Analysis of Feature Extraction and Channel Compensation in a GMM Speaker Recognition System

Abstract

In this paper, several feature extraction and channel compensation techniques found in state-of-the-art speaker verification systems are analyzed and discussed. For the NIST SRE 2006 submission, cepstral mean subtraction, feature warping, RelAtive SpecTrAl (RASTA) filtering, heteroscedastic linear discriminant analysis (HLDA), feature mapping, and eigenchannel adaptation were incrementally added to minimize the system's error rate. This paper deals with eigenchannel adaptation in more detail and includes its theoretical background and implementation issues. The key part of the paper is, however, the post-evaluation analysis, undermining a common myth that ldquothe more boxes in the scheme, the better the system.rdquo All results are presented on NIST Speaker Recognition Evaluation (SRE) 2005 and 2006 data.

Full text

Page 1

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR1
Analysis of feature extraction and channel
compensation in a GMM speaker recognition system
Luk´ aˇ s Burget?, Member, IEEE, Pavel Matˇ ejka, Member, IEEE, Petr Schwarz, Member, IEEE,
Ondˇ rej Glembek, Student Member, IEEE, and Jan “Honza”ˇCernock´ y, Member, IEEE
Abstract—In this paper, several feature extraction and channel
compensation techniques found in state-of-the-art speaker veri-
fication systems are analyzed and discussed. For the NIST SRE
2006 submission, Cepstral Mean Subtraction, Feature Warping,
RASTA filtering, HLDA, Feature Mapping and Eigenchannel
Adaptation were incrementally added to minimize the system’s
error rate. The paper deals with Eigenchannel Adaptation
in more detail, and includes its theoretical background and
implementation issues. The key part of the paper is however
the post-evaluation analysis, undermining a common myth that
“the more boxes in the scheme, the better the system”. All results
are presented on NIST SRE 2005 and 2006 data.
Index Terms—Speaker recognition, GMM, Feature Warping,
RASTA, HLDA, Feature Mapping, Eigenchannel Adaptation.
EDICS Category: SPE-SPKR
I. INTRODUCTION
In the NIST 2006 Speaker Recognition Evaluation [1], the
Brno University of Technology (BUT) participated with its
own submission and also contributed to systems developed by
the STBU1consortium. Both the BUT and STBU primary sys-
tems were fusions of several individual subsystems, namely:
systems based on Gaussian Mixture Modeling (GMM) [2], and
systems based on sequence kernel Support Vector Machines
(SVM) classifying either GMM mean supervectors [3] or
vectors constructed from Maximum Likelihood Linear Regres-
sion (MLLR) transformations [4], which are transformations
commonly used in speech recognition for speaker adaptation.
In this paper, we provide an analysis of the BUT GMM system
that took part in both the BUT and STBU primary systems,
and which was also submitted as a BUT stand-alone secondary
system. The overall description of the BUT and STBU systems
can be found in [5], [6].
The BUT GMM system is based on a standard Univer-
sal Background Model-Gaussian Mixture Modeling (UBM-
All authors are with Speech@FIT, Faculty of Information Technology, Brno
University of Technology, Czech Republic. Luk´ aˇ s Burget is the corresponding
author.
This work was partly supported by by European projects AMIDA (IST-
033812) and Caretaker (FP6-027231), by Grant Agency of the Czech Republic
under projects No. 102/05/0278 and GP102/06/383 and by the Czech Ministry
of Education under project No. MSM0021630528. The hardware used in this
work was partially provided by CESNET under projects No. 119/2004, No.
162/2005 and No. 201/2006.
We are grateful to the anonymous reviewers for their valuable comments
on the first version of this paper. Many thanks to Ken Froehling from the
Faculty of Electrical Engineering and Communication of BUT for help with
the English of the paper.
1BUT, TNO Human Factors (The Netherlands), Spescom DataVoice (South
Africa) and the University of Stellenbosch (South Africa).
GMM) paradigm [2] and employs a number of techniques
that have previously proven to improve GMM modeling ca-
pability and help fight against the main problem in speaker
verification - diversity in channel and acoustic conditions.
These techniques are: Cepstral Mean Subtraction, Feature
Warping [7], RelAtive SpecTrAl (RASTA) filtering [8], Het-
eroscedastic Linear Discriminant Analysis (HLDA) [9], Fea-
ture Mapping [10] and Eigenchannel Adaptation [11]. The aim
of this paper is to analyze the importance of the individual
techniques in terms of their contribution to overall system
performance.
The paper is organized as follows: A detailed description
of the BUT GMM speaker recognition system is provided
in section II. Section III documents building the system and
reports the improvements in performance obtained by adding
individual techniques. Section IV presents our post-evaluation
activity and analyzes the importance of the individual tech-
niques in the full system. The result obtained by fusing the
GMM system with the SVM-based systems are presented in
section V. We conclude the paper in section VI.
II. SYSTEM DESCRIPTION
A. Features
The features used in the system are Mel-frequency cepstral
coefficients (13 MFCC coefficients including C0, 20 ms win-
dow, 10 ms shift, 23 bands in a Mel filter bank). To compensate
for channel mismatch in different conversations, three simple
feature processing techniques were successively applied: the
cepstral mean over the whole conversation is subtracted from
the features, Feature Warping [7] (3 sec window, warping
into a normal distribution) is applied and finally temporal
trajectories of individual feature vector coefficients are filtered
using a standard RASTA filter [8]2. After this processing,
each feature vector is augmented with its first, second and
third order derivatives. This results in 52 dimensional feature
vectors containing information about the context of 13 frames.
B. Segmentation
At this stage, non-speech frames are discarded and only
speech frames are considered in the following stages of train-
ing models and verification. Speech/non-speech segmentation
is performed by our Hungarian phoneme recognizer [12],
2Cepstral Mean Subtraction has no effect after the application of Feature
Warping and RASTA filtering as both techniques also ensure the mean
removal. However, it will be interesting to see the effectiveness of these
techniques compared to Cepstral Mean Subtraction alone.

