you have to do um that well yeah i eyepiece entity it's not a so parallel acoustic model adaptation for improving alphabetic language recognition um in general um phonotactic um language recognition system um that you move to complement the first one you start our recognition of one ten in which then maybe what a single phone recogniser or a little thing uh for recognising patterns we wish oh we use it for the uh for that information extraction and the second one you say and classifier that oh use the extracted oh from type information two to distinguish between target language um in politics uh language recognition the idea of feature down i would first implication is that well i don't but example include a using parallel recognise a uniform and and the second yourself using multiple high level in the phone lattice decoding two we use the speaker and um section i used to be but in the uh speech data generally i involving the telephone speech something like that um and now adaptation and speaker adaptive training S A T a parallel to to a phone lattice decoding is used has gone and it must be posted seriously so in this piece of our work um we would like to investigate different types of uh adaptation techniques and we with um quantitatively master that i was working i between two sets of phonotactic features and finally oh we investigate but the hello acoustic model adaptation can provide for the oh feature diversification and in particular the we will work on the mean only mllr the station and the variance on the and a rotation yeah slows down it struck general structure of a all three um food addict a language recognition system that you want a two component that i mentioned before the parallel phone recogniser and also the backend uh in the back and we can use a oh vectors space modelling or at the end where modelling uh you know about in our experiment we use the uh we have to model double curved space modelling i'm sorry the the reason there's some problem on it i don't know the but uh anyway oh it was so so there is a lot um a value school on the if of a yes uh model and then we would like to combine them to get up in the school that and that and in fact they say at here and the F represent different different our phone recogniser and we combine the school and also we have so i have a phone recogniser so we combine the at school and you know our work we were uh uh we should use wall all features are for i diversification using a different uh model adaptation you can see that yeah at a phone recogniser oh and for each formica and so we have to if uh mobile application so for yeah yeah um organiser and maybe is that we use that eight he was two and then they are to have well score from the reassembled and you know experiment where you try to set a the and you go to one that we that means we we use are a single form organised you know experiments but all of this and see we we were uh for the whole experiment we find that uh using the other from a fellow that uh well know that the patient yeah i can still get into when we use a paraffin recognise um to to further up to we use the speaker and the session induced variation oh we use the N R or and the um uh that uh i patient uh in in in the phone lattice decoding um the transformation can be for me data but these two impatient yeah eight B and H is the transform to be computed and the meal and uh signal is the gaussian mean ankle very informative yeah so well the different types of adaptation technique we test by the way we also test the each radius and not the patient and also oh adaptation with multiple regression classes that's how we found that not all of this uh improvement can be found at so we did a report the results in details you know people you know well that was the mobile application to class uh decoding using wap and the post process is first of all we generate a single bad so sequence and then we estimate the transform eighty and all eight and then based on the transformed but i i was the model we generate the the format up a second uh uh who's model adaptation in the test uh test data we cannot fight uh speaker adaptive training in the training data all the of the uh phone recogniser oh and in in which of that feature level and all times well is a pilot to each other um uh training utterance in a uniform recogniser and do we test our experiment oh three types of adaptation technique we have right in the U S N um vector space um although i can um the phone like this is uh on is a commercial two to to expect that and run a and he's expert you are very much um we use that and all that uh and rambled on tree and then it is converted to a high dimensional a remote that features that contains on unigram bigram line trigram forms uh for uh statistic and this the size of this L I dimensional phonotactic feature the the uh that the dimension S is determined by the name brand uh all the and and also the phone set size she after we generate uh the high dimension phonotactic feature we put it into the svm training for the S R O the reassemble inside moreover we also define the diversity pitching to to to set up between two or phonotactic feature oh using at that you uh you could be yeah idea is that um between uh that the the feature C A S E be based on their nonzero uh and bram a statistic and you are but you