Using Discrete Probabilities With Bhattacharyya Measure for SVMBased Speaker Verification
ABSTRACT Support vector machines (SVMs), and kernel classifiers in general, rely on the kernel functions to measure the pairwise similarity between inputs. This paper advocates the use of discrete representation of speech signals in terms of the probabilities of discrete events as feature for speaker verification and proposes the use of Bhattacharyya coefficient as the similarity measure for this type of inputs to SVM. We analyze the effectiveness of the Bhattacharyya measure from the perspective of feature normalization and distribution warping in the SVM feature space. Experiments conducted on the NIST 2006 speaker verification task indicate that the Bhattacharyya measure outperforms the Fisher kernel, term frequency loglikelihood ratio (TFLLR) scaling, and rank normalization reported earlier in literature. Moreover, the Bhattacharyya measure is computed using a dataindependent squareroot operation instead of datadriven normalization, which simplifies the implementation. The effectiveness of the Bhattacharyya measure becomes more apparent when channel compensation is applied at the model and score levels. The performance of the proposed method is close to that of the popular GMM supervector with a small margin.
 [Show abstract] [Hide abstract]
ABSTRACT: Spoken language recognition refers to the automatic process through which we determine or verify the identity of the language spoken in a speech sample. We study a computational framework that allows such a decision to be made in a quantitative manner. In recent decades, we have made tremendous progress in spoken language recognition, which benefited from technological breakthroughs in related areas, such as signal processing, pattern recognition, cognitive science, and machine learning. In this paper, we attempt to provide an introductory tutorial on the fundamentals of the theory and the stateoftheart solutions, from both phonological and computational aspects. We also give a comprehensive review of current trends and future research directions using the language recognition evaluation (LRE) formulated by the National Institute of Standards and Technology (NIST) as the case studies.Proceedings of the IEEE 05/2013; 101(5):11361159. · 6.91 Impact Factor  [Show abstract] [Hide abstract]
ABSTRACT: Optimization techniques have been used for many years in the formulation and solution of computational problems arising in speech and language processing. Such techniques are found in the BaumWelch, extended BaumWelch (EBW), Rprop, and GIS algorithms, for example. Additionally, the use of regularization terms has been seen in other applications of sparse optimization. This paper outlines a range of problems in which optimization formulations and algorithms play a role, giving some additional details on certain application problems in machine translation, speaker/language recognition, and automatic speech recognition. Several approaches developed in the speech and language processing communities are described in a way that makes them more recognizable as optimization procedures. Our survey is not exhaustive and is complemented by other papers in this volume.IEEE Transactions on Audio Speech and Language Processing 01/2013; 21(11):22312243. · 1.68 Impact Factor  SourceAvailable from: Kong Aik Lee
Conference Paper: Spoken Language Recognition in the Latent Topic Simplex.
INTERSPEECH 2011, 12th Annual Conference of the International Speech Communication Association, Florence, Italy, August 2731, 2011; 01/2011
Page 1
Accepted for IEEE TASL
1
Abstract—Support vector machines (SVMs), and kernel
classifiers in general, rely on the kernel functions to measure the
pairwise similarity between inputs. This paper advocates the use
of discrete representation of speech signals in terms of the
probabilities of discrete events as feature for speaker verification
and proposes the use of Bhattacharyya coefficient as the
similarity measure for this type of inputs to SVM. We analyze the
effectiveness of the Bhattacharyya measure from the perspective
of feature normalization and distribution warping in the SVM
feature space. Experiments conducted on the NIST 2006 speaker
verification task indicate that the Bhattacharyya measure
outperforms the Fisher kernel, term frequency loglikelihood
ratio (TFLLR) scaling, and rank normalization reported earlier
in literature. Moreover, the Bhattacharyya measure is computed
using a dataindependent squareroot operation instead of data
driven normalization, which simplifies the implementation. The
effectiveness of the Bhattacharyya measure becomes more
apparent when channel compensation is applied at the model and
score levels. The performance of the proposed method is close to
that of the popular GMM supervector with a small margin.
Index Terms— Bhattacharyya coefficient, speaker verification,
support vector machine, supervector.
I. INTRODUCTION
PEAKER verification is the task of verifying the identity of
a person using his/her voice [1]. The verification process
typically consists of extracting a sequence of shortterm
spectral vectors from the given speech signal, matching the
sequence of vectors against the claimed speaker’s model, and
finally comparing the matched score against a verification
threshold. Recent advances reported in [18] show an
emerging trend in using support vector machines (SVMs) for
Copyright (c) 2010 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubspermissions@ieee.org.
Manuscript received December 11, 2009. The associate editor coordinating
the review of this manuscript and approving it for publication was Dr. Nestor
Becerra Yoma.
Kong Aik Lee, Chang Huai You, and Haizhou Li are with the Institute for
Infocomm Research, Agency for Science, Technology and Research
(A*STAR), Singapore. (email: kalee@i2r.astar.edu.sg; echyou@i2r.a
star.edu.sg; hli@i2r.astar.edu.sg). The work of Haizhou Li was partially
supported by Nokia Foundation.
Tomi Kinnunen is with the School of Computing, University of Eastern
Finland, Finland (email: tkinnu@cs.joensuu.fi). The work of T. Kinnunen
was supported by the Academy of Finland (project no. 132129,
“Characterizing individual information in speech”).
Khe Chai Sim is with the School of Computing, National University of
Singapore, Singapore (email: simkc@comp.nus.edu.sg).
speaker modeling. One reason for the popularity of SVM is its
good generalization performance.
The key issue in using SVM for classifying speech signals,
which have a varying number of spectral vectors, is how to
represent them in a suitable form as SVM can only use input
of a fixed dimensionality. A common approach is to map the
sequences explicitly into fixeddimensional vectors known as
supervectors. Classifying variablelength speech sequences is
thereby translated into a simpler task of classifying the
supervectors. For instance, in [3] speech vectors are mapped
to a highdimensional space via timeaveraged polynomial
expansion. In [4], speech vectors are used to train a Gaussian
mixture model (GMM) via the adaptation of a socalled
universal background model (UBM). The supervector is then
formed by concatenating the mean vectors of the adapted
GMM. In [5], supervectors are formed by stacking the
likelihood scores with respect to a cohort of anchor models on
a perutterance basis. In [6], the maximum likelihood linear
regression (MLLR) transform is used to form the supervectors
comprising of the transform coefficients. It should be
mentioned that the term “supervector” was originally used in
[4, 9] to refer to the GMM supervector. Here, we use similar
term in a broader sense referring to any fixeddimensional
vector that represents a speech sequence as a single point in
the vector space, having a much higher dimensionality than
the original input space.