Page 2

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR2
where all phoneme classes are linked to speech classes.
A postprocessing with two rules based on the short time
energy of the signal is applied: 1) If the average energy in
a speech segment is 30dB less than the maximum energy in
the conversation side, then the segment is labeled as silence.
2) If the energy in the opposite conversation side3is bigger
than the maximum energy minus 3dB in the processed side,
the segment is also labeled as silence.
C. HLDA
As the next step, we have employed Heteroscedastic Linear
Discriminant Analysis (HLDA), which is also in common
use in speech recognition systems. HLDA provides a linear
transformation that can de-correlate the features and reduce
the dimensionality while preserving the discriminative power
of features. The theory of HLDA is described in detail in [9],
[13]. HLDA needs classes to estimate its class-covariance
statistics (which are then used to estimate the transformation
matrix). For this purpose, GMM with 2048 Gaussian compo-
nents is trained on test data from SRE2004 and the feature
frames aligned with individual GMM mixture components
are considered as classes. HLDA transformation reducing the
dimensionality from 52 to 39 is estimated. GMM is then
updated in the new HLDA space (by projecting collected class-
covariance and mean statistics through HLDA transformation).
Features are also projected into HLDA space and GMM
is re-estimated (still only on SRE2004 test data) by few
additional Expectation-Maximization (EM) iterations to obtain
the Universal Background Model (UBM).
D. Feature Mapping
To further compensate for channel mismatch, Feature Map-
ping [10] was applied to all enrollment and test conversations.
Feature Mapping requires a set of models, each adapted from
UBM using data of particular acoustic condition (channel).
We have used 14 such models: 6 models were adapted for
3 channels (cell,cord,stnd) and 2 genders given the labels
from 2004 test data. The remaining 8 models were initially
adapted for 4 channels (cdma, cord, elec, gsmc) and 2 genders
using the TNO Feature Mapping labels used in SRE-2005.
However, these 8 models were then iteratively used to re-
cluster the training data in an unsupervised fashion and again
adapted using the new clustering (20 iterations lead to stable
clustering) [14].
E. Training speaker model and verification
Each speaker model is obtained by a traditional relevance
Maximum A-Posteriori (MAP) adaptation [15] of UBM using
enrollment conversation. Only means are adapted with a
relevance factor τ = 19.
In the verification phase, standard Top-N Expected Log
Likelihood Ratio (ELLR) scoring [15] is used to obtain a
verification score, where N = 10 in our system. However,
for each trial, both the speaker model and UBM are adapted
3In NIST SRE2006 evaluations, our system participated only in the primary
condition, where two separate recordings for the two sides of each phone
conversation are available.
to the channel of test conversation using simple Eigenchannel
Adaptation [11] prior to computing the log likelihood ratio
score. Note, that when T-norm [16] is used to normalize the
score, each T-norm model is also adapted to the channel of
relevant tested conversation.
F. Eigenchannel subspace estimation
We adopted the term ‘eigenchannel’ as used in speaker
recognition from Kenny [17]. It was introduced to the NIST
SRE by SDV in 2004 [11], revisited by Kenny and Vogt [18]
in SRE 2005, and again by several sites in various forms in
SRE 2006.
Let supervector be a MD dimensional vector constructed
by concatenating all GMM mean vectors and normalized by
corresponding standard deviations. M is the number if Gaus-
sian mixture components in GMM and D is dimensionality
of features. Before Eigenchannel Adaptation can be applied,
we must identify directions in which the supervector is mostly
affected by a changing channel. These directions, which we
will refer to as eigenchannels, are defined by columns of
MD × R matrix V, where R is the chosen number of
eigenchannels (R = 30 in our system). The matrix V is given
by R eigenvectors of average within class covariance matrix,
where each class is represented by supervectors estimated on
different segments spoken by the same speaker.
More precisely, we have selected all (310) speakers from
NIST SRE2004 data for which at least two conversations
are available. For each speaker, i, and all his conversations,
j = 1,...,Ji, UBM is adapted to obtain a supervector,
sij. The corresponding speaker average supervector given by
si =?Ji
and resulting vectors form columns of MD × J matrix S,
where J is the number of all conversations from all selected
speakers (J = 2961 in our case). Eigenchannels (columns of
matrix V) are given by R eigenvectors of MD×MD average
within speaker covariance matrix4
R largest eigenvalues. Unfortunately, for our system, where
MD = 2048 × 39 = 79872, direct computation of these
eigenvectors is unfeasible. A possible solution is to compute
eigenvectors, V?, of J × J matrix
are then given by V = SV?. In case the maximum a-
posteriori (MAP) criterion is used for Eigenchannel adaptation
(see below), the length of each eigenchannel must be also
normalized to the average within speaker standard deviation of
supervectors along the direction of the eigenchannel (i.e. each
eigenvector obtained in the previous step must be multiplied
by the square root of the corresponding eigenvalue). This
normalization is irrelevant in the case of maximum likelihood
(ML) criterion.
j=1sij/Ji is subtracted from each supervector, sij,
1
JSSTcorresponding to
1
JSTS; eigenchannels
G. Eigenchannel Adaptation
Once the eigenchannels are identified, a speaker model (or
UBM) can be adapted to the channel of a test conversation by
shifting its supervector in the directions given by eigenchan-
nels to better fit the test conversation data. Mathematically,
4Note that matrix
over columns of S is guaranteed by the subtraction of the speaker average
supervectors described above.
1
JSSTis a true covariance matrix as the zero mean

