0:00:16everybody i in a single button so that a new feature for automatic speaker verification
0:00:26the data for us to show that automatic specific speaker verification system
0:00:33are all the spoofing
0:00:38especially when no we are in an ideal scenario where the spoofing we cannot be
0:00:44you know in advance so the need the all the for the generalize the going
0:00:50to measure
0:00:51so we propose an feature example sample transform
0:00:56and the t v is a
0:00:59task automation
0:01:02and different from a for their trust one because we suppose that fourier transform me
0:01:06like frequency resolution secrecy security uses our body a time frequency the solution that means
0:01:14i your time resolution for i frequencies and the higher frequency resolution for lower frequencies
0:01:22we combine the c d more or less with the traditional a cepstral analysis but
0:01:30we found the problem but using the discrete cosine transform or you know the identity
0:01:37applied before so
0:01:39for two reasons results secrecy a dct have different skate one dramatic another one is
0:01:46the dinner and dramatically dct basis are no longer than one
0:01:50so we found a total shock joe's the only for me to sing but they're
0:01:54not but for thirty twenty scale over the speaker i
0:01:58with a linear frequency scale
0:02:02these is the some form comparison of a results a nice peaceful for database and
0:02:08we know that
0:02:13we found that no i x from a tax or the system and earlier are
0:02:18excellent error rate for unknown i so the thing to me
0:02:25best performance especially for a s then that young federation seem to use these your
0:02:32and where a we obtain a seven percent over relative improvement
0:02:38embodied the exposing detection performance
0:02:43a part of to be in our work seeks
0:02:47we that seventy two relative improvement
0:02:49so for more details are wait for you
0:02:54sure some possible