This paper advocates the use of discrete events (or symbols)
and their probabilities to construct supervectors. Discrete
events arise naturally in modeling many types of data, for
example, letters, words, and DNA sequences. Speech signals
can also be represented as sequences of discrete symbols by
using a quantizer. Notably, highlevel feature extraction (e.g.,
idiolect, phonotactic, prosody) usually produces discrete
symbols. For instance, in [10] speech signals are converted
into sequences of phone symbols and then represented in
terms of phone ngram probabilities. Discrete probabilities are
also useful in modeling prosodic feature sequences [11]. In [7,
8], we investigated the use of discrete acoustic events derived
using the UBM. In this paper, we show that various discrete
representations mentioned above can be summarized under the
maximum a posteriori (MAP) parameter estimation
framework [12]. Since they can be unified within similar
framework, an SVM kernel designed for one discrete
representation would be useful for the others. Another
practical virtue of discrete representation is that the estimation
Using Discrete Probabilities with Bhattacharyya
Measure for SVMbased Speaker Verification
Kong Aik Lee, Chang Huai You, Haizhou Li, Tomi Kinnunen, and Khe Chai Sim
S
Page 2
Accepted for IEEE TASL
2
of discrete distribution is simple and any arbitrarily shaped
distribution is possible since it is nonparametric.
Another challenge concerning the use of supervectors with
SVM is feature normalization – the process where the
elements of a feature vector are scaled or warped prior to
SVM modeling. Feature normalization in the SVM kernel
space is closely related to the similarity measure of
supervectors. In this study, since the supervectors represent
the probability distributions of discrete events, we propose
using Bhattacharyya coefficient [13] as the similarity measure.
The Bhattacharyya measure is symmetric as opposed to other
probabilistic measures such as KullbackLeibler (KL)
divergence [14], which is nonsymmetric and has to be
simplified and approximated substantially to arrive at a
symmetric kernel. While the Bhattacharyya measure is
simpler, dataindependent and more effective, we will also
show how it is related to and different from the Fisher kernel
[2], term frequency loglikelihood ratio (TFLLR) [10], and
rank normalization [15] proposed earlier for similar form of
supervectors.
The remainder of this paper is organized as follows. We
introduce the MAP framework for the estimation of discrete
probabilities in Section II. Using the UBM as a soft quantizer,
we describe the process of constructing supervectors using
discrete probabilities and show its relevance to the Fisher
kernel in Section III. We analyze the Bhattacharyya measure
from feature normalization perspective in Section IV. The
performance evaluation is reported in Section V. Finally,
Section VI concludes the paper.
II. ESTIMATION OF DISCRETE PROBABILITIES
Let some discrete events
possible outcomes, and let
the ith event
i
ω =
{}
12
,,,
T
= x xx
…
X
, our goal is to estimate the probabilities
{
12
,,
ω ωΩ =
}
,
M
ω
…
of observing individual events in X .
Using the maximum a posteriori (MAP) estimation criterion
[12], the solution is given by the mode of the posterior
distribution, as follows
{}
,1,2,,
i
Se iM
==
…
have M
i ω be the probability of observing
( )
i
P e
. Given a speech segment
ie , i.e.,
( ) ( )
g
{}
argmax log
Ω
P
Ω =
?
ΩΩ
X
, (1)
where ( )
g Ω is the prior distribution of the parameters Ω. Let
=∑x
X
()
: ~1
t
t
i
i
e
n
(2)
denote the number of occurrences of the event
where the notation ~
t
x
1,2,,tT
=
…
, is encoded as the event
can be expressed as
⎧
Ω =Ω
⎨
⎩
⎧
=Ω
⎨
⎩
⎧
= Ω +
⎨
⎩
ie in X ,
tx , for
ie
denotes that the vector
ie . The MAP estimate
( )
()
( )
( )
X
( )()
1
1
∑
1
argmax log
Ω
~,
argmax log
Ω
argmax log
Ω
log.
i
T
ttt
t
M
n
i
i
M
ii
i
gPeeS
g
gn
ω
ω
=
=
=
⎫
⎬
⎭
Ω∈
⎫
⎬
⎭
⎫
⎬
⎭
∏
∏
x
?
X
(3)
Since the likelihood function
distribution, the prior density can be assumed as a Dirichlet
distribution (i.e., a conjugate prior for the parameters of the
multinomial distribution) [12], as follows
()
P
Ω
X
follows a multinomial
( )
Ω =
1
1
i
M
ci
i
gK
ν
ω
−
=
∏
, (4)
where
distribution and
the MAP estimate can then be solved, subject to the
constraints
1
1
i
=
∑
and
Lagrange multipliers, to give
iν are the set of positive parameters for the Dirichlet
K is a normalization factor. Using (4) in (3),
c
M
iω=
0
i ω ≥
, using the method of
(
n
)
()
1
1
1
i
⎡
⎣
i
i
M
jj
j
n
ν
ω
?
ν
=
+−
=
+−
⎤
⎦
∑
X
X
,
1,2,,iM
=
…
. (5)
The MAP estimate of discrete probabilities in (5) is given
by the sum of the observed statistics
of the prior distribution. For a flat prior, whereby
(5) reduces to the popular maximum likelihood (ML) estimate:
in and the parameters
iν
01
iν − =
,
(
T
)
i
i
n
ω =
?
X
,
1,2,,
iM
=
…
, (6)
since
higher belief in the observed statistics than the prior
information.
We assume in (5) that a rule of correspondence has been
defined between the speech feature vectors and the set of
events such that
()
1,2,
i
=
of observing those events in X . In (2), speech feature vectors
are quantized as discrete events or symbols on a frameby
frame basis. These events may correspond to the codewords of
a vector quantization (VQ) codebook [14] or the Gaussian
densities in a UBM as shown in the next section. The discrete
events may also correspond to abstract linguistic units such as
phonemes, syllables, words, or subsequences of n symbols
(i.e., ngrams). For instance, in spoken language recognition
[16] and speaker recognition utilizing highlevel features [10,
11], the events represent ngrams of phones, words or some
prosodic features. In these methods, phone recognizers or
prosodic feature extractors are used to discover the events set
from the speech signals.
()
1
M
in
=
i
T
=
∑
X
. ML estimate is used when we have
in X for
,
M
…
represent the counts
III. CONSTRUCTING SUPERVECTOR USING DISCRETE
PROBABILITIES
A. UBM as Soft Quantizer
A universal background model, or UBM, is a GMM trained
to represent a speakerindependent distribution [17]. In this
regard, the UBM is usually trained, using the expectation
maximization (EM) algorithm [14], from tens or hundreds of
hours of speech data gathered from a large number of
speakers.
A UBM, denoted by Θ, with M mixture components is
characterized by the following probability density function:
Page 3
Accepted for IEEE TASL
3
()
()
1
,
M
iii
i
p
λ
=
Θ =∑
xx μ Σ
N
, (7)
where
i Σ is the covariance matrix of the ith Gaussian component.
The mixture weights satisfy the constraint
covariance matrices are assumed to be diagonal in this paper.
Let each of the Gaussian densities represent a discrete event
ie . Given a speech segment X , the number of occurrences of
event
probabilities
iλ is the mixture weight,
iμ is the mean vector, and
∑
1
1
M
iλ
=
i
=
and the
ie is computed by accumulating the posterior
()
(
N
)
()
1

x μ Σ
,
,
,
itii
t
M
jtjj
j
P i
λ
λ
=
Θ =∑
x μ Σ
x
N
(8)
evaluated for the ith Gaussian component for the whole
utterance, as follows
()
()
1
,
T
it
t
n P i
=
=Θ
∑
x
X
.
