Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

Discussion of Segmentation

Methods

The Daugman Algorithm

The algorithm developed by Daugman [2–5] is used in

the current public deployments of iris recognition sys-

tems. Indeed, two of the iris image preprocessing steps

introduced in [2], an integro-differential operator for

edge detection and the pseudo-polar coordinate trans-

form which transforms (or ‘‘unwraps’’) the iris ring

into a doubly-dimensionless coordinate system, have

been incorporated into various other proposed recog-

nition methods. Therefore it is natural to begin with

the off-axis segmentation in [5].

The segmentation begins with approximating the

pupil, limbic, and eyelid boundaries using the integro-

differential operator from [2]:

max

r;x

ðÞ

ðrÞ

r;x

Iðx; yÞ

2pr



: ð1Þ

Here I(x, y) are the image grayscale values, G

(r)is

a smoothing function such as a Gaussian of scale s,

and the contour integral is along circles given by

center (x

) and radius r (but the circles are replaced

by arcs when searching for eyelids). Thus this opera-

tor produces initial circular estimates for the pupil

and limbic boundaries from the maximum blurred

partial derivative of the image grayscale values with

respect to a radial variable. To zero in on the actual

boundary, the plots of radial gradients for the pupil

and limbic boundaries form two ‘‘snakes’’, like those

shown in Fig. 2, each of w hich is then approxi-

mated by a discrete Fourier series. The iris ring is thus

bounded by smooth closed curves which would project

behind occluding eyelids, but in general the inner

and outer boundary will be neither exactly circular

nor elliptical.

The pupil boundary contour provides the informa-

tion necessary to compute the gaze angle from ‘‘Four-

ier-based trigonometry’’: using the Fourier coefﬁcients

to ﬁt an oriented ellipse to the pupil boundary. An

ellipse is parameterized by the equations x(t) ¼ Acos(t),

y(t) ¼ B sin(t), where A < B in the case of a vertical

major axis. If the gaze is deviated upward or down-

ward, the pupil boundary would appear rotated by

some ang le y, so the more general parameterization is

an oriented ellipse, given by

XðtÞ¼½A cos

y þ B sin

ycosðtÞ

þ½ðB  AÞcos y sin ysinðtÞ;

Y ðtÞ¼½ðB  AÞcos y sin ycosðtÞ

þ½B cos

y þ A sin

ysinðtÞ:

By comparison of these expressions with the lower

frequency terms of the Fourier series which describes

the pupil boundary contour, the values of A, B, and y

are uniquely determined. This is sufﬁcient to deter-

mine sideward and upward/downward gaze directions,

and to rotate the image back to orthographic view with

an afﬁne transformation. An example image before

and after rotation is shown in Fig. 3.

The last step in the segment ation process is the

detection of eyelashes, which are found by the realiza-

tion that the presence of eyelashes overlying the iris

results in too many dark pixels in the upper half of

the iris, compared with the grayscale distribution

in the lower half of the iris. In such cases, when the

grayscale distributio n in the top half of the iris

shows multimodal mixing, the suspect pixels are elimi-

nated by thresholding. Figure 2 shows the iris pixel

histograms for eyelash detection, the segmentation

Segmentation of Off-Axis Iris Images. Figure 2

Reprinted from [5], Fig. 4, pp. 1171. Iris image (ICE-1 file

239766) overlaid with pupil, iris-sclera, and eyelid boundaries,

and detected eyelashes. Also shows iris pixel histograms

(lower right), snakes (the two lower left plots), and resulting

Iris Code (top left). In the top snake plot the two fuzzy bands

are the curvature map for the iris-sclera boundary, and the

dotted curve shows its Fourier series approximation are for

1160

Segmentation of Off-Axis Iris Images

result, and the iris template after feature extraction

(i.e., the IrisCode [2–4]).

Alternative Segmentation Approaches

Some approaches to off-axis iris segmentation originate

from research for

▶ eye tracking applications. There are

many published eye tracking methods with various

segmentation procedures, any of which could be bor-

rowed or built upon for use in iris recognition. For

present purposes, we mention the eye tracking meth-

ods of Zhu and Yang [6], and the related iris recogni-

tion paper by Sung et al. [7], which uses an initial

step of eye corner detection (by ﬁlters which detect

sideways V-shapes); then Sobel edge detection to ﬁnd

the nearly vertical left and right ‘‘sides’’ of the iris–

sclera boundary, which also provides the center of the

iris and approximate center of the pup il; intensity

thresholding to identify dark regions; and ﬁnally,

▶ bi-

nary morphology (size and shape considerations) to

choose the pupil disk from among the dark masses

which lie w ithin the previously determined iris

boundary.

Zuo et al. [8] likewise use a combination of inten-

sity and location for pupil segmentation, after a

▶ circular Hough transform which generates the list

of pupil candidates. Additional morphology (ﬁnding a

convex hull) eliminates indentations in the pupil mass

caused by specular reﬂection at the edge of the pupil.