have to use you uh means that the set of anger and statistic which is nonzero in blue both C N C P and and you use those uh size of the set you our system has been talking about it uh using the thirty second tar in two thousand i snap and this uh language recognition evaluation you michelle fourteen target languages are involved in the detection cost um the system determine whether the target language is spoken in the speech uh huh and at least equal error rate which is calculate the from the eer of each target target language could easily ported oh we use this that i've page uh he are used to ensure that oh is target language has very has an equal contribution to the match on examination people of a single organiser is used you know one and um forty nine uh dimension mfcc feature or standard three state left to right hmm thirty two gaussian components per state is used in all acoustic model um for the training data fifteen hours of uh switchboard one set or the uh english uh data use use that to train do some recogniser and a full um on the phone loop grammar used used in the decoding of all the training data of the target languages we use the close friend ooh so uh corpora and also the training data set of uh this uh L R E zero two thousand and seven training data in those in the first experiment we compare if and if adaptation techniques and with this for uh uh what model but and these uh the yes i uh speaker independent and S A T multiple model um but so first of all oh we found that uh or adaptation techniques for white input but oh you can see that a system able we didn't do any adaptation technique and all the others we use on different kind of adaptation technique and maybe using A C T model and what S S I S I phone model yeah now adaptation and and mean only and uh adaptation performed the best and also you can find that a further improvement can be can be obtained when we use a yeah i say to you for model secondly are we test whether um to phonotactic system with different types of add that to uh also more uh uh that that that model provide complementary information to the uh to each other and better the corresponding system user um cookbook for white a further system uh input what by considering oh curacy whistle at the table that eight sis phonotactic system uh we can combine them oh can can generate twenty eight possible uh to assist on a user and then we plot yeah their corresponding average uh featured a varsity and also that oh be out in the fields the system and you can find that that's a you can also that system using mean only mean only adaptation and bayesian adaptation i i my here both of them um they can provide relatively higher uh oh diversity and also you can see the trend all over all twenty eight possible combination you can see that when you are uh when you update oh higher oh feature directly and then you can all take uh low uh yeah you know the last experiment or refuse to a system using mean only and that the only adaptation that need system at a cheaper so eighty and eighty four and B she too and petri you can see the result you need only here and then the fusion result and we also that just use a lot so system with uh can provide uh obvious improvement for example uh when when aside model use use and a tree and a four is used it can all hold form to the system be one which are S A T model is used and also when we use A S A T model um p2p plus P V we can provide a four door um improvement and you know vol when you compare this result using S A P model and comparing with uh a one before any uh adaptation techniques we can provide overall uh around forty percent relative improvement one two seven uh we have studied a different types of C and uh and and uh adaptation techniques for the phonotactic language recognition oh yeah yeah that's true uh illustrate oh yeah that a mistake model adaptation and we found that um and then only and no adaptation which polite and uh the phonotactic feature so i can provide a complementary information to the one using mean only mllr cation and our ongoing work include uh to see the interaction with a recogniser fun and and also we we investigate more sophisticated adaptation technique and that's all all all my temptation fig let's see you could use hmmm you mean for a second all test data yeah yes text we used hmmm fig no i didn't do it but uh in a room where motif on the first exactly you know but that that would be a problem if you we test it on the feedback and all kinds i control yeah but is likely to be no and in this movie also sure to think about this moving paul so that i i thought about the most of you data that yeah you see this no hmmm with extreme hmmm yeah yeah sure sure mixture sure sure exactly yeah but i it is in this moment in all in the very study we found that even using the simple most convenient uh commas a new method we can still get some improvement but school of course you are right we can do some more in public uh interpolation we have some some uh like that a universal adaptation trans is one yeah and your your you sorry you hmmm hmmm i oh you you mean i using from a practical acoustic or well as well uh to yeah hmmm oh oh you mean a and five test diffusion with a system no i didn't yes hmmm yeah sure sure sure but i didn't make a number so that depends on it yeah questions okay