(9)
The UBM quantizes the input vectors into discrete symbols,
much similar to the VQ codebook except that the codewords
are now modeled as Gaussian densities. Since the Gaussian
densities can be overlapped, rather than partitioned, soft
membership can be computed based on the Bayes rule as
given in (8). The UBM Θ together with (9) thereby define the
set of discrete events, S, and the rule of correspondence
between the feature vectors and the events.
Finally, to obtain the MAP estimate, the parameters
(5) are set to
iν in
1
ii
M
ντλ − = ⋅⋅
, (10)
where
parameter τ has to be greater or equal to 0. This is known as
the τinitialization method in [12]. Feasible values for τ range
from 0 to 1, which we have found effective for this
application. Equation (10) controls the broadness of the prior
density ( )
g Ω in (4) with the parameter τ. When τ is large, the
prior density is sharply peaked around the UBM weights
in which case the resulting MAP estimate approaches
Conversely, if τ is small the MAP estimate approaches the ML
estimate. In particular, the MAP estimate in (5) reduces to the
ML estimate in (6) for
0
τ =
.
iλ are the weights of the UBM and the controlled
iλ ,
iλ .
B. Constructing Supervector
The discrete probabilities Ω?
conveniently represented in functional form as
the variable h represents any event in S such that
()
ii
P he
ω
== ? for hS
∈
. That is, the function
probability mass function (PMF) [14]. We can express the
PMF in vector form as
{
12
,,
ω ω
= ??
}
,
M
ω ?
P h , where
…
can be
( )
( )
P h is the
( )
P e
()()
[][]
TT
1212
,,,,,,
MM
P e P e
ω ω
?
ω
?
==
p
?
……
, (11)
where the superscript T denotes transposition. The vector p
has a fixed dimensionality, M, equivalent to the cardinality of
the event set S. It represents the speech segment X in terms
of the distribution of discrete events observed in X . These
attributes fulfill our requirement of supervector representation.
For the case of UBM, the relation between Ω? and X is given
by (5) and (9), and the dimensionality of the supervector is
determined by the number of Gaussian densities in the UBM.
C. Fisher Information
The concept of mapping sequences into supervectors is
commonly interpreted as a sequence (or dynamic) kernel. The
earliest example of sequence kernel can be traced back to [18]
in which the Fisher kernel was proposed. The Fisher kernel
maps a sequence into a supervector by taking the derivatives
of the loglikelihood function with respect to the parameters
of the model. Let the model be the UBM as defined in (7).
Taking the derivative of the loglikelihood function with
respect to the weights
i
by the duration T , we obtain
iλ , for
1,2,,M
=
…
, and normalizing
()
()
1
log
,
T
t
t
i
i
i
p
P i
λ
F
TT
λ
=
∇Θ
Θ
==
⋅
∑
x
X
. (12)
We deliberately write (12) in terms of
in (8), to establish the connection to our earlier discussion.
Using (9) in (12) and normalizing by the respective Fisher
information [18] we arrive at
()
,
t
P i
Θ
x
, as defined
()
i
i
i
nT
F
I
′=
X
, (13)
where the constant
information definition as
defined in (12).
Notice that
. Hence, Fisher mapping essentially boils down to the ML
estimate of discrete distribution, with additional normalization
factors depending on the Fisher information. The supervector
[cf. (11)] is now given by
iλ has been absorbed as part of the Fisher
()
{
iii
IEF
}
2
λ=
, where
iF is as
()
inT
X
gives the ML estimate as shown in (6)
T
1
I
2
FISHER
′
12
,,,
M
I
M
I
ω
?
ω
?
ω
?
⎡
⎣
⎤
⎥
⎦
=⎢
p
…
.
(14)
In practice, the Fisher information
estimated by replacing the expectation with sample average
computed from a large background corpus. The Fisher
information normalizes individual dimensions of the
supervector to the same scale corresponding to the mean
square value of the discrete probabilities estimated from the
background samples. By so doing, all dimensions are treated
equally in the meansquare sense when used as inputs to
SVM.
Recall that the Fisher mapping in (12) was obtained by
taking the derivative of the loglikelihood function with
respect to the weights of the UBM. In addition to the weights,
taking the derivative with respect to the mean vectors and
covariance matrices, as originally proposed in [18], increases
the dimensionality of the supervector. These additional
dimensions are not considered in this paper as they do not
correspond to any discrete probability interpretation, which is
iI ,
1,2,,iM
=
…
, are
Page 4
Accepted for IEEE TASL
4
the focus of the paper. A full account of using this form of
supervector for speaker verification can be found in [2].
IV. THE BHATTACHARYYA MEASURE
The Bhattacharyya coefficient [13] is commonly used in
statistics to measure the similarity of two probability
distributions. It is computed by integrating the square root of
the product of the two distributions. For discrete distributions,
the Bhattacharyya coefficient is given by
( )( )
ab
h S
∈
P h P h
ρ=∑
, (15)
where S is the set of discrete events. The coefficient ρ lies
between zero and unity, where
distributions are fully overlapped, while
case of nonoverlapping distributions.
Using
()
,ai i a
P he
ω== ?
and
Bhattacharyya measure can be written in vector form as
1
ρ = indicates that the two
0
ρ = occurs for the
()
bi
P he
=
, i b
ω= ?
in (15), the
T
,,
1
M
i ai ba
′
b
′
i
ρ ω ω
?
=
==
∑
p p
?
, (16)
where the supervector is now given by
T
BHAT
′
12
,,,
M
ω
?
ω
?
ω
?
⎡⎤
⎦
=⎣
p
…
. (17)
Clearly, the Bhattacharyya measure is symmetric and it
represents an inner product in the supervector space. Hence, it
can be used as a legitimate kernel function [19] in SVM.
From feature normalization perspective, the squareroot
operator has an effect in normalizing the contribution of
individual dimensions to the inner product. As shown by the
solid curve in Fig. 1, higher gain is applied to rare events, i.e.,
those events with lower probabilities. The gain reduces
gradually (so as the slope of the curve) when the input
approaches unity. By so doing, the situation where rare events
are outweighed by those with higher probabilities is avoided.
The squareroot operator could also be interpreted as a
warping function, where the horizontal axis is shrunk for
inputs close the zero and stretched for inputs close to unity, as
depicted Fig. 1. This is different from the normalization
scheme in the Fisher kernel, where constant scaling is applied
to individual dimension based on the Fisher information
estimated from a background corpus.
A. Term Frequency LogLikelihood Ratio (TFLLR)
Term frequency loglikelihood ratio (TFLLR) was
introduced in [10] for the scaling of ngram probabilities.
Since each ngram can be regarded as a discrete event
ngram probabilities can be expressed as PMF and supervector
as given in (11). In this regard, the event set S consists of all
unique ngrams. The TFLLR scales individual dimensions of
the supervector (i.e., the ngram probabilities) in proportion to
the square root of the inverse ngram probabilities computed
from a large background corpus. Denoting the background
probabilities as
ie , the
iλ , the supervector is now given by
T
12
TFLLR
′
12
,,,
M
M
ω
λ
ω
λ
ω
λ
⎡
⎣
⎤
⎥
⎦
=⎢
p
???
