We are pleased to announce that the following papers have been accepted
to The Speaker and Language Recognition Workshop (IEEE - Odyssey 2008 --
** Reda Dehak, Najim Dehak, Patrick Kenny, Pierre Dumouchel. Kernel
Combination for SVM Speaker Verification.
We present a new approach for constructing the kernels used to build
support vector machines for speaker verification. The idea is to
construct new kernels by taking linear combination of many kernels such
as the GLDS and GMM supervector kernels. In this new kernel combination,
the combination weights are speaker dependent rather than universal
weights on score level fusion and there is no need for extra-data to
estimate them. An experiment on the NIST 2006 speaker recognition
evaluation dataset (all trial) was done using three different kernel
functions (GLDS kernel, linear and Gaussian GMM supervector kernels). We
compared our kernel combination to the optimal linear score fusion
obtained using logistic regression. This optimal score fusion was
trained on the same test data. We had an equal error rate of $\simeq
5,9\%$ using the kernel combination technique which is better than the
optimal score fusion system ($\simeq 6,0\%$).
** Reda Dehak, Najim Dehak, Patrick Kenny, Pierre Dumouchel. Comparison
Between Factor Analysis and GMM Support Vector Machines for Speaker
We present a comparison between speaker verification systems based on
factor analysis modeling and support vector machines using GMM
supervectors as features. All systems used the same acoustic features
and they were trained and tested on the same data sets. We test two
types of kernel (one linear, the other non-linear) for the GMM support
vector machines. The results show that factor analysis using speaker
factors gives the best results on the core condition of the NIST 2006
speaker recognition evaluation. The difference is particularly marked on
the English language subset. Fusion of all systems gave an equal error
rate of 4.2% (all trials) and 3.2% (English trials only).
** Patrick Kenny, Najim Dehak, Reda Dehak, Vishwa Gupta, Pierre
Dumouchel. The Role of Speaker Factors in the NIST Extended Data Task.
We tested factor analysis models having various numbers of speaker
factors on the core condition and the extended data condition of the
2006 NIST speaker recognition evaluation. In order to ensure strict
disjointness between training and test sets, the factor analysis models
were trained without using any of the data made available for the 2005
evaluation. The factor analysis training set consisted primarily of
Switchboard data and so was to some degree mismatched with the 2006 test
data (drawn from the Mixer collection). Consequently, our initial
results were not as good as those submitted for the 2006 evaluation.
However we found that we could compensate for this by a simple
modification to our score normalization strategy, namely by using 1000
z-norm utterances in zt-norm.
Our purpose in varying the number of speaker factors was to evaluate the
eigenvoiceMAP and classicalMAP components of the inter-speaker
variability model in factor analysis. We found that on the core
condition (i.e. 2?3 minutes of enrollment data), only the eigenvoice MAP
component plays a useful role. On the other hand, on the extended data
condition (i.e. 15?20 minutes of enrollment data) both the classical MAP
component and the eigenvoice component proved to be useful provided that
the number of speaker factors was limited. Our best result on the
extended data condition (all trials) was an equal error rate of 2.2% and
a detection cost of 0.011.