good afternoon that would be present in the were carried out that the at the
university of mister
and i present in the age related voice disguise anything about speaker verification and as
we know when we ask that in the performance of a
of automatic speaker verification system we want to assess
inter speaker variability and we wanted to more
with h really that these guys and for that we define it that's the intentional
modification of the speaker voice sounded like the younger one and all the first and
so forth these we believe that the corpus with six stick speakers all native finnish
in that's to the environment and the speech is read text in finnish and in
english
and for these we got twenty six segments on each voiced i natural voice for
voicing your voice into different sessions for the speakers at the same time we do
the prior to recording with that too smart phones just to check this channel differences
also
and so we for the performance evaluation we use just a
mfcc feature fifty four dimensional and to a s p system base gmm ubm and
i-vector system with cosine and b lda score and ported that's there are we used
to going for the three conditions basically because we have one not recognition and then
the disguised condition would be different all and
own voice
we got this number of trials for each of these conditions and the channels as
i already mentioned so they training they that all these microphone data from the natural
voice and the corresponding disguised voice was used for testing in the in the disguise
conditions
so this is nick a big of the results and we can see for the
i-vector be lda a system we can see i degradation between the natural voice baseline
of the system and then they these guys all and these guys don't we can
see this degradation in for more details on the experimental setup and
the data please this poster seven