…
. (18)
For discrete probabilities derived from the UBM quantizer,
the background probabilities
weights of the UBM since the weights are estimated from a
large background corpus.
The TFLLR deemphasizes frequent events and emphasizes
rare events. This is similar in spirit with the Bhattacharyya
measure, except that individual dimension is subjected to
constant scaling instead of warping. Hence, TFLLR scaling
falls into the same category as the Fisher kernel from the
perspective of feature normalization in the supervector space.
iλ correspond directly to the
B. Rank Normalization
In [15], rank normalization was proposed for normalizing
the supervectors of ngram probabilities. Elements of the
0.01 0.09 0.160.25 0.360.49
Input
0.640.81 1.0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Output
Fig. 1. The horizontal axis is warped according to a squareroot function.
0 0.050.100.15 0.200.250.30 0.350.40 0.450.50
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Input (×102)
(a)
Output
00.050.100.150.200.250.300.350.400.450.50
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Input (×102)
(b)
Output
Fig. 2. Warping functions produced by rank normalizing individual
dimension of the supervector for (a) male and (b) female populations. The
solid curves were obtained by ensemble averaging the warping functions
across all dimensions, where the size of UBM was set to M = 1000 in the
experiment.
Page 5
Accepted for IEEE TASL
5
supervector are processed separately where a warping
function is used for mapping each dimension to a uniform
distribution over the interval from zero to unity. The warping
function is nonparametric and is derived from the concept of
cumulative density function (CDF) matching, similar to that
used in histogram equalization (HEQ) [20] and feature
warping [21]. Since the CDF of the targeted uniform
distribution is linear (and monotonically increasing) in the
interval [0, 1], CDF matching amounts to a procedure where
each element of the supervector is replaced by its rank in a
background corpus. Denoting the rank of
supervector is now given by
i ω ? as
ir , the
T
12
rank
′
,,,
M
R
r
R R
rr
⎡
⎣
⎤
⎥
⎦
=⎢
p
…
, (19)
where R is the number of reference samples in the background
data. Let
rank
background set
{
:
ii
rbB b
=∈
iB be the set of R values for the ith dimension. The
ir of
iB whose values are smaller than
< ?
i ω ? is given by the number of elements in the
i ω ? :
}
i
ω
, (20)
where ⋅ denotes the cardinality of a set and
In Section II, we assume that the probability estimates Ω?
follow a Dirichlet distribution. This assumption implies that
each element
distribution [22], which is warped to a uniform distribution via
rank normalization. The normalization stretches the high
density areas of the feature space and shrinks it in areas of low
density. The warping functions are shown in Fig. 2 for
individual dimensions and their ensemble average.
Comparing Fig. 2 to Fig. 1, it can be seen that the warping
function closely resembles the squareroot curve in the sense
that the input axis is shrunk for values closer to the origin and
stretched at the other end. However, there is an important
difference regarding computational complexity. Since rank
normalization is nonparametric, the background sets
to be stored and therefore computation of (20) is far more
expensive than the squareroot operation. Recall that the
Bhattacharyya coefficient has a dynamic range bounded
between zero and unity. This is unachievable with rank
normalization, where the inner product of the supervectors in
(19) generally results in unpredictable dynamic range. Similar
problem happens for the Fisher kernel and TFLLR scaling.
This has profound impact on the performance of the SVM, as
shown in the next section.
iBR
=
.
i ω ? , when treated independently, follows a beta
iB have
V. EXPERIMENTS AND RESULTS
A. Experimental Setup
The experiments were carried out on the NIST 2006
speaker recognition evaluation (SRE) task [23]. The core task
consists of 810 target speakers, each enrolled with one side of
a fiveminute conversation, which roughly contains two
minutes of speech. There are 3616 genuine and 47452
imposter trials, where test utterances are scored against target
speakers of the same gender. All speech utterances were first
preprocessed to remove silence and converted into sequences
of 36dimensional feature vectors, each consisting of 12 Mel
frequency cepstral coefficients (MFCCs) appended with deltas
and double deltas. Relative spectral (RASTA) filtering [24]
and utterancelevel mean and variance normalization were
performed. We use two wellknown metrics in evaluating the
performance of the speaker verification systems – equal error
rate (EER) and the minimum detection cost function
(MinDCF) [23]. The EER corresponds to the decision that
gives equal false acceptance rate (FAR) and false rejection
rate (FRR). The MinDCF is defined as the minimum value of
the function 0.1×FRR + 0.99×FAR.
The speaker verification system was designed to be gender
dependent1. Two genderdependent UBMs were trained using
data drawn from the NIST 2004 dataset. The same dataset was
used to form the background data for SVM training. The
commonly available libSVM toolkit [25] was used for this
purpose. HTK toolkit [26] was used for training the UBMs.
B. ML vs. MAP Estimation
This section investigates the difference between ML and
MAP estimation and the influence of the parameter τ on the
system performance. We increased the value of τ in (10) from
0 to 1.0 with a step size of 0.1. Recall that setting
to the ML estimate in (6). Similar procedure was repeated for
various sizes of UBM from 128 to 4096. In all the
1 Gender information is provided and there is no crossgender trial in NIST
SREs. Genderdependent systems have shown better result than gender
independent systems in past evaluations.
0
τ =
leads
128256 512
Size of the UBM, M
(a)
1024 20484096
7
8
9
10
11
12
13
14
15
Equal Error Rate (EER)
ML
τ = 0.1
τ = 0.2
τ = 0.3
τ = 0.4
τ = 0.5
τ = 0.6
τ = 0.7
τ = 0.8
τ = 0.9
τ = 1.0
0.11.0 10.0100.0
10
12
14
16
18
MAP controlled parameter, τ
(b)
Equal Error Rate (EER)
Fig. 3. Performance comparison between ML and MAP in terms of EER. (a)
The value of τ was increased from 0 to 1.0 with a step size of 0.1 for UBM
with various sizes. Letting τ = 0 leads to the ML estimation. The curves for τ
= 0.7 and τ = 0.8, which are farther apart from the ML curve, are highlighted
in red. (b) The EER was evaluated as a function of τ with value increases
from 0.1 to 100. Significant degradation in EER can be observed for τ greater
than 10.
Page 6
Accepted for IEEE TASL
6
experiments Bhattacharyya measure was used for normalizing
the supervectors and tnorm [27] was performed at the score
level (see Section V.D for more details about tnorm). The
results are presented in Fig. 3(a). It can be seen that, for
0.1,0.2,,1.0
τ =
…
, MAP estimation gives lower EER as
compared to the ML estimation of discrete probabilities. This
observation is consistent across different UBM sizes. The
empirical results in Fig. 3(a) also show that the optimum value
of τ varies for different UBM size M. For instance, the
optimum τ for M = 2048 is 0.8, but the value changes to 0.7
for larger UBM with M = 4096. To further investigate the
influence of τ on MAP estimation, we gradually increased the
value from 0.1 to 100. Fig. 3(b) shows the EER as a function
of τ. Notice that we took the ensemble average over different
UBM sizes in order to smooth out the empirical noise from
individual curves. It can be observed that EER increases
drastically for τ greater than 10. Larger value of τ pushes the
MAP estimate toward the prior weights. This weakens the
effect of the observed statistics, which contain speaker
characteristics. In particular, using a very large τ in (10) and
(5) would cause the MAP estimation to give the prior weights
as the probabilities estimate, and thus losing all speaker
related information. In general, the parameter τ has to be
optimized empirically for a given set of data conditions
(duration, signaltonoise ratio, etc.).