Finally, the pupil boundary is declared to be an ellipse

ﬁtted to the resulting region. Fitting an ellipse to the

pupil boundary is convenient for the eventual unwrap-

ping of the iris, but the ellipse is used here also to ﬁnd

the limbic boundary: the operator (1) is modiﬁed so

that, instead of families of circles, the contours of

integration are ellipses of the same eccentricity and

orientation as the pupil boundary.

Binary morphology may be the most ubiquitous

idea in pupil segmentation; the structuring elements

are easy to construct and combine in many convenient

and intuitive ways. To cite a third example, Ross and

Shah [9] ﬁnd the pupil using thresholding to identify

dark regions, median ﬁltering to reduce noise, and a

circle-ﬁtting procedure to locate the circle containing

the maximum number of black pixels, which is de-

clared to be the pupil boundary. The main innovation

in [9] is that the limbic boundary and eyelids are

detected using geodesic active contours. The boundary

is described as the zero level set of an embedding

function, where the embedding function is initialized

as the signed distance from any point in the image to a

circle concentric with the pupil boundary, but of

slightly larger radius. The active contour process

makes the embedding function evolve according to

gradients of the image grayscale values and the behav-

ior of the stopping function. The resulting boundary

contour traces the iris–sclera boundary arcs and the

iris–eyelid boundary, creating the correct iris mask.

Two sample images are shown in Fig. 4. Finally, the

boundary is projected behind the eyelid by ﬁtting

a circle to the iris–sclera po rtion of the boundary

contour, as this is needed for unwrapping.

In each of the segmentation methods mentioned

so far, the segmentation process is performed prior

to gaze angle determination: segmentation provides

the information required to ﬁnd the gaze angle using

the eccentricity idea. However, the segmentation pro-

cess and gaze angle determination can be intertwined

in other ways. In each of the two methods proposed by

Segmentation of Off-Axis Iris Images. Figure 3 From [5],

Fig. 3, p. 1170. ICE -1 file 244858, shown before and after

the rotation process which corrects the iris to orthographic

Segmentation of Off-Axis Iris Images

1161

Schuckers et al. [10], for instance, the rotation angle is

incremented to create a series of possible image trans-

forms. The angle is then selected according to an opti-

mizing procedure. One of the ways to use this is by the

application of camera calibration techniques to create

model planes: plotting and tracking reference points

from multiple eyes (image classes) at different angles,

as a training process to transform other images. The

other approach in [10] is to process the image through

a projective transform incremented w ithin a range

of pitch and yaw angles. The rotation angle pair is

then deﬁned as the one which maximizes the operator

(1) for the pupil, or in other words, the pitch and yaw

combination which best transform s the pupil bound-

ary back to a circle. This optimization process was

also used by Dorairaj et al. in [11].

Summary

Segmentation is always the critical step for correct

feature encoding . In the off-axis context, accurate seg-

mentation is required not only to know which pixels to

keep and which to throw away, but also (in most

approaches) to ascertain the gaze angle, from which

we can undo the effects of rotation before feature

encoding takes place.

The most common themes and techniques that

have been applied to on-axis segmentation appear

just as often in off-axis segmentation, notably edge

detection from large gradients, such as the Daugman

operator (1), and various combinations of threshold-

ing and binary morphology often brought to bear

on ﬁnding the pupil: thresholding to identify dark

regions, dilation and erosion for noise removal,

convexity to remove specular reﬂection, and assorted

ways to distinguish the pupil mass from the other dark

regions by its size and near-roundness. The same issues

that can plague on-axis segmentation are here as well,

but it is to be expected that they will be exacerbated

by the noncooperative image capture: blur, glare,

eyelashes, and eyelids conspire to thwart off-axis seg-

mentation and recognition.

There are no cut-and-dry ‘‘right answers’’ in iris

segmentation; the iris–sclera boundary in particular is

fairly subjective. Segmentation is often evaluated by

visual inspection or after its incorporation into a com-

plete recognition algorithm. In the latter case, the

performance of segmentation is affected by the choice

of the feature extraction process, the unwrapping pro-

cess (assumi ng unwrapping is required), and in the

off-axis case speciﬁcally, the results reﬂect the interac-

tions between the segmentation, the unwrapping, and

the determination of the gaze angle. Ultimately it

is true that segmentation must perform within a

start-to-ﬁnish recognition process, so all-up testing is

a necessary step, but it is important to recognize the

difﬁculty of cross-comparing segmentation perfor-

mance statistics in isolation.