Page 3

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR3
this can be expressed as finding the channel factors, x, that
maximize the following MAP criterion:
p(O|s + Vx)N(x;0,I),
(1)
where s is a supervectorrepresenting the model to be adapted5,
p(O|s + Vx) is the likelihood of the test conversation given
the adapted supervector (model) and N(·;0,I) denotes a
normally distributed vector. Assuming a fixed occupation of
the Gaussian mixture components by test conversation frames,
ot,t = 1,...,T, it can be shown [11] that x maximizing
criterion (1) is given by:
x = A−1
M
?
m=1
VT
m
T
?
t=1
γm(t)ot− µm
σm
,
(2)
where Vmis M×R part of matrix V correspondingto the mth
mixture component, γm(t) is the probability of occupation
mixture component m at time t, µmand σmare the mixture
component’s mean and standard deviation vectors and
A = I +
M
?
m=1
VT
mVm
T
?
t=1
γi(t).
(3)
In our implementation, occupation probabilities, γm(t), are
computed using UBM and assumed to be fixed for given test
conversation. This allows us to pre-compute matrix A−1only
once for each test conversation. For each frame, only Top-
N occupation probabilities are assumed not to be zero. In
the following ELLR scoring, only the same top-N mixture
components are also considered. All these facts ensure that
adapting and scoring different speaker or T-norm models on a
test conversation can be performed very efficiently.
Eigenchannel Adaptation can be also performed by max-
imizing ML criterion instead of MAP criterion. This corre-
sponds to dropping the prior term, N(x;0,I), in criterion (1)
and term I in equation 3. In our experiments, there is always
enough adaptation data (test conversations contain approx-
imately 2.5 minutes of speech) making the prior term in
MAP criterion negligible. Therefore, we have not found any
differences in performance when using the two criteria.
Our system uses a very simple scheme of modeling channel
variability that affects only the verification phase. However,
more sophisticated schemes can be considered. In [19] the
verification phase is equivalent to that described here, however,
modeling channel variability is considered also in training
speaker models. This may become important especially when
speaker models are trained using more than one enrollment
conversation.
A very elaborate scheme can be found in [17], where mod-
eling channel variability is considered in all phases: training
background model, training speaker models and verification.
Instead of finding eigenvectors, channel subspace V is ob-
tained also by maximizing MAP criterion similar to (1). For
enrollment data, instead of finding MAP point estimates of
model parameters, posterior probabilities of model parameters
are considered and integrated over to obtain the likelihood
score for a test conversation.
5Note again that by our definition, a supervector is a mean supervector
normalized by the corresponding standard deviations.
III. BUILDING THE SYSTEM
In the following experiments, results will be presented for
“1-side training, 1-side test, all trials” condition from SRE2005
NIST evaluation, which we have used for system development,
and for primary condition (1-side training, 1-side test, English
only trials) from SRE2006 NIST evaluation. In the tables,
results are presented in terms of EER (Equal Error Rate) and
Cmin
SRE2006 primary condition, performances are also presented
in the form of DET (Detection Error Tradeoff) curves.
Table I and figure 1 document the process of building our
system. It shows line-by-line the improvementsin performance
obtained by successively adding different techniques. Our
starting point was GMM system with 2048 Gaussian mixture
components, features were 13 MFCC coefficients augmented
with their deltas and processed by cepstral mean subtraction.
The error rate of this system is very high and is almost
halved by simply adding RASTA filtering. Replacing RASTA
with Feature Warping improved the performance; however,
a further small gain was obtained from the combination of
both techniques. The application of RASTA filtering on top
of Feature Warping appeared to be slightly more advanta-
geous than doing it in the opposite order. In the next two
steps, features were also augmented with double-delta and
triple-delta coefficients. While adding double-deltas is clearly
beneficial for both SRE2005 and SRE2006 evaluation sets,
the advantage of adding triple-deltas, which we have seen
during development on SRE2005 data, was not confirmed on
SRE2006.
The following three steps, each significantly improving
the system performance, were: projection of 52 dimensional
features into 39 dimensional HLDA space, application of our
14 classes Feature Mapping and Eigenchannel Adaptation.
So far, all the presented results were obtained without
normalizing the verification scores by any standard tech-
nique, such as T-normalization or Z-normalization (Z-norm/T-
norm) [16]. As can be seen in Table I, T-norm was not effective
in improving the performance of our full system. We have
also experimented with Z-norm and ZT-norm, nevertheless,
results obtained with all normalization techniques were mixed
and unconvincing. This contradicted the conclusions drawn
in [17], [18], [19], where Z-norm or ZT-norm was found
necessary for making channel variability modeling techniques
really effective.
Most of GMM based speaker verification systems, for which
the results are published by various sites, use less than 2048
Gaussian components. The last line of table I show results
for a system with the usual number of only 512 Gaussian
components, which is otherwise identical to our full system.
It can be seen that the performance of a 2048 component
system is superior to this smaller one.
Detas defined by SRE2006 NIST evaluation rules [1]. For
IV. POST-EVALUATION ANALYSIS
In the previous section, we have shown how adding individ-
ual techniques improves system performance. However, it will
be even more interesting to see whether and how the individual
techniques are important in the full system.

Page 4

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR4
SystemSRE2005
EER
26.6%
14.3%
12.4%
11.2%
10.6%
9.7%
7.3%
4.6%
4.6%
4.9%
TABLE I
SRE2006
EER
23.8%
11.8%
10.0%
9.1%
9.3%
8.2%
6.2%
4.0%
4.0%
4.7%
Cmin
Det
.089
.055
.052
.047
.047
.042
.033
.020
.020
.026
Cmin
Det
.088
.059
.051
.049
.048
.041
.032
.020
.018
.024
MFCC+∆, CMS, 2048 G.
+ RASTA
+ Feature Warping
+ ∆∆
+ ∆∆∆
+ HLDA (52→39)
+ Feature Mapping
+ Eigenchannel Adapt.
+ T-norm
Full system, 512 Gauss.
THE IMPROVEMENTS IN PERFORMANCE OBTAINED BY SUCCESSIVELY
ADDING DIFFERENT TECHNIQUES.
0.05 0.1 0.2 0.5 1 2 5 10 20 40
0.05
0.1
0.2
0.5
1
2
5
10
20
40
False Alarm probability (in %)
Miss probability (in %)


2048 Gauss., 13 MFCC + delta, CMS
with RASTA
with Feature Warping
with Feature Warping and RASTA
+ double delta
+ triple delta
+ HLDA (52 to 39 dim. reduction)
+ Feature mapping (14 classes)
+ Eigenchannel Adaptation
512 instead of 2048 Gaussians
+ T−norm
Fig. 1.
techniques.
DET curves showing improvement in successive adding different
A. The importance of RASTA and Feature Warping
Table II and figure 2 present results obtained with the
baseline full system6and two of its modifications leaving out
either RASTA filtering or Feature Warping. While Feature
Warping turns out to be an important part of the system,
leaving out RASTA filtering even slightly improves the system
performance. This may support the conclusions in [8], where
RASTA was found to discard important speaker information
lying under its cut-off frequency and a filter more appropriate
for speaker verification was designed.
B. Analyzing the effect of HLDA
The left half of Table III shows the effect of HLDA for
systems without the following Feature Mapping and Eigen-
channel Adaptation. The first two results (already presented
in Table I) demonstrate the effectiveness of HLDA at this
6System with 2 Gender Feature Mapping (see below) is used as a baseline
system in this experiment for efficiency reasons.
0.05 0.1 0.2 0.5 1 2 5 10 20 40
0.05
0.1
0.2
0.5
1
2
5
10
20
40
False Alarm probability (in %)
Miss probability (in %)


