first i will give a quick overview of i-vectors
after that i will
only some of the methods for hand recounts and start the
of the i-vector eyes them estimate scores by
limited the
duration of recordings
then i will
describe a simple preprocessing weighting scheme which uses duration information as a measure of
i wrecked or a oral ability
then i will describe some experiments and the results
followed by concluding remarks
in theory each decision should be made to dependent on the amount of data available
and the same should hold also in the case of speaker recognition since
we usually have recordings of different lengths
in practice this is usually not the case mainly due to practical reasons since panic
of uncertainty increases the article we agreed to make and computational complexity
and also
the gain in performance
cohen
can be not that could be not so significant especially if the recordings are sufficiently
long
in the case of
i-vector challenge
the
i-vectors were extracted from recordings of different lengths
and to the duration follows log normal distribution this suggests
that
we should see some improvement
if the duration information is taken into account
i-vector is defined as a map point estimate of keeping the variable of factor analysis
model
and it serves as a compact representation of speech utterance
the posterior covariance encodes the answer t
of the i-vector or
estimate
which is caused by a limited to duration of the recordings
usually
the i sort the
is discarded to and comparing i-vectors for example in the
be lda model
nevertheless there have been proposed some solutions how to the
take the uncertainty into account for example be a day with uncertainty propagation
where and then we should note term is added to the model
which models
which explicitly models the
duration variability
another one
is score calibration using different
duration is a quality measure
and yet another recycle i-vector scaling where the length normalisation is modified this to account
for the
uncertainty of i-vectors
and those solutions are not directly applicable or at least not
easily applicable in the context of i-vector challenge
scenes
the data for we can start
reconstructing the posterior covariance is not available
and also there is no development data that could be used for
optimising the calibration parameters
so is there another possibility how to use duration information
to as a measure of i-vector a
rely reliability
prior to
comparing the i-vectors are usually preprocessed
among more common preprocessing methods are pca lda and do within class covariance normalization
in which the basic step is to calculate mean and
covariance matrix s
we implicitly assume
in those calculations that
each the i-vector is equally all i-vectors are equally reliable
some to account for the difference in a reliability of i-vectors
re
proposed a simple weighting scheme in of each other
in which the to could contribution of each i-vector is multiplied by its corresponding duration
so
to verify the
soundness of the proposed idea
we implemented that the baseline system right in which we compare it
the standard pca with
the weighted version of the pca
the results showed that weighted version of peace
pca
produce slightly better results than a standard one
we also wanted to
try within class covariance normalisation
but
in order to
the apply within class covariance normalization
we need to have labeled to date time which was not the case in the
challenge
so we needed to perform unsupervised the clustering we
but
experiment that with the different clustering algorithms but that the end to be selected k-means
with cosine distance and four thousand clusters
unfortunately the results are worse for within class covariance normalization then for a pca but
at least the
the weighted version was
slightly a cat of the standard one
we tried also several different classifiers and the best results were at used
with a logistic regression but only after reading remove the
length normalisation of from the processing pipeline
in that case within class covariance normalisation
gave better results then pca and all spend the can
weighted towards and was
score the better than standard one
we try to further improve the results by additional fine-tuning
so we put the duration as it and additional feature of i-vectors we excluded clusters
with small official score
we were is the roles of
target and test i-vectors
and do you can't do you want to the hyper parameters of logistic regression
we did this fine tuning we were able to improve
the previous result
for a little bit more
so this was also our
best submitted result
to conclude we
present that
a simple preprocessing weighting scheme which uses do duration information is a measure of i-vector
a reliability
we at you would quite reason the bus six sets
with a clustering in the case of within class covariance normalization
but okay but cat
nearly no success with clustering in the case of the lda
which suggests that we had a is more susceptible for labeling errors
and the last remark we found out that length normalization does not help logistic regression
thank you
okay
just empirical results but maybe somebody s can comment that i don't know
nicole
with on the side
but we with at the same results as logistic regression icons otherwise
did you generate what we did a clustering or you just one clustering stage we
tried
different things also two
to iterate the clustering but didn't six it
this was also experiments clear sets of four thousand because we didn't the get then
you improvements by
changing