The handling of off-axis segmentation is a relatively

new problem within the iris recognition community;

the methods discussed in this chapter were mostly

published in 2006–2007. For future work, a few obvi-

ous avenues of improvement are apparent. First, it has

been an educational experience for the author of this

chapter to explore a little more of the eye tracking

world, since they have been hard at work on quick

and robust segmentation at arbitrary angles for several

years. One such paper [6] was mentioned in this chap-

ter, but one could ask whether we have gleaned every-

thing we can from this and other image processing

applications. Second, there are ver y few off-axis

image databases. Most of the research discussed was

performed on the West Virginia Universit y database

[12], the NIST Iris Challenge Evaluation (ICE) data-

base, or small sets of images collected locally by the

researchers. Of course, the more databases that can be

assembled, the better. Lastly, considerations about the

three-dimensional shape of the iris have until now

taken a back seat to other questions. This is certainly

understandable, since these modeling issues would not

affect the comparison of two images taken at the same

angle, which is the traditional on- axis iris recognition

assumption. Similarly, corneal refraction will affect

Segmentation of Off-Axis Iris Images. Figure 4 Iris

segmentation result using geodesic active contours, from

[9], Fig. 10a-b, shown on images from the West Virgina

University off-angle iris image database [12]. Copyright

2006 IEEE.

1162

Segmentation of Off-Axis Iris Images

off-axis recognition performance, a preliminary study

of which was undertaken in [13]. When it comes to the

more challenging real-world iris recognition situa-

tions, even seemingly small improvements in iris mod-

eling may prove very helpful.

Related Entries

▶ Iris Encoding and Recognition

▶ Iris Features and Anatomy

▶ Iris Recognition, Overview

▶ Iris Segmentation Using Active Co ntours

References

1. Bonney, B.L., Ives, R.W., Etter, D.M., Du, Y.: Iris pattern

extraction using bit-planes analysis. In: Proceedings of the 38th

Asilomar Conference on Signals, Systems and Computers,

pp. 582–586 (Nov. 2004)

2. Daugman, J.G.: High conﬁdence visual recognition of persons

by a test of statistical independence. IEEE Trans. Pattern Anal.

Mach. Intell. 15(11), 1148–1161 (1993)

3. Daugman, J.G.: Biometric personal identiﬁcation system

based on iris analysis. US Patent 5,291,560, US Patent Ofﬁce,

Washington, D.C., 1 Mar 1994

4. Daugman, J.G.: How iris recognition works. IEEE Trans. Circuits

Syst. Video Technol. 14(1), 21–30 (2004)

5. Daugman, J.: New methods in iris recognition. IEEE Trans Syst

Man Cybern B 37(5) (2007)

6. Jie Zhu, Jie Yang: Subpixel eye gaze tracking. In: Proceedings of

the Fifth IEEE International Conference on Automatic Face and

Gesture Recognition, pp. 124–129, Washington, D.C. (May 2002)

7. Eric Sung, Xilin Chen, Jie Zhu, Jie Yang: Towards non-

cooperative iris recognition systems. In: Proceedings of the

2002 Seventh IEEE International Conference on Control,

Automation, Robotics, and Vision, pp. 990–995, (Dec. 2002)

8. Jinya Zuo, Kalka, N.D., Schmid, N.A.: A robust iris segmentation

procedure for unconstrained subject presentation. In: Special

Session on Research at the Biometric Consortium Conference,

2006 Biometrics Symposium.

9. Ross, A., Shah, S.: Segmenting non-ideal irises using geodesic

active contours. In: Special Session on Research at the Biometric

Consortium Conference, 2006 Biometrics Symposium.

10. Schuckers, S.A.C., Schmid, N.A., Abhyankar, A., Dorairaj, V.,

Boyce, C.K., Hornak, L.A.: On techniques for angle compensa-

tion in nonideal iris recognition. In: IEEE Trans. Syst. Man

Cybern. B 37(5) (2007)

11. Dorairaj, V., Schmid, N.A., Fahmy, G.: Performance evaluation

of non-ideal iris based recognition system implementing global

ICA encoding. In: Proceedings of the IEEE International Con-

ference on Image Processing, vol. 3, pp. 285–288 (2005)

12. Crihalmeanu, S., Ross, A., Hornak, L., Schuckers, S.: A Protocol

for Multibiometric Data Acquisition Storage and Dissemination,

Technical Report No. 5396, West Virginia University (2007)

13. Price, J.R., Gee, T.F., Paquit, V., Tobin, K.W. Jr.: On the efﬁcacy of

correcting for refractive effects in iris recognition. In: IEEE

Conference on Computer Vision and Pattern Recognition

(June 2007)

Semi-Supervised

Part of data are labeled, but the rest are unlabeled.

Algorithms are designed for separation of data under-

lying different classes, by taking the use of both labeled

information and unlabeled information.

▶ Linear Dimension Reduction

Sensitivity Analysis in Biometric

Systems

Sensitivity analysis in application to biometric systems

investigates the inﬂuence of varia tions of certain para-

meters on the recognition performance of biometric

systems. As an example, in the case of iris sample

synthesis the selected parameters may include but are

not limited to parameters characterizing ﬁbers, collar-

ette, and g lobal features of iris images.

▶ Iris Sample Synthesis

Service-Oriented Architecture

Service-oriented architecture (SOA) are software

architectures in which reusable services are deployed

onto application servers and then consumed by clients

in different applications or business processes [1]. SOA

is a paradigm for organizing and utilizing distributed

capabilities that may be under the control of different

ownership domains. It provides a uniform means to

offer, discover, interact with, and use capabilities to

Service-Oriented Architecture

1163

produce desired effects consistent with measurable

preconditions and expectations.

