Mistral : open source biometric platform.
[AB-113]

Eric Charton
Laboratoire Informatique
d’Avignon
339, chemin des Meinajaries
Avignon, France
eric.charton@univ-
avignon.fr
Anthony Larcher
Laboratoire Informatique
d’Avignon
339, chemin des Meinajaries
Avignon, France
anthony.larcher@univ-
avignon.fr
Christophe Levy
Laboratoire Informatique
d’Avignon
339, chemin des Meinajaries
Avignon, France
christophe.levy@univ-
avignon.fr
Jean-Francois Bonastre
y
Laboratoire Informatique
d’Avignon
339, chemin des Meinajaries
Avignon, France
jean-
francois.bonastre@univ-
avignon.fr
ABSTRACT
Mistral is an open source software for biometrics applica-
tions. This software, based on the well-known UBM/GMM
approach includes also the latest speaker recognition devel-
opments such as latent factor analysis, unsupervised adap-
tation or SVM supervectors. The software performance is
highlighted in the framework of the NIST evaluation cam-
paigns.
Categories and Subject Descriptors
H.4 [Information Systems Applications]: Miscellaneous;
D.2.8 [Software Engineering]: Metrics|complexity mea-
sures, performance measures
Keywords
GMM, Biometric, Speaker recognition, Face recognition
1. INTRODUCTION
As the scientic competition is becoming intense in the
biometric eld, an isolated team of researcher would have
diculties to elaborate a biometric tool at a state-of-the-art
Mistral Project is funded by the French National Research
Agency (ANR) - Programme Technologies Logicielles 2006
yOcial coordinator of the Mistral Consortium.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SAC’10 March 22-26, 2010, Sierre, Switzerland.
Copyright 2010 ACM 978-1-60558-638-0/10/03 ...$10.00.
level. The Mistral project issues from this observation. Mis-
tral is an high quality open-source, free biometric platform
(distributed under LGPL license), supported by an ecient
and active scientic community. The main originality of the
Mistral project is to use a unique statistical engine for vari-
ous modalities of biometric applications.
2. ARCHITECTURE OF MISTRAL
A complete biometric application often includes a sta-
tistical engine (e.g. GMM models) and a wide variety of
pre/post processing tools like those dedicated to data prepa-
ration and score calculation. The Mistral platform oers a
complete biometric tool performing at the state-of-the-art
level in highly competitive tasks such as international eval-
uation campaigns like the NIST-SRE [2].
Work environment
Mistral is a fully congurable scientic tool which can be
used through a graphic user interface oering ergonomic fa-
cilities. In the perspective of Mistral use by student or novice
users, a Java interface is also available. First visual interface,
called Mistral Cong, allows beginners to use a set of pre-
dened parameters. This aims to decrease the risk of bad
settings and to give a textual explanation of each parame-
ter. This Java interface allows also automatic generation of
experiments sets (scripts, folder conguration).
3. INTERNET PORTAL AND COMMUNITY
All Mistral applications and tools can be accessed through
an Internet portal1. It includes also a set of engineering soft-
ware applications to help developers community (UML tools
for modelization, Doxygen for documentation, SVN for dis-
semination, . . . ). All those elements federate a community
of users from numerous scientic laboratories, faculties and
1http://mistral.univ-avignon.fr

companies around the world (mostly from Europe, North
America, China).
4. MISTRAL FUNCTIONALITIES
Objective of Mistral is to give high level libraries (Alize
and Mistral tools) and ready-to-use modules, for any user
wishing to manage a task of speaker identication and veri-
cation, speaker diarization, face and language recognition.
The architecture includes a set of tools, based on two soft-
ware libraries: the Alize library dedicated to feature and
model management, and the Mistral tools library providing
useful functions mainly for post-processing.
Statistical models of Mistral
For basic applications, Mistral proposes a complete GMM
statistical engine (Gaussian Mixtures) fully congurable by
user. The models can then be completed by the two main
sets of functionalities included in libraries:
The Alize library allows data reading and synchroniza-
tion, algorithms (EM/ML/MAP) for models training,
score computation, discriminant classiers, session vari-
ability modeling.
The Mistral tools provides ready to use functions for SVM
classication, Latent Factor Analysis, features/scores
normalization.
