10th Annual Conference of the
International Speech Communication Association.
Faundez-Zanuy, M., 2006. Biometric security technology.
In IEEE Aerospace and Electronic Systems Magazine
No. 21, pp. 15-26.
Fauve, B., Evans, N., Mason, J., 2008. Improving the
performance of text-independent short duration.
Odyssey.
Fauve, B.G.B., Matrouf, D., Scheffer, N., Bonastre, J. F.,
Masin, J. S. D., 2007. State-of-the-Art Performance in
Text-Independent Speaker Verification Through
Open-Source Software. In IEEE Transactions on
Audio Speech and Language Processing.
Gish, H., Schmidt, M., 1994. Text-independent speaker
identification. In IEEE Signal Processing Magazine,
Vol. 1, pp. 18-32.
Kenney, P., Boulianne, G., Ouellet, P., Dumouchel, P.,
2007. Joint Factor Analysis versus Eigenchannels in
Speaker Recognition. In IEEE Transactions on Audio,
Speech, and Language Processing, Vol. 15.
Larcher, A., et al., 2013. ALIZE 3.0 - Open source toolkit
for state-of-the-art speaker recognition. In Interspeech.
Lu, H., Brush, A., Priyantha, B., Karlson, A., Liu, J., 2011.
SpeakerSense: Energy Efficient Unobtrusive Speaker
Identification on Mobile Phones. In The Ninth
International Conference on Pervasive Computing.
Mokhov, S., Clement, I., Sinclair, S., Nicolacopoulos, D.,
2002. Modular Audio Recognition Framework.
Department of Computer Science and Software
Engineering, Concordia University.
http://marf.sourceforge.net.
NIST, 2014, www.nist.gov.
Petrovska-Delacrétaz, D., Chollet, G., Dorizzi, B., Jain,
A., 2009. Guide to Biometric Reference Systems and
Performance Evaluation. Springer.
Rabiner, L., Schafer, R., 2010. Theory and Applications of
Digital Speech Processing 1st (first) Edition . Prentice
Hall.
Reynolds, D. A., Rose, R., 1995. Robust Text-Independent
Speaker Identification Using Gaussian Mixture
Speaker Models. In IEEE Transactions on Speech and
Audio Processing, Vol. 3, No. 1.
Sharafeddin, M., Sharafeddin, H, Akkary, H., 2014.
Android-Voice-IDentification-App-using-SPRO-
ALIZE-LIARAL. github.com/umbatoul/Android-
Voice-IDentification-App-using-SPRO-ALIZE-
LIARAL.
Srikanth, N, Hegde, R. M., 2010. On line client-wise
cohort set selection for speaker verification using
iterative normalization of confusion matrices. In
EUSIPCO European Signal Processing Conference,
pp. 576-580.
Spro, 2004. http://www.irisa.fr/metiss/guig/spro/.
Trewin, S., Swart, C., Koved, L., Matino, J., Singh, K.,
Ben-Davic, J., 2012. Biometric Authentication on a
Mobile Device: A Study of User Effort, Error, and
Task Disruption. In ACSAC'12 Proceedings of the
28th Annual Computer Security Applications
Conference pp. 159-168.
VoiceVerificationSystemforMobileDevicesbasedonALIZE/LIA_RAL
255