with Feature Warping and RASTA
no No Weature Warping
no RASTA
Fig. 2.The importance of RASTA filtering and Feature Warping.
SystemSRE2005
EER
4.5%
4.4%
5.1%
TABLE II
SRE2006
EER
3.8%
3.8%
4.3%
Cmin
Det
.019
.019
.020
Cmin
Det
.020
.019
.021
Full system
No RASTA
No Feature Warping
THE IMPORTANCE OF RASTA AND FEATURE WARPING.
stage. The dimensionality reduction from 52 to 39 was chosen
as exactly the same scheme had already been proven to be
effective for speech recognition [20]. Since it was not clear
whether this scheme is optimal for our speaker verification
system, reductions to various dimensionalities were examined
and the best results were obtained without any dimensionality
reduction7(last line of Table III).
The situation is different in the right half of Table III,
where Feature Mapping and Eigenchannel Adaptation are
used. Performances of systems using HLDA are still supe-
rior to the one that leaves HLDA out; however, the system
with dimensionality reduction outperforms the one without
reduction. The possible explanation is that the significant
increase in GMM (and supervector) size makes it impossible
to robustly estimate eigenchannels given the limited number
of supervectors available for their estimation. The summary of
HLDA and MLLT results can be also found in figure 3.
C. Eigenchannels vs. Feature Mapping
The left half of Table IV and dotted DET curves in figure 4
show the effect of Feature Mapping for systems without
the following Eigenchannel Adaptation. The first two results
(already presented in Table I) demonstrate the effectiveness
7HLDA without dimensionality reduction is often referred to as a Maximum
Likelihood Linear Transform (MLLT) [21]

Page 5

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR5
SystemWithout channel comp.
SRE2005
EER
Cmin
Det
10.6%.047
9.7%.042
8.7%.038
With channel comp.
SRE2005
EER
Cmin
Det
5.1% .024
4.5%.019
4.6%.023
SRE2006
EER
9.3%
8.2%
7.5%
TABLE III
SRE2006
EER
5.0%
3.8%
4.2%
Cmin
Det
.048
.041
.037
Cmin
Det
.025
.020
.021
No HLDA
HLDA 52 → 39
HLDA 52 → 52
THE EFFECT OF HLDA ON SYSTEM PERFORMANCE.
0.05 0.1 0.2 0.5 1 2 5 10 20 40
0.05
0.1
0.2
0.5
1
2
5
10
20
40
False Alarm probability (in %)
Miss probability (in %)


no HLDA with FM and Eigenchannel Adapt.
no HLDA no FM and Eigenchannel Adapt.
HLDA 52 → 39 with FM and Eigenchannel Adapt.
HLDA 52 → 39 no FM and Eigenchannel Adapt.
HLDA 52 → 52 with FM and Eigenchannel Adapt.
HLDA 52 → 52 no FM and Eigenchannel Adapt.
Fig. 3.
reduction) on system performance.
The effect of HLDA and MLLT (HLDA without dimensionality
of Feature Mapping at this stage. In the third line, the
performance of a system using Feature Mapping based on only
two models adapted on male and female specific data is shown.
This allows us to compensate for the fact that our system uses
only a single UBM instead of the usual approach where two
genders are handled separately using two UBMs. Although
such 2-gender Feature Mapping significantly outperforms the
system leaving Feature Mapping out, it still reaches only about
half of the gain in performance compared to 14 classes Feature
Mapping used in our final system.
The right half of Table IV and solid DET curves in figure 4
show similar results for systems applying also Eigenchan-
nel Adaptation. We can see that without Feature Mapping,
EigenchannelAdaptation causes an impressive improvementin
system performance (more than 50% relative in both EER and
Cmin
after the Eigenchannel Adaptation is applied, which allows
us to simplify the verification system considerably by leaving
Feature Mapping out. In fact, the use of our 14 classes Feature
Mapping causes even slight degradation in the performance. It
was surprising for us that even 2-gender Feature Mapping did
not turn out to be effective, as eigenchannels are not trained to
model the directions of differences between male and female
specific models.
Detpoints). There is no advantage in using Feature Mapping
0.05 0.1 0.2 0.5 1 2 5 10 20 40
0.05
0.1
0.2
0.5
1
2
5
10
20
40
False Alarm probability (in %)
Miss probability (in %)