The size of the UBM (i.e., the cardinality of the discrete
event set) has a great impact on the performance as shown in
Fig. 3(a). It can be seen that the EER reduces as the size of the
UBM increases. Similar trend can be observed for the ML and
MAP with different values of τ. These results motivate us to
use larger UBM, and therefore larger event set, for discrete
probabilities modeling in the next and subsequent sections.
C. Computational Speed Up: Gaussian Selection
For a large UBM (large as compared to the dimensionality
of the feature vectors) an input vector will be close only to a
few Gaussian components. We can therefore compute the
probabilities of a small subset of components located in the
vicinity of the input vector; the remaining components are
assumed to have zero probability. Gaussian selection
technique [28] as described below can then be used to speed
up the probability computation in (8).
A smaller GMM, referred to as the hash model [28], is
trained with the same training data as the UBM. Mahalanobis
distance is then computed for each pair of Gaussian
components of the UBM and the hash model. Since the
Gaussian components may have different covariance matrices,
their average is used in computing the Mahalanobis distance.
A shortlist is then generated for each component of the hash
model. The shortlist contains indices of those components of
the UBM having the closest distance to a particular
component of the hash model. For a given input vector, we
first determine the top scoring component in the hash model.
The probabilities of the components in the shortlist of the top
scoring component are then computed.
Let H be the size of the hash model and Q be the length of
the shortlists. Gaussian selection results in M/(H + Q) times
faster computation, where M is the order of the UBM. Table I
shows the average accuracy of the Gaussian selection
technique for different sizes of UBM. We evaluate the
accuracy by comparing the probability estimates obtained with
and without Gaussian selection in terms of the Bhattacharyya
measure (16) averaged over 100 random samples, which were
selected from our development data. Formally,
,,
11
1
K
ˆ
Average accuracy
KM
i k i k
ki
ω
?
ω
==
⎛
⎜
⎝
⎞
⎟
⎠
=
∑ ∑
, (21)
where {
probability estimates obtained with and without Gaussian
selection for the kth utterance, respectively, and
the number of utterances. We fixed the ratio M/H to be 32,
and Q = 4H, which results in 6.4 times faster computation.
Also, we used
10
ν − = in the experiment to mitigate the
influence of prior on the result. Considerably high accuracy
()
99%
≈
is achieved with Gaussian selection technique for all
cases listed in Table I.
}
1, 2,,
,,,
kk M k
ω
?
ω
?
ω
?
…
and {}
1,2,,
ˆˆˆ
,,
kk M k
ωωω
…
are the
100
K =
is
D. Nuisance Attribute Projection and SVM Training
Since a supervector represents a speech utterance as a
single point in the vector space, it becomes possible to remove
unwanted variability, due to different handsets, channels and
phonetic content, from the supervector by linear projection.
Let E be an MbyN matrix representing the unwanted
subspace that causes the variability. Nuisance attribute
projection (NAP) [29] removes the unwanted variability from
a supervector via a projection to the subspace complementary
to E, as follows
(
p I EE
)
T
′′′
=−
p . (22)
NAP assumes that the variability is confined in a relatively
low dimensional subspace such that N ≪ M. The columns of E
are the eigenvectors of the withinspeaker covariance matrix
TABLE I
ACCURACY
OVER 100 SAMPLES FOR DIFFERENT SIZES OF UBM AND HASH MODEL. THE
COMPUTATION SPEEDUP IS FIXED AT ×6.4 FOR ALL COMBINATIONS.
100%
ρ ×
OF GAUSSIAN SELECTION TECHNIQUE AVERAGED
Size of the
UBM, M
Size of the hash
model, H
Average accuracy (×100%)
Female
98.59
99.19
99.48
99.62
99.68
Male
98.19
99.13
99.46
99.58
99.68
1024
2048
4096
8192
16384
32
64
128
256
512
TABLE II
COMPARISON OF EER AND MINDCF FOR SUPERVECTORS WITH AND
WITHOUT CHANNEL COMPENSATION AND SCORE NORMALIZATION.
Supervector
Raw, p
+ NAP
+ NAP + tnorm
Bhattacharyya,
+ NAP
+ NAP + tnorm
% EER
8.57
7.05
6.60
7.25
5.53
4.98
MinDCF (×100)
3.90
3.31
3.15
3.42
2.73
2.46
BHAT
′ p
Page 7
Accepted for IEEE TASL
7
estimated from a development dataset with a large number of
speakers, each having several training sessions.
In (22),
′ p denotes the supervectors that have been
normalized with any of the methods mentioned listed in
Sections III and IV. SVM modeling is then performed in the
supervector space that has been properly scaled or warped and
compensated for session variability. The discriminant function
of an SVM [19] can be expressed in terms of the supervector
as follows
()
( )
p
T
1
L
lll
l
fy
αβ
=
′′′′ ′′
=+
∑
pp
, (23)
where L is the number of support vectors,
assigned to the lth support vector with its label given by
{}
1, 1 and β is the bias parameter.
Table II shows the results with and without feature
normalization, channel compensation and score normalization.
Notably, 15.40% relative improvement in EER and 12.31%
relative improvement in MinDCF are obtained by applying the
Bhattacharyya measure on the raw discrete probabilities.
Further improvement (24.55% in EER and 21.90% in
MinDCF) is obtained with NAP and tnorm which
compensate for session variability at the model and score
levels, respectively. Feature normalization is essential for
effective SVM modeling. The reason is that SVMs are not
invariant to linear transformations, i.e., any form of scaling
would cause some of the dimensions to dominate the overall
decision.
For the experiments in Table II, the UBM has a model size
of
16384
M =
, while the hash model for Gaussian selection
has model size of 512. For the MAP estimation, the parameter
τ in (10) is set to 0.1. For the NAP, the projection matrix has
a rank of 60 and was derived from NIST 2004 and 2005 SRE
datasets. For the score normalization [27], tnorm cohorts
were selected from NIST 2005 SRE dataset. We use the same
configuration for subsequent experiments. The overall process
from supervector construction to SVM training is summarized
and illustrated in Fig. 4. Also included in the figure are
references to equations used at each stage.
l α are the weights
ly ∈ − +
E. Comparison of Normalization Methods
We compare the performance of the Bhattacharyya
measure, Fisher kernel, TFLLR
normalization using the same configuration as mentioned
scaling and rank
above in Table III (see the upper panel). Fig. 5 shows the
detection error tradeoff (DET) curves. It can be seen that the
Fisher kernel and TFLLR scaling perform better than just
using the raw discrete probabilities, which indicates that
kernel normalization is important. Comparing these results to
the Bhattacharyya measure, on the other hand, shows that the
squareroot operator is more appropriate than constant scaling
in the Fisher kernel and TFLLR scaling. The Bhattacharyya
measure performs consistently better than the rank
normalization in terms of EER and MinDCF. The
effectiveness of the rank normalization depends on the extent
the supervectors matches the background distribution.
