so we are presenting here our work using what we can have a four gram
units
using require program known and of course for language identification
so
where you know what we it doing with a regular and recurrent neural network is
to use phonemes as input to see now when i think with indication
begin and that of the number of phonemes
and then we have also incorporated in the context information use in a uniform slide
function trigrams
comparing them and their fusion all of them
and so we are proposing the concatenation of this in a descent phonemes in our
in our system
so this architecture apply to this language and the via some system
is based on phonotactic system i prepare landmark detection
so we have for its phonetic recognisers in the bruno recognizers we obtain a sequence
of phonemes
and in evaluations for each utterance we a compute like an entropy metric provided by
their like the network
and this entropy scores are calibrated than used later
we also present a word but don't with funded hubert representations a it used in
order to reduce the vocabulary in this a neural network using k-means to group a
similar from grants
and we have were with this keeper model at the phoneme level and we had
a relative improvement of seven percent
hence the despicable to read the text
also in the work we present like the study of the most role of and
their right brown report in that no one of course parameters
so here is the list of parameters have been working with
here in their results we can see this you have rate in our database used
in comparing the nice ones the diphones triphones and then we can see a the
fusion of them and the comparison with the work these landed pprlm
and a fusion with that and the standard acoustic system based on mfccs
and
c different portions finally where we can see that a
this approach also provides complementary information so there are a final improvements in our global
system
that's it