14 classes FM + Eigenchannel Adapt.
14 classes FM only
2 gender FM + Eigenchannel Adapt.
2 gender FM only
Eigenchannel Adaptation only
no FM, no Eigenchannel Adaptation
Fig. 4.The importance of Feature Mapping and Eigenchannel Adaptation.
020406080100
2
4
6
8
10
number of eigenchannels
EER [%]


SRE 2005
SRE 2006
Fig. 5.
adaptation.
The dependency of EER on the number of eigenchannels used for
D. Number of eigenchannels
The number of eigenchannels was chosen to be R = 30 for
our system submitted to SRE2006 NIST evaluations. Figure 5
shows the dependency of EER on the number of eigenchannels
used for adaptation. A similar trend has also been observed
for Cmin
Det
values. It can be seen that our system is not very
sensitive to the exact selection of the number of eigenchannels.

Page 6

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR6
System Without Eigenchannel Adapt.
SRE2005
EER
Cmin
Det
9.7%.042
7.3%.033
8.5%.037
TABLE IV
With Eigenchannel Adapt.
SRE2005
EER
Cmin
Det
4.6%.019
4.6%.020
4.5% .019
SRE2006
EER
8.2%
6.2%
7.6%
SRE2006
EER
3.8%
4.0%
3.8%
Cmin
Det
.041
.032
.036
Cmin
Det
.020
.020
.020
No Feature Mapping
14 classes Feature Mapping
2-gender Feature Mapping
THE IMPORTANCE OF FEATURE MAPPING AND EIGENCHANNEL ADAPTATION.
SystemSRE2005
EER
4.4%
SRE2006
EER
3.6%
Cmin
Det
.017
Cmin
Det
.018 no T-n., no RASTA, no FM, 50 EA
TABLE V
RESULTS OF THE FINAL TUNED AND SIMPLIFIED SYSTEM.
V. FUSING WITH SVM BASED SYSTEMS
The performance of the GMM system was also tested in
combination with speaker recognition systems based on a
different classification paradigm – Support Vector Machines
(SVM). Figure 6 contains a summary of results for SRE2006
primary condition. Results are presented for BUT stand-alone
systems as well as for fused systems that were BUT and STBU
submissions into the SRE2006 NIST evaluations.
These systems (from the worst to the best) are:
• SVM-MLLR, where MLLR and constrained MLLR (CM-
LLR) speaker adaptation matrices from a speech recog-
nition system are classified by SVM. Two variants are
shown: with and without T-norm
• SVM-GMM, where GMM supervectors are classified by
SVMs. Two variants are shown: with and without T-norm
• GMM is the full system described in this paper. Two
variants (already presented in Table I) are shown: with
and without T-norm
• BUT02 is a fusion of 3 systems: GMM, SVM-GMM and
SVM-MLLR, all with T-norm applied
• BUT01 (BUT primary system) is a fusion of 6 systems:
GMM, SVM-GMM and SVM-MLLR, each in two vari-
ants: with and without T-norm
• STBU1-N is fusion of 10 systems from the partners in
the STBU consortium.
• STBU1-U (STBU primary system) is fusion of the same
10 systems, plus one more SVM-GMM system imple-
menting unsupervised adaptation to test data according
to SRE2006 NIST evaluation rules [1].
A detailed description of different systems can be found in [5],
[6]. The fusion was performed using linear logistic regression
implemented in the FoCal toolkit8and it is also described and
commented on in [6].
VI. CONCLUSION
BUT GMM system contains nothing more than techniques
that were already published – its main contribution is in a
thorough analysis and discussion of these techniques in a full
speaker recognition system. Starting in the feature extraction,
8www.dsp.sun.ac.za/∼nbrummer/focal/
0.05 0.1 0.2 0.5 1 2 5 10 20 40
0.05
0.1
0.2
0.5
1
2
5
10
20
40
False Alarm probability (in %)
Miss probability (in %)