▶ Biometric Identity Assurance Services

Service Provider

In general, a business organization that provides con-

sumers with speciﬁed services, for a fee or sometimes

in response to a request. In the context of a federated

registered traveler program, service providers contract

with sponsoring entities (e.g., airports) to provide

enrollment and/or veriﬁcation services. These ser vices

generally include marketing, application processing,

applicant registration (including data collection/

biometric capture), data management/protection, card

production and issuance/re-issuance, and overall pro-

gram administration including fee collection. Veriﬁca-

tion services include provision of manned kiosks at

travel lanes (e.g., screening checkpoints, passport con-

trol) with hardware/software to perform card reading/

validation, biometric veriﬁcation (capture/matching),

security, and revocation checking.

▶ Registered Traveler

Session and Channel Variabilities

▶ Session Effects on Speaker Modeling

Session Effects on Speaker Modeling

DRISS MATROUF,JEA N-FRANC¸OIS BONASTRE

CERI, University of Avignon, Avignon, France

Synonyms

Session and channel variabilities; Factor analysis;

Nuisance attribute projection, Support vector

machine; Score normalization

Definition

In spite of the research efforts in the ﬁelds of speaker

feature characterization and speaker modeling, the

speaker recognizer must face a problem involving the

change of acoustic conditions, which vary in

an unforeseeable way, from one recording to another.

This phenomenon is generally referred to as session

mismatch or session variability. It is one of the greatest

sources of automatic speaker recognition (ASR)

performance degradation. The term session variability

encompasses a number of phenomen a including trans-

mission channel effects, environment noise (other

people, cars, TV, etc.), differing room acoustics (hall,

park, etc.), the microphone position relative to the

mouth, and the variability introduced by the spea-

ker himself. These sources of mismatch have the

potential to increase the rate of errors of ASR. It is

not easy to separately solve the problem of session

variability caused by each of these sources thus, all

the solutions suggested in literature deal with the prob-

lem of session mismatch in its globality. These pro-

posed solutions perform at various levels of the ASR

(feature space, model space and score space).

Introduction

In speaker recognition most of the errors are due not to

the similarity among voiceprints of different speakers,

but to the intrinsic variability of different utterances of

the same speaker. This awkward variability for the

speaker recognition is called session variability. One of

the largest challenges in speaker recognition applica-

tions is dealing with session variability [1]. Session

variability, also known as channel variability, is deﬁned

as the varia bility exhibited by given speaker from one

recording session to another and is caused by several

factors: channel effects, transducer characteristics, en-

vironment noise and intraspeaker variability. When

this mismatch is produced between sessions extracted

for enrolment and veriﬁcation phases, performance of

speaker recognition system is highly degraded and

numerous techniques has been investigated to com-

pensate session variability in several domains: In fea-

ture domain, compensation attempts at removing the

channel effects from the feature vectors prior to mod-

eling or testing. In model domain, compensation

attempts to transform speaker models to minimize

1164

Service Provider

the effects of varying channels. In score dom ain, com-

pensation attempts to remove model score scales and

shifts caused by varying input channel conditions.

Feature Domain Compensation

In literature concerning the variability of the session

we ﬁnd two t ypes of compensation: unsuper vised and

supervised. In the unsupervised case, speaker features

are transformed to accommodate the channel variation

without a priori knowledge of the channel character-

istics. On the other hand, superv ised compensation

estimates session-channel characteristics based on a

priori knowledge of all possible channels.

Unsupervised Compensation: There are three types

of unsupervised compensation. The ﬁrst type use the

temporal variability of feature vectors. It exploits the

fact that the channel effe cts vary slowly with respect to

the cepstral vectors (vocal tract characteristics evolu-

tion). For example, cepstral mean subtraction (CMS)

[2, 3] subtracts the cepstral mean of an utterance from

each of the cepstral vectors.

▶ RASTA-ﬁltering [4]

removes the slow varying components (corresponding

to the channel) from the sequence of cepstral vectors

by applying a bandpass ﬁlter. The second type of

unsupervised compensation transforms the distorted

features such that acoustic environments have minimum

effect on the distribution of the transformed features.

For example, in feature warping [6], observed features

are mapped to a target distribution (e.g., standard

normal) such that they follow the target distribution

over a sliding window of feature vectors. In short-time

Gaussianization [7], a linear transformation is applied

to the distorted features before mapping them to a

normal distribution. The transformation aims to dec-

orrelate the feature vectors making them more suitable

for diagonal covariance GMMs. It was found that

short-time Gaussianizat ion is superior to feature warp-

ing, especially at a low false acceptance rate. The third

type makes use of the statistical difference between the

clean acoustic models and the distorted speech to

estimate a transformation matrix to map the distorted

vectors to ﬁt the clean model. This type of technique

involves blind stochastic feature transformation [8].