It is also possible to use bigram (i.e., subsequences of two
Gaussian indexes) probabilities to construct supervectors and
to compare the performance of various normalization
methods. For a UBM of size M′, bigram probability
modeling leads to a set of M
=
128
M′ =
be the size of the UBM, the supervector of bigram
probabilities will have a dimensionality of
lower panel of Table III shows the performance of bigram
supervector using different normalization methods. We used
exactly the same training data and parameter settings for all
the experiments in Table III. It can be seen that the bigram
supervector gives poorer accuracy compared with the unigram
supervector. This is likely due to the fact that we have
significantly reduced the size of the UBM to
the resulting bigram supervector has the same dimensionality
as the unigram supervector. For most lowlevel acoustic
MM
′′
×
discrete events. Let
16384
M =
. The
128
M′ =
so that
Event
Extraction and
Count Accumulation
SVM
Training
[Eq. (23)]
Nuisance Attribute
Projection (NAP)
[Eq. (22)]
MAP estimation
[Eq. (1), (5)]
UBM [Eq. (7), (8), (9)],
phone recognizer,
prosodic feature
extractor, etc.
()
in X
X
BHAT
′ p
NAP matrix, E
Discrete probabilities are
estimated [Eq. (5), (10)]
and stacked to form a
supervector [Eq. (11)]
p
BHAT
′′
p
Background
corpus
Background
corpus
Bhattacharyya
measure
[Eq. (17)]
Speaker
model
Number of occurrences
of discrete events
Speech utterance Raw supervector
Normalized supervector
Normalized and compensated
supervector
Fig. 4. The speech utterance, X, is mapped to a supervector of discrete probabilities p, normalized in accordance with the Bhattacharya measure, and
channel compensated prior to SVM modeling. Similar mapping operation is performed on the training utterance of the target speaker, all the utterances
in the background corpus (as indicated by the dotted lines), and test utterances (not shown in the figure).
TABLE III
COMPARISON OF EER AND MINDCF FOR DIFFERENT NORMALIZATION
METHODS AND SUPERVECTORS. NAP AND TNORM WERE APPLIED USING
EXACTLY THE SAME DATASET.
Unigram supervector
Raw
Fisher kernel
TFLLR scaling
Rank normalization
Bhattacharyya measure
% EER
6.60
6.47
6.25
5.23
4.98
MinDCF (×100)
3.15
3.11
3.07
2.63
2.46
Bigram supervector
Raw
Fisher kernel
TFLLR scaling
Rank normalization
Bhattacharyya measure
% EER
15.49

13.66
13.77
11.99
MinDCF (×100)
5.96

5.85
5.68
4.80
Page 8
Accepted for IEEE TASL
8
quantizers, e.g., UBM and VQ codebook, the event set that
contains only the unigrams can be made sufficiently large.
This is different from highlevel events, e.g., phones [10] and
prosodic features [11], which usually rely on bigram or
trigram to form a larger event set.
It is worth noting that bigram supervector gives better result
when the same UBM is used for deriving the unigram
supervector. This can be seen from Fig. 3(a), where the EER
is around 14.5% for unigram supervector with M = 128,
compared to 11.99% of EER (in the last row of Table III) for
bigram supervector with M = 128×128. Clearly, bigram
probabilities are useful but not as effective as simply
increasing the UBM size to obtain unigram supervector
having the same dimensionality.
Regarding the normalization of the bigram supervectors,
our conclusion is clear – Bhattacharyya measure performs
consistently better than other normalization methods for both
unigram and bigram supervectors. A McNemar’s statistical
test [14, 30] was conducted to see if the EER and MinDCF of
the Bhattacharyya measure are significantly better than other
normalization methods. The pvalues obtained were all less
than 0.05, which means that the improvements are significant
with a confidence level of 95%. Notice that we do not provide
results for Fisher kernel in the lower panel of Table III as the
kernel is not readily applicable to bigram probabilities.
F. Comparison and Fusion of Supervectors
Finally, we evaluate the performance of the supervector of
discrete probabilities (with the Bhattacharyya measure) in
comparison with the GLDS kernel [3] and GMM supervector
[4]. For the GLDS kernel, we used all monomials up to the
third order. The resulting supervectors have a dimensionality
of 9139. For the GMM supervector, the UBM consists of 512
mixtures leading to supervectors of dimensionality 18432.
Recall that the supervector of discrete probabilities has a
comparable dimensionality of
M =
16384
. The datasets used
for UBM training, SVM background data, NAP and tnorm
are the same for all systems.
Table IV shows the EER and MinDCF. Fig. 6 shows the
DET curves. The Bhattacharyya system exhibits competitive
performance compared to the other two systems, with the
GMM supervector being the best. We fused the Bhattacharyya
system with the other two at the score level using equal
weights summation. The fusion with the GLDS gives relative
improvement of 15.06% in EER and 8.13% in MinDCF over
the best single system. Some improvement can also be
observed for the fusion with the GMM supervector, which
amounts to 5.35% and 0.95% relative reduction in EER and
MinDCF, respectively, over the best single system. A
McNemar statistical test [14, 30] was conducted to see if the
TABLE IV
COMPARISON OF EER AND MINDCF FOR DIFFERENT SUPERVECTORS. NAP
AND TNORM WERE APPLIED USING EXACTLY THE SAME DATASET. THE
FUSION RESULTS WERE OBTAINED VIA LINEAR COMBINATION WITH EQUAL
WEIGHTS AT THE SCORE LEVEL.
Supervector
Bhattacharyya (proposed)
GLDS
GMM supervector
% EER
4.98
5.39
4.30
MinDCF (×100)
2.46
2.67
2.10
Bhattacharyya + GLDS
Bhattacharyya + GMM
supervector
4.23 2.26
4.07 2.08
TABLE V
values
OF THE FUSION COMPARED TO THE SINGLE BEST SYSTEM FOR OPERATING
POINTS AT EER AND MINDCF. A value p
ARE OBSERVING A SIGNIFICANT DIFFERENCE IN THE PERFORMANCE AT A
CONFIDENCE LEVEL OF 95%.
p
OF MCNEMAR’S TESTS ON THE DIFFERENCES IN THE PERFORMANCE
LESS THAN 0.05 MEANS THAT WE
Operating point:
Bhattacharyya + GLDS
Bhattacharyya + GMM
supervector
The best single system
EER
2.2 10−
×
1.8 10−
×
MinDCF
1.5 10−
×
6.8 10−
×
16
2
4
1
0.2 0.5 1 2 5 10 20 40
0.2
0.5
1
2
5
10
20
40
False acceptance rate (%)
False rejection rate (%)
GLDS
Bhat
GMM SV
GLDS + BHAT
GMM SV + BHAT
Fig. 6. DET curves showing a comparison of the performance of the GLDS
kernel, GMM supervector, and the fusion with the proposed Bhattacharyya
system.