STBU1−U (adapted)
STBU1−N (non−adapted)
BUT01
BUT02
GMM (BUT03)
GMM with T−norm
SVM−GMM
SVM−GMM with T−norm
SVM−MLLR
SVM−MLLR with T−norm
Fig. 6.Fusion of GMM system with SVM based systems.
the main conclusion is that RASTA did not help in the full
system. On the other hand, HLDA significantly improved its
performances, although we know that there is still work to
be done (different dimensionality reductions examined with
the full system, not using triple-deltas, etc.). In fighting the
channel variability, even the simple Eigenchannel Adaptation
turned out to be very effective, erasing the advantages of Fea-
ture Mapping, which is actually not important when applied
together with Eigenchannel Adaptation. Table V presents the
results of the final tuned and simplified system, containing
50 eigenchannels, no T-norm, no RASTA and no Feature
Mapping. All the conclusions may, however, not hold for other
than 1-side training, 1-side test condition examined in this
work. Our current and future work aims at these conditions as
well as at using the described GMM system as an excellent
baseline for further experiments.
REFERENCES
[1] “The NIST year 2006 speaker recognition evaluation plan,” 2006, Avail-
able from: http://www.nist.gov/speech/tests/spk/2006/sre-06 evalplan-
v9.pdf.
[2] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, “Speaker verification
using adapted gaussian mixture models,” Digital Signal Processing,
vol. 10, no. 1–3, pp. 19–41, 2000.
[3] A. Solomonoff, W. Campbell, and I. BoardmanCampbell, “Advances in
channel compensation for SVM speaker recognition,” in Proc. ICASSP,
vol. I, Philadelphia, PA, USA, Mar. 2005, pp. 629–632.

