0:00:15a good morning
0:00:17i'm gonna present
0:00:19in your approach or relatively new approach to extract i-vectors
0:00:24the idea is to extract phonetically compensated
0:00:28i-vectors
0:00:30by using my scenery that is quite close to the g a the jfa must
0:00:36ignore e
0:00:37so
0:00:40it's
0:00:40the nn will be involved a slightly differently than just under two or three ways
0:00:48meaning of a bottleneck approach
0:00:50or the use the use of the d n n's
0:00:56instead of ubm
0:00:57in this case we gonna use the ubm
0:00:59and
0:01:01we consider remote these model as a probabilistic extension to the subspace
0:01:07gmm
0:01:09and the core of the idea is to treat the phonetic variability a side and
0:01:15nuisance variability
0:01:17and
0:01:19we assume for it to do that we assume that at each frame
0:01:23a week this super vector that corresponds to each frame to each observation can be
0:01:29decomposed
0:01:30into an i-vector by corresponds to the combination of speaker and channel
0:01:36and
0:01:38plus
0:01:39and which an analysis variability that captures
0:01:43the phonetic variability
0:01:46k and that's where the d n and games
0:01:50can't and that provides an extra supervision
0:01:54okay so my telling us which seen on is probably
0:01:58i which corresponds to this particular frame
0:02:02probabilistically of course
0:02:04we form of the variational bayes algorithm to train the model
0:02:08and
0:02:10practically estimated to subspaces
0:02:14and
0:02:16as it all you the nn provides these extra supervision
0:02:20which differentiated from the channel factors that we have enough jfa model so please come
0:02:27to the possible to discuss the