Supervised Compensation: Supervised techniques

are based on channel type detection during recogni-

tion. Examples in this category are feature mapping [9],

spectral-magnitude matching [10], and stochastic

feature transformation [11]. In feature mapping, the

handset type of the testing utterance is identiﬁed by

a handset detector. Feature vectors are then mapped

to the channel-independent space based on the closest

Gaussian in the channel-dependent GMM. In spectral-

magnitude matching [ 10], a nonlinear polynomial

mapper is trained t o minimize the mean-squared spec-

tral magnitude error between speech arising from elec-

tret and carbon-button handsets. This mapping is

shown to be good at minimizing mismatches caused

by phantom formants, bandwidth widening, and spec-

tral ﬂattening due to channel nonlinearity. Stochastic

feature transformation (SFT) [11] aims to transform

the distorted features to ﬁt the clean speech models

by selecting the most appropriate pre-computed trans-

formation matrix. It has been shown that SFT can be

extended to nonlinear feature transformation to over-

come the nonlinear distortion [ 11].

Featur e mapping [9] is a supervised technique

that relies on a handset detector which contains the

information of all possible channels that the users

may use during veriﬁcation. In feature mapping, the

transformation is based on the top-1 Gaussian only.

Speciﬁcally, let GMM L

CDi

¼fp

CDi

; m

CDi

; S

CDi

j¼1

an M-mixture channel-dependent GMM for channel

i and GMM L ¼ {p

, m

, S

}

j¼1

, be an M-mixture

channel-independent root model. The mapping of a

distorted vector y in the channel-independent space

is given by

x ¼ðy  m

CDi

þ m

ð1Þ

x is the corresponding frame in the channel-independent

space. To account for the effect of other Gaussian com-

ponents on the transformed features, the transformation

should be based on a weighted average of all Gaussian

components, which leads to probabilistic feature

mapping (PFM).

Score Domain Compensation

Since the study of Li and Porter [40], various kinds of

score normalization techniques have been proposed in

the literature. Several of these are brieﬂy described in

the following section.

Score normalization techniques have mainly been

derived from the study of Li and Porter [40]. In their

paper, large variances had been observed from both

Session Effects on Speaker Modeling

1165

distributions of client scores (intraspeaker scores) and

impostor scores (interspeaker scores) during speaker

veriﬁcation tests. As a result of these observations,

the authors proposed solutions based on impostor

score distribution normalization in order to reduce

the overall score distribution variance (both client

and impostor distributions) of the spe aker veriﬁcation

system. The basic idea of the normalization technique

is to center the impostor score distribution by applying

on each score generated by the speaker veriﬁcation

system the following normalization. Let L

(X) denote

the score for speech signal X and speaker model l. The

normalized score

ðXÞ is then given as follows:

ðXÞ¼

ðXÞm

ð2Þ

where m

and s

are the normalization parameters for

speaker l. Both of these parameters are estimated from

a (pseudo) impostor score distribution. Different pos-

sibilities are available to compute the impostor score

distribution: Znorm, Hnorm, Tnorm, HTnorm,

Cnorm, Dnorm. All these normalization techniques

are decsribed in [1]. This family of normalization

techniques is the most commonly used in ASR. They

are di rectly derived from Eq. 2 in whi ch the scores are

normalized by subtracting the mean and then dividing

by the standard deviation. In the following the princi-

ple of the most popular is given:

Znorm The zero normalization (Znorm) technique

is directly derived from the work done in [12]. It w as

quickly incorporated in most speaker veriﬁcation

systems in the middle of the nineties. In practice, a

speaker model is tested against a set of speech signals

produced by a group of impostors resulting in an impos-

tor similarity score distribution. Speaker-dependent

mean and variance-normalization parameters are esti-

mated from this distribution and used to normalize

similarity scores yielded by the speaker veriﬁcation sys-

tem when running using Eq. 16. One of the advantages of

Znorm is that the estimation of the normalization para-

meters can be performed ofﬂine during speaker model

training.

Hnorm By observing that, for telephone speech,

most of the client speaker models respond differently

according to the handset type used during testing data

recording, Reynolds [13] proposed a variant of Znorm

technique referred to as handset normalization

(Hnorm) to deal with handset mismatch between the

training and testing phases. Here, handset-dependent

normalization parameters are estimated by testing each

speaker model against handset-dependent speech signals

produced by impostors. During testing, the type of hand-

set relating to the incoming speech signal determines the

set of parameters to use for score normalization.

Tnorm Still based on the estimate of mean and

variance parameters to normalize impostor score dis-

tribution, test-normalization (Tnorm), proposed in

[14], differs from Znorm by the use of impostor mod-

els instead of test speech signals. During testing, the

incoming speech signal is classically compared with

claimed speaker model as well as with a set of impostor

models to estimate impostor score di stribution and

normalization parameters consecutively. If Znorm is

considered as a speaker-dependent normalization tech-

nique, Tnorm can then be termed test-dependent. As the

same test utterance is used during both testing and

normalization parameter estimation, Tnorm avoids

mismatch between test and normalization utterances

which is a possible issue encountered when using

Znorm. Conversely, Tnorm has to be performed online

during testing.