Page 7

IEEE TRANS. ON AUDIO, SPEECH AND LANGUAGE PROCESSING, VOL. X, NO. Y, MONTH YEAR7
[4] A. Stolcke, L. Ferrer, S. Kajarekar, E. Shriberg, and A. Venkataraman,
“MLLR transforms as features in speaker recognition,” in Proc. Eu-
rospeech, Lisbon, Portugal, Sep. 2005, pp. 2425–2428.
[5] P. Matˇ ejka, L. Burget, P. Schwarz, O. Glembek, M. Karafi´ at, F. Gr´ ezl,
J. ˇCernock´ y, D. A. van Leeuwen, N. Br¨ ummer, and A. Strasheim,
“STBU system for the NIST 2006 speaker recognition evaluation,” in
Proc. ICASSP, Honolulu, Hawaii, USA, Apr. 2007.
[6] N. Br¨ ummer, L. Burget, J.ˇCernock´ y, O. Glembek, F. Gr´ ezl, M. Karafi´ at,
D. A. van Leeuwen, P. Matˇ ejka, P. Schwarz, and A. Strasheim, “Fusion
of heterogeneous speaker recognition systems in the STBU submission
for the NIST speaker recognition evaluation 2006,” IEEE Trans. on
Audio, Speech and Language Processing, 2007, submitted.
[7] J. Pelecanos and S. Sridharan, “Feature warping for robust speaker
verification,” in Proc. A Speaker Odyssey, Crete, Grece, 2001, pp. 213–
218.
[8] S. van Vuuren and H. Hermansky, “On the importance of components
of the modulation spectrum for speaker verification,” in Proc. ICSLP,
vol. 7, Sydney, Australia, May 1998, pp. 3205–3208.
[9] N. Kumar, “Investigation of silicon-auditory models and generalization
of linear discriminant analysis for improved speech recognition,” Ph.D.
dissertation, John Hopkins University, Baltimore, 1997.
[10] D. A. Reynolds, “Channel robust speaker verification via feature map-
ping,” in Proc. ICASSP, vol. II, Apr. 2003, pp. 53–56.
[11] N. Br¨ ummer, “Spescom DataVoice NIST 2004 system description,” in
Proc. NIST Speaker Recognition Evaluation 2004, Toledo, Spain, Jun.
2004.
[12] P. Schwarz, P. Matˇ ejka, and J.ˇCernock´ y, “Hierarchical structures of
neural networks for phoneme recognition,” in Proc. ICASSP, France,
May 2006, pp. 325–328.
[13] M. J. F. Gales, “Semi-tied covariance matrices for hidden markov
models,” IEEE Transactions Speech and Audio Processing, vol. 7, pp.
272–281, 1999.
[14] M. Mason, R. Vogt, B. Baker, and S. Sridharan, “Data-driven clustering
for blind feature mapping in speaker verification,” in Proc. Eurospeech,
Lisbon, Portugal, Sep. 2005, pp. 3109–3112.
[15] D. A. Reynolds, “Comparison of background normalization methods
for text-independent speaker verification,” in Proc. Eurospeech, Rhodes,
Greece, Sep. 1997, pp. 963–966.
[16] R. Auckenthaler, M. Carey, and H. Lloyd-Tomas, “Score normalization
for text-independent speaker verification systems,” Digital Signal Pro-
cessing, vol. 10, pp. 42–54, 2000.
[17] P. Kenny and P. Dumouchel, “Disentangling speaker and channel effects
in speaker verification,” in Proc. ICASSP, vol. 1, Montreal, Canada, May
2004, pp. 47–40.
[18] R. Vogt, B. Baker, and S. Sridharan, “Modelling session variability
in text-independent speaker verication,” in Proc. Eurospeech, Lisbon,
Portugal, Sep. 2005, pp. 3117–3120.
[19] R. Vogt and S. Sridharan, “Experiments in session variability modelling
for speaker verification,” in Proc. ICASSP, vol. 1, Toulouse, France, May
2006, pp. 897–900.
[20] T. Hain, L. Burget, J. Dines, G. Garau, M. Karafiat, M. Lincoln,
I. McCowan, D. Moore, V. Wan, R. Ordelman, and S. Renals, “The
2005 AMI system for the transcription of speech in meetings,” in Proc.
NIST Rich Transcription 2005 Spring Meeting Recognition Evaluation,
Edinburgh,UK, Jul. 2005.
[21] R. Gopinath, “Maximum likelihood modeling with Gaussian distribu-
tions for classification,” in Proc. ICASSP, vol. II, Seattle, Washington,
USA, May 1998, pp. 661–664.
Luk´ aˇ s Burget (Ing. [MS]. Brno University of Tech-
nology, 1999, Ph.D. Brno University of Technology,