0.2 0.5 1 2 5 10 20 40
0.2
0.5
1
2
5
10
20
40
False acceptance rate (%)
False rejection rate (%)
Raw
Fisher
TFLLR
Rank
BHAT
Fig. 5. DET curves showing a comparison of various normalization methods
(or kernels) on the supervector of discrete probabilities.
Page 9
Accepted for IEEE TASL
9
differences in EER and MinDCF are significant. The pvalues
are shown in Table V. Clearly, the improvement in EER is
significant at a confidence level of 95% for both fusions since
the pvalues are less than 0.05. The improvement in MinDCF
for the first fusion is also significant with 95% confidence,
which, however does not hold for the second fusion.
The fusion of Bhattacharyya and GMM supervector is less
successful because the same datasets were used for UBM
training, SVM background data, NAP and tnorm. This was
purposely done so as to have controlled comparisons. We
would like to emphasize that, even though we use the UBM as
the quantizer in this paper, the occurrence counts of the
discrete events could come from a phone recognizer, a
prosodic feature extractor or a speech recognition system as
shown in Fig. 4. We anticipate that, had we chosen such a
completely different frontend, we would likely observe
higher fusion gain. This is a point for future research.
VI. CONCLUSIONS
Speech signals can be represented in terms of the
probability distribution of acoustic, idiolect, phonotactic or
some highlevel discrete events. Formulated under the
maximum a posteriori (MAP) estimation framework, we have
demonstrated the usefulness of modeling speech signals as
discrete distributions for SVMbased speaker verification. We
further proposed and analyzed the use of Bhattacharyya
coefficient as the similarity measure between supervectors
constructed from the discrete probabilities. From the
perspective of feature normalization in the supervector space,
the Bhattacharyya measure warps the distribution of each
dimension with a squareroot function, a much simpler and
dataindependent operation, yet leading to higher accuracy
compared to the Fisher kernel, TFLLR scaling, and rank
normalization. Experiments conducted on the NIST 2006 SRE
showed that relative reduction in EER was 15.40 % with the
Bhattacharyya measure and 24.55 % when used in
conjunction with NAP and tnorm. These results suggest that
the Bhattacharyya measure is a strong candidate for measuring
the similarity between discrete distributions with SVM
classifier. The proposed
performance to the stateoftheart GMM supervector
approach. Their fusion gave 5.35% relative improvement in
EER, even though the improvement in MinDCF was marginal.
It is worth emphasizing that, even though the current work
uses a UBM quantizer to construct supervectors, this is not
necessarily the case; the proposed method can be used with
other types of frontend quantizer. In future, it would be
interesting to compare how much we would benefit by using
the proposed method with a different frontend quantizer such
as a phone recognizer or a prosodic feature extractor. We also
expect the method to be readily applicable for spoken
language recognition, and applications beyond speech
technology that operate on discrete symbols, such as natural
language processing (NLP) and bioinformatics.
method gives comparable
REFERENCES
[1] T. Kinnunen and H. Li, “An overview of textindependent speaker
recognition: from features to supervectors,” Speech Communication, vol.
52, no. 1, pp. 1240, Jan. 2010.
[2] V. Wan and S. Renals, “Speaker verification using sequence
discriminant support vector machines,” IEEE Trans. Speech Audio
Process., vol. 13, no. 2, pp. 203210, Mar. 2005.
[3] W. M. Campbell, J. P. Campbell, D. A. Reynolds, E. Singer, and P. A.
TorresCarrasquillo, “Support vector machines for speaker and language
recognition,” Computer Speech and Language, vol. 20, no. 23, pp. 210
229, 2006.
[4] W. M. Campbell, D. E. Sturim, and D. A. Reynolds, “Support vector
machines using GMM supervectors for speaker recognition,” IEEE
Signal Processing Lett., vol. 13, no. 5, pp. 308311, May 2006.
[5] Y. Mami and D. Charlet, “Speaker recognition by location in the space
of reference speakers,” Speech communication, vol. 48, no. 2, pp. 127
141, 2006.
[6] A. Stolcke, S. S. Kajarekar, L. Ferrer, and E. Shriberg, “Speaker
recognition with session variability normalization based on MLLR
adaptation transforms,” IEEE Trans. Audio, Speech, and Lang. Process.,
vol. 15, pp. 19871998, Sep. 2007.
[7] K. A. Lee, C. You, H. Li, and T. Kinnunen, “A GMMbased
probabilistic sequence kernel for speaker recognition,” in Proc.
INTERSPEECH, pp. 294297, 2007.
[8] K. A. Lee, C. You, H. Li, T. Kinnunen, and D. Zhu, “Characterizing
speech utterances for speaker verification with sequence kernel SVM,”
in Proc. INTERSPEECH, pp. 13971400, Sep. 2008.
[9] P. Kenny, M. Mihoubi, and P. Dumouchel, “New MAP estimates for
speaker recognition,” in Proc. EUROSPEECH, pp. 29642967, 2003.
[10] W. M. Campbell, J. P. Campbell, T. P. Gleason, D. A. Reynolds, and W.
Shen, “Speaker verification using support vector machines and high
level features,” IEEE Trans. Audio, Speech, and Language Process.,
vol. 15, no. 7, pp. 20852094, Sep. 2007.
[11] E. Shriberg, L. Ferrer, S. Kajarekar, A. Venkataraman and A. Stolcke,
“Modeling prosodic feature sequences for speaker recognition,” Speech
Communication, vol. 46, no. 34, pp. 455472, Jul. 2005.
[12] C. H. Lee and Q. Huo, “On adaptive decision rules and decision
parameter adaptation for sutomatic speech recognition,” Proceedings of
the IEEE, vol. 88, no. 8, pp. 12411269, Aug. 2000.
[13] T. Kailath, “The divergence and Bhattacharyya distance measures in
signal detection,” IEEE Trans. Commun. Technol., vol. COM15, no. 1,
pp. 52 – 60, Feb. 1967.
[14] X. Huang, A. Acero, and H. W. Hon, Spoken Language Processing: a
Guide to Theory, Algorithm, and System Development. NJ: Prentice
Hall, 2001.
[15] A. Stolcke, S. Kajarekar, and L. Ferrer, “Nonparametric feature
normalization for SVMbased speaker verification,” in Proc. ICASSP,
pp. 15771580, 2008.
[16] M. A. Zissman, “Comparison of four approaches to automatic language
identification of telephone speech,” IEEE Trans. Speech Audio Process.,
vol. 4, no. 1, pp. 3144, Jan. 1996.
[17] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, “Speaker verification
using adapted Gaussian mixture models,” Digital Signal Processing, vol.
10, no. 13, pp. 1941, 2000.
[18] T. S. Jaakkola and D. Haussler, “Exploiting generative models in
discriminative classifisers,” in Advances in Nueral Information
Processing Systems 11, M. S. Kearns, S. A. Solla, and D. A. Cohn, Eds.
Cambridge, U.K.: MIT Press, 1998.
[19] V. Kecman, Learning and Soft Computing: Support Vector Machines,
Neural Networks, and Fuzzy Logic Models. MA: MIT Press, 2001.