Model Domain Compensation

Gaussian Mixture Models (GMM) used in a GMM -

UBM framework is perhaps one of the most common

conﬁgurations found in speaker veriﬁcation systems

[1]. For a several years now, new techniques that take

session variability (or speaker intra-variability) into

account have emerged providing a signiﬁcant increase

in system performance.

Among the approaches aimed at reducing the effect

of session variability, feature mapping was often used

alongside channel-labelled data with the assumption of

a discrete channel space. The novelty brought by the

factor analysis model is that it assumes the channel (or

session) variability space to be continuous. In this

model, the session variability effect is incorporated in

the speaker model throug h session-dependent GMM

mean supervectors offsets, constrained in a low dimen-

sional subspace.

A speaker model can be decomposed into three

different components: a speaker-session-independent

component, a speaker dependent component and a

session dependent component. A GMM mean super-

vector is deﬁned as the concatenation of the GMM

component means. Let D be the dimension of the

1166

Session Effects on Speaker Modeling

feature space and MD the dimension of a mean super-

vector where M is the number of Gaussian in the

GMM. A speaker and session independent model is

usually estimated in speaker veriﬁcation to represent

the inverse hypothesis: the UBM model. Let this model

be parameterized by y ¼ {m, S, a}. In the followi ng,

(h, s) indicates the session h of the speaker s. The factor

analysis model, in our case the eigenchannel MAP

estimator, can be w ritten as:

ðh; sÞ

¼ m þ Dy

þ Ux

ðh; sÞ

; ð3Þ

where m

(h,s)

is the session-speaker dependent super-

vector mean, D is MD  MD diagonal matrix, y

the

speaker vector (a MD vector), U is the session varia-

bility matrix of low rank R (a MD  R matrix) and

(h,s)

are the channel factors, a R vector (theoretically

(h,s)

does not dependent on s). Both y

and x

(h,s)

are

normally distributed among Nð0; IÞ. D satisﬁes

the following equation I ¼ tD

1

D where t is

the relevance factor required in the standard MAP

adaptation (DD

represents the a priori covariance

matrix of y

The success of the factor analysis model relies on a

good esti mation of the U matrix. This requires a sufﬁ-

ciently high amount of data in which a high number of

different recordings per speaker are available.

The veriﬁcation task is deﬁned as follows. A speaker

tar

is enrolled by the system with his training data Y

tar

Given a sequence of speech frames Y¼fy

...y

and the speaker s

tar

, the speaker veriﬁcation task con-

sists of determining if Y was spoken by s

tar

or not.

Using the factor analysis decomposition in both train-

ing and testing data, one can write:

ðh

tar;

tar

¼ m þ Dy

tar

þ Ux

tar

;

ðh

test;

test

¼ m þ Dy

test

þ Ux

test

;

ð4Þ

where the speaker s

tar

in the training data and s

test

the testing data have been distinguish ed. To address

session variability, the strategy adopted by [15, 16]

assumes that the test speaker has the same identity as

the target speaker, i.e. y

test

¼ y

tar

. The channel com-

ponent Ux

test

of the test segment is estimated under

this assumption. Indeed, the session component of the

target model Ux

tar

is replaced by the one estimated

from the test data Ux

test

. The world model in the score

equation remains unchanged. This strategy has several

drawbacks: the target speaker model is changed for

each test and signiﬁcant performance gains can only

be achieved when score normalization techniques are

employed. In [17], a hybrid domain normali zation

strategy is proposed which aims to withdraw the ses-

sion component in the test and training data. This can

be formulated as,

tar

¼ m þ Dy

tar

; m

test

¼ m þ Dy

test

: ð5Þ

In this strategy, speakers are assumed to be different

and are treated separately. When using a LLR-based

veriﬁcation approach, the speaker veriﬁcation score is

an expected log-likelihood ratio:

LLKðYjm

ðh

tar

ÞLLK ðYjmÞ; ð6Þ

where LLK (j) indicates the average of the log-likeli-

hood function over all frames. Here, all GMMs have the

same covariance matrices as well as the same mixture

weights (both dropped from the equation for clarity).

Two session compensation approaches can be adopted.

The ﬁrst approach involves performing compensation

at the frame level, where session compensation can be

seen as a front-end process. The second approach is a

hybrid compensation, where the session variability is

subtracted from the target speaker model (model do-

main) and the compensation in the testing data is

performed at the frame level (feature domain). The

following formula is used to remove the session effect

for each frame t (also successfully used by [18]):

t ¼ t 

g¼1

ðtÞfU  x

test

½g

ð7Þ

A similar approach, NAP (nuisance attribute projec-

tion) [5] is also used to reduce the impact of channel

(or session) [5]. NAP is dedicated to the

▶ SVM super-

vector (support vector machine) based systems [22],

which called the

▶ NAP-SVM approach.

