a good morning
i'm gonna present
in your approach or relatively new approach to extract i-vectors
the idea is to extract phonetically compensated
i-vectors
by using my scenery that is quite close to the g a the jfa must
ignore e
so
it's
the nn will be involved a slightly differently than just under two or three ways
meaning of a bottleneck approach
or the use the use of the d n n's
instead of ubm
in this case we gonna use the ubm
and
we consider remote these model as a probabilistic extension to the subspace
gmm
and the core of the idea is to treat the phonetic variability a side and
nuisance variability
and
we assume for it to do that we assume that at each frame
a week this super vector that corresponds to each frame to each observation can be
decomposed
into an i-vector by corresponds to the combination of speaker and channel
and
plus
and which an analysis variability that captures
the phonetic variability
k and that's where the d n and games
can't and that provides an extra supervision
okay so my telling us which seen on is probably
i which corresponds to this particular frame
probabilistically of course
we form of the variational bayes algorithm to train the model
and
practically estimated to subspaces
and
as it all you the nn provides these extra supervision
which differentiated from the channel factors that we have enough jfa model so please come
to the possible to discuss the