[20] Á. de la Torre, A. M. Peinado, J. C. Segura, J. L. PérezCérdoba, M. C.
Benítez, and A. J. Rubio, “Histogram equalization of speech
representation for robust speech recognition,” IEEE Trans. Speech Audio
Process., vol. 13, no. 3, pp. 355366, May. 2005.
[21] J. Pelecanos and S. Sridharan, “Feature warping for robust speaker
verification,” in Proc. Odyssey, 2001.
[22] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, NY:
Wiley, 2001.
[23] The NIST Year 2006 Speaker Recognition Evaluation Plan, National
Institute of Standards and Technology, Mar. 2006.
[24] H. Hermansky and N. Morgan, “RASTA processing of speech,” IEEE
Trans. Speech Audio Process., vol. 2, no. 4, pp. 578589, Oct. 1994.
Page 10
Accepted for IEEE TASL
10
[25] C. C. Chang and C. J. Lin, LIBSVM: a library for support vector
machines, 2001. Software available at http://www.csie.ntu.edu.tw/
~cjlin/libsvm
[26] S. Young, G. Evermann, M. J. F. Gales, T. Hain, D. Kershaw, X. Liu, G.
Moore, J. Odell, D. Ollason, D. Povey, V. Valtchev, and P. Woodland,
The HTK Book, version 3.4, Cambridge, UK, 2006.
[27] R. Auckenthaler, M. Carey, and H. LloydThomas, “Scorenormalization
for textindependent speaker verification,” Digital Signal Process., vol.
10, pp. 4254, Jan. 2000.
[28] R. Auckenthaler and J. S. Mason, “Gaussian selection applied to text
independent speaker verification,” in Proc. Odyssey, 2001.
[29] A. Solomonoff, W. M. Campbell, and C. Quillen, “Channel
compensation for SVM speaker recognition,” in Proc. Odyssey, pp. 57
62, 2004.
[30] L. Gillick and S. J. Cox, “Some statistical issues in the comparison of
speech recognition algorithms,” in Proc. ICASSP, pp. 532535, 1989.
Kong Aik Lee received the B.Eng. (first class
honors) degree from University Technology
Malaysia in 1999, and the Ph.D. degree from
Nanyang Technological University, Singapore, in
2006.
He is currently a Senior Research Fellow with
Human Language
Institute for Infocomm Research (I2R), Singapore.
His research focuses on statistical methods for
speaker and spoken language recognition, adaptive
echo and noise control, and subband adaptive
filtering. He is the leading author of the book:
Subband Adaptive Filtering: Theory and Implementation (Wiley, 2009).
Chang Huai You received the B.Sc. degree in
physics and wireless from Xiamen University,
China, in
communication and electronics engineering from
Shanghai University of Science and Technology,
China, in 1989, and the Ph.D. degree in electrical
and electronic
Technological University (NTU), Singapore, in
2006.
From 1989 to 1992, he was an Engineer with
Fujian Hitachi Television Corporation, Ltd.,
Fuzhou City, China. From 1992 to 1998, he was an Engineering Specialist
with Seagate International, Singapore. He joined the Centre for Signal
Processing, NTU, as a Research Engineer in 1998 and became a Senior
Research Engineer in 2001. In 2002, he joined the Agency for Science,
Technology, and Research, Singapore, and was appointed as a Member of
Associate Research Staff. In 2003, he was appointed as a Scientist with the
Institute for Infocomm Research (I2R), Singapore. In 2006, he was appointed
as a Senior Research Fellow with I2R. Since 2007, he has been a Research
Scientist with Human Language Technology department of I2R. His research
interests include speaker recognition, language recognition, speech
enhancement, speech recognition, acoustic noise reduction, array signal
processing, audio signal processing, and image processing. He is a reviewer of
many international conferences and journals.
Dr. You was the recipient of Silver Prize of EEE Technology Exhibition at
NTU for his “Intelligent Karaoke Project” as a major designer and project
leader in 2001.
Haizhou Li (M’91–SM’01) is currently the
Principal Scientist and Department Head of Human
Language Technology at the Institute for
Infocomm Research. He is also the Program
Manager of Social Robotics at the Science and
Engineering Research Council of A*Star in
Singapore.
Dr Li has worked on speech and language
technology in academia and industry since 1988.
He taught in the University of Hong Kong (1988
1990), South China University of Technology
(19901994),
Technology department,
1986, the M.Eng. degree in
engineering from Nanyang
and Nanyang Technological
University (2006present). He was a Visiting Professor at CRIN in France
(19941995), and at the University of New South Wales in Australia (2008).
As a technologist, he was appointed as Research Manager in AppleISS
Research Centre (19961998), Research Director in Lernout & Hauspie Asia
Pacific (19992001), and Vice President in InfoTalk Corp. Ltd (20012003).
Dr Li's current research interests include automatic speech recognition,
speaker and language recognition, and natural language processing. He has
published over 150 technical papers in international journals and conferences.
He holds five international patents. Dr Li now serves as an Associate Editor of
IEEE Transactions on Audio, Speech and Language Processing, and Springer
International Journal of Social Robotics. He is an elected Board Member of
the International Speech Communication Association (ISCA, 20092013), a
Vice President of the Chinese and Oriental Language Information Processing
Society (COLIPS, 20092011), an Executive Board Member of the Asian
Federation of Natural Language Processing (AFNLP, 20062010), and a
Member of the ACL.
Dr Li was a recipient of National Infocomm Award of Singapore in 2001.
He was named one the two Nokia Visiting Professors 2009 by Nokia
Foundation in recognition of his contribution to speaker and language
recognition technologies.
Tomi Kinnunen received the M.Sc. and Ph.D.
degrees in computer science from the University
of Joensuu, Finland, in 1999 and 2005,
respectively.
He worked as an associate scientist at the
Speech and Dialogue Processing Lab of the
Institute
Singapore and as a senior assistant at the
Department of Computer Science and Statistics,
University of Joensuu, Finland. He works
currently as a postdoctoral researcher in the
University of Eastern Finland, Joensuu, Finland, and his research is funded by
the Academy of Finland. His research areas cover speaker recognition and
speech signal processing.
Khe Chai Sim received the B.A. and M.Eng.
degrees in Electrical and Information Sciences
from the University of Cambridge, England in
2001. He then received the M.Phil degree in
Computer Speech, Text and Internet Technology
from the same university in 2002 before joining
the Machine Intelligence laboratory, Cambridge
University Engineering Department in the same
year as a research student.
Upon completion of his Ph.D. degree in 2006,
he joined the Institute for Infocomm Research,
Singapore as a research engineer. Since 2008, he is the Assistant Professor at
the School of Computing, National University of Singapore, Singapore. His
research interests include statistical pattern classification, automatic speech
recognition, speaker recognition and spoken language recognition. He has also
worked on the DARPA funded EARS project from 2002 to 2005 and the
GALE project from 2005 to 2006.
for Infocomm Research (I2R),
View other sources
Hide other sources
 Available from Kong Aik Lee · May 28, 2014
 Available from joensuu.fi
 Available from psu.edu