0:00:15have run some to learn from university of east and women today same here two
0:00:19presents the last it binds
0:00:21in using the planning phone number and its applications
0:00:24so how would you simple we chuckle reflects a hand graph feature in
0:00:30in direct approach
0:00:31by our deep-learning that will
0:00:32so what we use be plantings first
0:00:35you given any feature itinerary in use a added to considering so we try to
0:00:40understand how the effect of convolution on the signal by reconstruct the signal from the
0:00:46art is no messy go very small print a tree trees so signal for via
0:00:51kind of read effects and fall be but
0:00:54become comparison is like glow the signals
0:00:58and yes the
0:01:00subsets in speech recognition this
0:01:02or you modify must in
0:01:05frame within we depending
0:01:07and next week
0:01:10deep-learning in bottleneck feature we improve the performance of language identification system a lot
0:01:16so this is a forty size of i was system
0:01:20is the end-to-end approach from audio file we extract
0:01:24no mfcc or filter bank feature and feeding to the network we custom probabilities for
0:01:29each language
0:01:31so far
0:01:32jennings this deep network
0:01:34well we need to address how overnight in well how to better if it from
0:01:38multiple watching take the design how to change the artistic the
0:01:42efficiently using early stopping regularization
0:01:45am optimization techniques
0:01:47and how to address the computation noise you with deep network
0:01:51and this is the results so we get improvement compared to single system and our
0:01:58republic system that improvement
0:02:00by using more advanced technical and what's normalization dropouts
0:02:05and fall we sees as the imbalanced dataset happen or negative interface i on the
0:02:12model so we use modify the corpus and using by just imposing to be
0:02:17or sampling and score calibration
0:02:19we closer to the bottleneck feature but
0:02:22still plenty of room for improvement