2004) is employed as an assistant professor at the
Faculty of Information Technology, University of
Technology, Brno, Czech Republic. The topic of his
PhD dissertation that he successfully defended in
November 2004 was: ”Complementarity of Speech
Recognition Systems and System Combination”.
From 2000 to 2002, he was a visiting researcher at
OGI Portland, USA under the supervision of Prof.
Hynek Hermansky. He is a member of IEEE and
ISCA. His scientific interests are in the field of speech processing, namely
acoustic modeling for speech recognition.
Pavel Matˇ ejka (Ing. [MS]. Brno University of Tech-
nology, 2001) is a PhD student at the Institute of
Radio-electronics, Faculty of Electrical Engineering
and Communication and Department of Computer
Graphics and Multimedia, FIT, BUT. He is planning
to submit his doctoral thesis “Language identifi-
cation based on phonetic cues” in summer 2007.
He has been with the Anthropic speech processing
group at the Oregon Graduate Institute of Science
and Technology, USA. He is a member of IEEE
and ISCA. His research interests include speaker
recognition, language identification, speech recognition, namely phoneme
recognition based on novel feature extractions (temporal patterns) and neural
networks. He has been active also in keyword-spotting and on-line implemen-
tation of speech processing algorithms. He was a finalist in the Student paper
contest at ICASSP2006 in Toulouse.
Petr Schwarz (Ing. [MS]. Brno University of Tech-
nology, 2001) is a PhD student of Speech processing
group at the Faculty of Information Technology
(FIT), BUT since September 2001 and is planning to
submit his doctoral thesis “Robust phoneme recog-
nition” in 2007. He has been with the Anthropic
speech processing group of the Oregon Graduate
Institute of Science and Technology, USA. He is a
member of IEEE and ISCA. His research interests
include speech recognition, namely phoneme recog-
nition based on novel feature extractions (temporal
patterns) and neural networks. He has been active also in keyword-spotting
and on-line implementation of speech processing algorithms.
Ondˇ rej Glembek (Ing. [MS]. Brno University of
Technology, 2005) was student at the Brno Univer-
sity of Technology, Faculty of Electrical Engineering
and Computer, later the Faculty of Information Tech-
nology from 1999. From September till December
2003, he was at the University of Joensuu, Finland as
a participant of the Socrates/Erasmus program. From
October till November 2004, he was working on
a project concerning wavelet transforms at Izhevsk
State Technical University, Izhevsk, Russia. From
2005, he is a PhD student in Speech@FIT - he is
concentrating on acoustic modeling for speech recognition, recognition of
Czech and STK toolkit development.
Jan “Honza” ˇCernock´ y (Ing. [MS] 1993 Brno
University of Technology (BUT); Dr. [PhD] 1998
Universite Paris XI and BUT) was with the Institute
of Radio-electronics, BUT (Faculty of Electrical
Engineering and Computer Science) as an assistant
professor from 1997. Since February 2002, he is
with the Faculty of Information Technology (FIT),
BUT as an Associate Professor (Doc.) and Deputy
Head of the Institute of Computer Graphics and
Multimedia. With Prof. Hynek Hermansky he is
leading the Speech@FIT group at FIT BUT. He
supervises several PhD students, and coordinates Speech@FIT activities
in several European and national projects. His research interests include
signal processing, speech processing (very low bit rate coding, verification,
recognition), segmental methods, data-driven determination of speech units
and speech corpora. He is a member of IEEE and ISCA and serves on the
board of the Czechoslovak section of IEEE.