Alize library provides statistical engine, and feature server
management of Mistral. The global architecture of Alize di-
vides the various tasks of the biometric modelization and
classication processes: [The feature server] manages access
to acoustic or raw features. [The model server] manages rep-
resentations of mixtures and their parameters. [The statis-
tical server] is in charge of calculations for models training,
likelihood estimation, score normalization.
4.1 Mistral applications
Mistral provides ready to use applications to create a com-
plete biometric experience, based on Alize and Mistral tools
libraries.
Speaker verification - SpkDet
Automatic speaker recognition experiments are achieved with
the SpkDet tool. This complete application allows user to
manage a complete experiment of speaker recognition, from
acoustic features preparation to modelization and score cal-
culation.
All those tools give and easy access to conguration op-
tions, allowing to select available options from the Alize and
Mistral tools libraries, including SVM conguration, Factor
Analysis or NAP.
SpkDet includes some complementary tools like energy
Detector (speech and non-speech detection in signal), GMM
visualization, histogram generation, and score fusion.
Segmentation and classification of speakers:
Segmentation application, called hSeg, allows segmentation
and hierarchical classication of speakers from a speech sig-
nal. Signal segments are low enough to allow classication
even in dicult conditions, like multiple speakers interven-
tion. The available methods include plain or diagonal Gaus-
sian with various evaluation metrics.
Langage detection - LangDet:
The LangDet application is a set of tools derived from Sp-
kDet. Those tools allow to build a complete language detec-
tion system, fully compatible with the NIST-LRE evaluation
campaign datas.
Facial biometrics:
Several modules of Mistral are dedicated to facial biometrics.
Local features approaches have been shown to be more ro-
bust than holistic ones in terms of pose variations or transla-
tions due to face localization. The use of such parameters in
the GMM/UBM paradigm has provided good performance
[3]. Local Principal Component Analysis (LPCA), eigen-
faces and 2D DCT extraction tools as well as some normal-
ization tools are part of the Mistral platform and could be
easily combined to the statistic tools.
5. MISTRAL PERFORMANCES
Mistral performances are evaluated on regular basis through-
out national and international scientic evaluation campaigns.
During the last 3 years, Mistral has been deployed by Mis-
tral Consortium members in NIST-SRE 2006-08 (speaker
detection), ESTER 2 (speaker segmentation).
Mistral is regularly evaluated with participation to several
national/international campaign like NIST-SRE. During the
NIST-SRE 2006 campaign, Mistral achieved an EER of 5%,
what placed this system close to the state-of-the-art systems
(cf. [1]). During NIST6SRE 2008 evaluations, Mistral was
at the state-of-the-art level (cf. [2]).
Mistral had been deployed in the ESTER 2 campaign:
speaker diarization task. The segmentation module of Mis-
tral obtained a 13 % error rate on the test corpus. cf. [4])
6. CONCLUSIONS
The aim of Mistral project is to develop a state-of-the-art
biometric system. The performance of Mistral platform is
highlighted during several national and international cam-
paigns.
The major originality of Mistral is to propose a single
recognition engine for multiple modalities: primarily voice
and face. Mistral is also available on dierent operating
systems (Linux / Windows / Mac OS). For users of Mistral,
adoption of a unique and free recognition engine allows to
focus on research and development eorts on specic aspects
of the studied modality, rather than on the development and
monitoring of complex low level software engines.
7. REFERENCES
[1] J.-F. Bonastre, N. Scheer, D. Matrouf, C. Fredouille,
A. Larcher, A. P. G. Pouchoulin, N. Evans, B. Fauve,
and J. S. Mason. Alize/spkdet: a state-of-the-art open
source software for speaker recognition. 2008.
[2] D. Matrouf, J.-F. Bonastre, C. Fredouille, A. Larcher,
S. Mezaache, M. McLaren, and F. Huenupan. LIA
GMM-SVM system description: NIST SRE. In NIST
SRE, Montreal (Canada), april 2008.
[3] C. McCool and S. Marcel. Parts-Based Face
Verication Using Local Frequency Bands. IEEE IAPR
International Conference on Biometrics (ICB),
page 72, 2009.
[4] NIST. Dn4.2: Report on benchmark-based evaluations,
2008. Muscle, European Project no. FP6-507752.