Experiments

Speaker veriﬁcation experiments are performed using

the NIST SRE 2005 as a development set and the 2006

database for the validation set using only male speak-

ers. These databases consist entirely of telephony

speech and are referred to as the 2005 and 2006 pro-

tocols. The 2005 protocol consists of 274 speakers,

9,012 tests (951 target tests, the remainder being im-

postor trials) while the 2006 protocol consists of 354

speakers, 9,720 tests (of which 741 were target tests).

Session Effects on Speaker Modeling

1167

(Please note that 2005 protocol corresponds to the core

condition labeled as det7 and the 2006 protocol corre-

sponds to the core condition labeled as det3.) Results

are given in terms of equal-error-rate (EER) and the

minimum DCF (an a posteriori decision). On average,

train and test utterances contain 2.5 min of speech

where around 30% of speech frames per speaker have

been retained. The intersess ion variability matrix is

trained on the NIST-SRE-2004 database with 2,938

examples from 124 speakers, taking approximately 20

iterations to reach convergence.

ThebaselinesystemisastandardGMM-UBMsystem

[19]. Frames are composed of 19 LFCC parameters,

their derivatives, and 11 second derivatives (the fre-

quency w indow is restricted to 300–3,400 Hz). A nor-

malization process is applied so that the distribution of

each cepstral coefﬁcient is 0-mean and 1-variance for a

given utterance. Table 1 shows the results of the base-

line system and T-norm scores (Z and ZT-norm do not

bring any improvement).

The following outlines the experimental results

obtained with the implementation described in

[17]. Table 2 shows the improvement with score

normalization. Znorm consistently brings about an im-

provement in both protocols, while Tnorm is only ef-

fective in 2006. Indeed, the DCFmin drops from 1.83 to

1.64 with Znorm in 2005 and from 1.61 to 1.18 with

ZTnorm in 2006. While the behavior is different on

both protocols, ZTnorm appears to be the most conﬁ-

dent choice for score normalization.

Table 1 shows the results of the baseline system and

T-norm scores (Z and ZT-norm do not bring any

improvement).

Conclusion

The majority of techniques presented in this chap-

ter fail to reach their goal for different reasons. For

example, feature mappi ng, speaker model synthesis

(SMS) [20], and H-Norm are sub-optimal because

they consider only ﬁnite and discrete sources of session

mismatch. In theory, assuming that the mismatch

belongs to a ﬁnite number of categories simpliﬁes deal-

ing with the session variability problem. However, when

the session characteristics are very different from those

considered, inappropriate normalisation is applied

which can cause recognition or veriﬁcation error.

When using this technique it is necessary to automati-

cally detect the appropriate category for each test utter-

ance. This process is somewhat error-prone having the

potential to cause veriﬁcation errors when the inappro-

priate normalisation in applied.

The second key deﬁciency is not actually modeling

the effects of session variability but simply attempting

to suppress them. Feature warping, T-Norm and

Z-Norm ﬁt into this category. These methods have

no knowledge of the speciﬁc conditions encountered

in a recording but use some a-priori information of the

effects caused by session variability.

For several years, new appr oachs have emerged that

take session variability (or speaker intra-variability) into

account, providing a signiﬁcant increase in sys-

tem performance. The main feature of these techniques is

the way in which the issue of mismatch in GMM-based

speaker veriﬁcation in addressed. This is achieved

through the explicit modelling of session variability in

both the training and testing procedures. These directly

model the mismatch between sessions in a constrained

subspace of the GMM speake r model means. This

replaces the discrete categorisation of techniques such

as featur e mapping and H-Norm with a continuous

Session Effects on Speaker Modeling. Table 2 Baseline

and score normalized results obtained when using the

factor analysis model (rank = 40). Znorm consistently

provides an improvement over the baseline results.

DCFmin ( 100), EER(%)

SRE-05 SRE-06

DCFmin EER DCFmin EER

Nonorm 1.83 4.42 1.61 2.97

Tnorm 1.84 4.72 1.29 2.83

ZTnorm 1.72 4.62 1.18 2.15

Znorm 1.64 4.21 1.46 2.33

Session Effects on Speaker Modeling. Table 1 Results of

the baseline GMM-UBM system on the 2005 and 2006

protocol. DCFmin ( 100), EER(%)

SRE-05 SRE-06

DCFmin EER DCFmin EER

Nonorm 3.83 7.15 3.88 6.79

Tnorm 3.05 8.52 2.9 5.7

1168

Session Effects on Speaker Modeling

vector -valued representation of the session conditions. A

key strength of this approach is the avoidance of data

labelling requirements due to the particular training

methods that are employed.

Related Entries

▶ Gaussian Mixture Models

▶ Speaker Matching

▶ Speaker Recognition, Overview

▶ Uiniversal Background Models

References

1. Bimbot, F., Bonastre, J.-F., Fredouille, C., Gravier, G.,

Magrin-Chagnolleau, I., Meignier, S., Merlin, T., Ortega-Garcia,

J., Petrovska, D., and Reynolds, D.A.: A tutorial on text-

independent speaker veriﬁcation. EURASIP J. Appl. Signal

Process., Special issue on biometric signal processing (2004)

2. Atal, B.S.: Effectiveness of LPC characteristics of the

speech waves for A.S.I and A.S.V. J. Acoust. Soc. Am. JASA

55 (1974)

3. Furui, S.: An analysis of long-term variation of feature para-

meters of speech and its application to talker recognition. Elec-

tron. Commun. 57-A, 34–42 (1977)

4. Hermansky, H., Morgan, N.: RASTA processing of speech. IEEE

Trans. Speech Audio Process. 2, 578–589 (1994)

5. Solomonoff, A., Campbell, W.M., and Boardman, I.: Advances

in channel compensation for SVM speaker veriﬁcation. In Pro-

ceedings of ICASSP (2005)

6. Pelecanos, J., Sridharan, S.: Feature warping for robust spea-

ker veriﬁcation. 2001: a Speaker Odyssey. The Speaker Recog-

nition Workshop, Chania, Greece, pp. 213–218. Chania,

Crete (2001)

7. Xaing, B., Chaudhari, U., Navratil, J., Ramaswamy, G., and

Gopinath, R.: Short-time Gaussianization for robust speaker

veriﬁcation. In: Proceedings of International Conference on

Acoustics Speech and Signal Processing (ICASSP), Vol. 1, pp.

681–684 (2002)

8. Yiu, M.C.K.K., Mak, M.W., Kung, S.: Blind stochastic feature

transformation for channel robust speaker veriﬁcation. J. VLSI

Signal Process. 42(2), 117–126 (2006)

9. Reynolds, D.: Channel robust speaker veriﬁcation v ia feature

mapping. In: Proceedings of International Conference on

Acoustics Speech and Signal Processing (ICASSP 2003), vol. 2.

Hong Kong, China (2003)

10. Quatieri, D.A.R.T.F., OLeary, G.C.: Estimation of handset non-

linearity with application to speaker recognition. IEEE Trans.

Speech Audio Process. 8(5), 567–584 (2000)

11. Kung, S.Y., Mak, M.W., and Lin, S.H.: Biometric Authentication:

A Machine Learning Approach. New Jersey: Prentice-Hall,

Upper Saddle River (2005)

12. Li, K.P., Porter, J.E.: Normalizations and selection of speech

segments for speaker recognition scoring. In: Proceedings of

International Conference on Acoustics Speech and Signal Pro-

cessing (ICASSP 98), pp. 595–598 (1988)

13. Reynolds, D.A.: The effects of handset variability on speaker

recognition performance: experiments on the switchboard cor-

pus. In: International Conference on Acoustics, Speech, and

Signal Processing ICASSP (1996)

14. Auckenthaler, R., Carey, M., Lloyd-Thomas, H.: Score normali-

zation for text-independent speaker veriﬁcation system. Digital

Signal Processing (DSP), a review journal – Special issue on

NIST 1999 speaker recognition workshop 10(1–3), 42–54 (2000)

15. Kenny, P., Boulianne, G., Ouellet, P., Dumouchel, P.: Factor

Analysis Simpliﬁed. In: Proceedings of International Conference

on Acoustics Speech and Signal Processing (ICASSP 2005), vol. 1.

Philadelphia, USA (2005)

16. Vogt, R., Baker, B., Sridharan, S.: Modelling Session Variability

in Text-Independent Speaker Veriﬁcation. In: Proceedings of

Interspeech, European Conference on Speech Communication

and Technology (Eurospeech 2005). Lisboa, Portugal (2005)

17. Matrouf, D., Scheffer, N., Fauve, B., Bonastre, J.-F.: A straight-

forward and efﬁcient implementation of the factor analysis

model for speaker veriﬁcation. In INTERSPEECH Conference.

Antewerp, Belgium (2007)

18. Vair, C., Colibro, D., Laface, P.P.: Channel factors compensation in

model and feature domain for speaker recognition. In: Odyssey’06,

the Speaker Recognition Workshop. San Juan, Puerto Rico (2006)

19. Reynolds, D.A., Quatieri, T.F., Dunn, R.B.: Speaker veriﬁcation

using adapted gaussian mixture models. Digit. Signal Process.

(DSP), a review journal – Special issue on NIST 1999 speaker

recognition workshop, 10(1–3), 19–41 (2000)

20. Teunen, R., Shahshahani, B., and Heck, L.: A model-based

transformational approach to robust speaker recognition. In:

International Conference on Spoken Language Processing,

pp. 495–498 (2000)

21. Hanson, B.A. et al.: Subband or cepstral domain ﬁltering for

recognition of Lombard and channel distorded speech. In Pro-

ceedings of ICASSP-93, pp. 79–82 (1993)

22. Collobert, R., Bengio, S.: SVM Torch: support Vector Machines

for large-scale regression problems. J. Mach. Learn. Res. 1,

143–160 (2001)

SFinGe

RAFFAELE CAPPELLI

Biometric System Laboratory – DEIS – University of

Bologna

Synonyms

Synthetic Fingerprint Generator

SFinGe

1169