models but i should each time but also in young can and it would be uh present a very uh she general so you G U okay uh uh not talk is right features based i'm are are speaker adaptation with that in models okay uh there it all line first uh that we introduce the back wrong of of the red it and the small amount data a the that patient problem uh this will include to part than the few uh the feature based the and are that patient and the by or models and then let's go to the combination of these two part so that's what we T um to combine the feature based i'm are are a and the body and models to do the adaptation work and then that's have a look at our experiment results and a last but it's a conclusion okay no the one of the is in the asr it the mismatch between the training data and the testing data that its distribution distribution between that this tour process so that's why we introduce the adaptation work oh a a we have a a a a people have proposed that will model based adaptation master uh we kind that we have three basic categories of the the first one is the speaker clustering this is the eigenspace space based master are and then it the best base man and the server one is the uh i i are so we can put oh on the model base of that patient but to no now we have for in on the red it that but that patient because in the model base of that should be have a lot of parameters to uh compute and then one uh and the speaker or uh when the number of the one B can you'd are the number of speakers so a new model parameter files we were you crazy further to grammatical is so this is not a a rapid adaptation so will be have to turn to the feature or us based adaptation but we can only use uh metrics uh beach we uh uh used on the of their we you the but feature so that we would be a much it it will be read or uh that's good to the small amount data for uh of a small amount data oh uh a means we have to not really to it's and number of parameters oh oh we have to compute are we have to find out uh reach uh a parameter or be factor it's of the most important so we can do some can space uh i guess but its based complexion and the rank uh and the rank the and battle uh and then we for one backed or it of the most important so this uh a then in prime her space side so we can also i variance some constraints on on a a a a a or more or all compassion for example uh we have uh Q are we have a a lot of story a help that are gonna all an or i'm our all our map are are so we have tried i have my i'll a mess there are so it also performs the uh from score that's so uh one be used i my i i are we should know of the prior you formation because i my means maximum up close to a so we have to now when we have to now uh the prior distribution if the assumed up a are distribution is the same as the row this mess a real key you a good performance but this is not or rest the real case so assuming we to to the prior distributions so uh uh we have to okay case speech and not though the prior information then we propose the by linear mass are so in or mess are we can uh us any are lee for one dollar and that i have and also important factor i some a like that a can it was and then we don't need the prior that the is be in here so last have a look at of the by or models okay let's first uh have a look at that i i R so uh this mass is very easy to expressed uh are an or waiting the after knowing the up so the feature and power time that time are able so a them knowing the of the relations we have a mac six uh uh work on these the metrics is a a and the we also have a a be used still the combine the metrics which is a and times and plus one matrix is that do so the past what means the uh ours so uh also following the traditional pipeline we have the all the other functions a us the a and then what we need to do in the following it's to maximise the uh seat still uh uh in this uh uh you this function uh a video it a is the low uh uh and the very well a both of the utterance for information metrics that a up a little and uh K a and E G as some but relations of the of the feature observed feature okay hmmm okay let's go to buy models uh you by in models we assume the of the rate the observed features depends on to a kind of of of a a a a factor as for example in speech recognition is we have a a a a of the relation but this observation may depends on the speaker and the same at ten hours day depends on the environment so we have to fact or wouldn't be need to do is to decompose a this um this feature why into this two factors i i and the B and mean they'll we have couple things pitch and a a B so that a video it's the coupling matrix so we that actually i a uh a a do it this or to the super metrics put uh because we have uh why is the a why it's a actors so and the super metrics stop a has the number of the number of mac in equivalent to the number i uh elements in by so this is the same actually could uh this is so close to match in a model because a a and a B I the because nobody oh i'd independent oh i i and B so and the the form a also a symmetric four but generally can not fun dollars that independent metrics stop video a you maybe you most cases is stop do have had to read a is a all this be assuming a double of read utterance or always have the right utterance with a so this is that is a matter eek but in models so we multiply apply i E and that we then be obtained the final or peak i i still a big i is the speaker to combination so that i that big a is speaker dependent the combination of factor i and the transfer matrix W so uh uh what a need to do is to obtain a and B after knowing the prior information of the transformation for example uh you know a a for we know a transformation matrix uh be she'd north has a in this the slide so oh we can use svd D to decompose a this metric that uh two two factors U and V U and V can be considered a as the speaker of faction and environment of action so in the middle part as is the uh is the coupling between this two part and ask you have to S P D as it also the singular bad it'll mac so this is uh uh something like a a you can but it was so i to run king and to rear at uh after re the single about bad or the singular battle according to the uh according to their size so a uh we can now the importance of a can or the important of a speaker information and the environment information so how to decrease the uh parameters space uh we can only at dot because only a of first uh maybe of false fly well first ten important and to single but it was uh according to the these we can uh estimate ten hours the layout pain the simplified fight U and V then uh be real uh and then be can decrease the parameter space i'd so then be multiple you and uh ask me can of ten and we can out end of uh the final form of the by in model a and B so uh was we need to do is to find out the a and B in the speech recognition case so um okay okay here is the pipeline for the combination of by the model and i i M a so that a blue is the uh a speaker transformation matrix in a i i'm a and uh it you close to a park that be as the buyers and uh a a i J at the speaker if for a a at the speaker the transform matrix so the final the motion is i times deep plus one so and and in the seconds type we find the average of the transform metrics stuff they are so uh and the we use each transfer matrix stop do my nose i do we have to still and the B can oh and uh um we can a the stack to that we can and the stack the transfer matrix still so these to the uh a high dimensional tree this that axe the different is that the matrix from different speakers together to uh compose these super metrics and then we put will that perform svd to find the speaker information and the maybe that you environment the related information so and the single about it was i uh so a a a and the B are the decompose the and the top loop the are or it is the by a uh it is that have a if it that have a leash for a is that have research the transfer matrix uh should be divided for a should we stuff the stops the track from the room at the the top lip so uh a i it just speaker dependent and B it's only environment a dependent that and so uh in the decoding stage at ten noise in a new uh up to a new be cursed at is is coming need to compute it a and be actually be V have or a up C D so the new you uh and you speaker information is only related to be i so how to compute i the uh for in the some a in the should in a tradition no i i'm our so see to a given is the uh of the or function and uh we replace the uh we sat key two i E and the B S P for you to this out the low of functions and uh then be re move with that root mean let's remote all terms this regard use of a uh uh and then be a ten and uh following forms so this is a a a a batteries three for about the mathematical our or uh uh uh operation okay okay then me last make uh last make the you have a here well uh of the up to their function be street respect to a a and make these uh the year to you to be there are so we can help and we kind of and the solution of these functions so uh i think that uh uh uh i think all these mathematical the operation can be found in a reference paper the name to the i'm i are i think of the read them by would land okay so uh okay this is the solution of this method or okay a problem the how to select a J J is the uh and you and the the after svd we have got a uh so is of single but it was and we only want to keep thus such the a subset of of these but it was so the the the set there's the estimation of this stuff that is J so we only need to keep the first the J important but it was so but how to select these J they are a group of master first one uh that selecting a according to the amount of the that checked that adaptation data so they are some experimental results shown that are shown that and the bad of J the log uh has a lot the relationship that be two in the experiment a the has a lot relationship between in is the the training amount of the adaptation data so we can use this relation log relationship and then be can also use the single but it was maybe uh a to a threshold uh uh can't one the singular better equates grammatical only and that this the base is this threshold then be kind C lack the single about it was before this threshold okay and we can also use you've atoms that you've V we have also route of staff so as uh only we can and use the we can pass maybe we can assume a so it better off J and then G minus one my notes one mine of fun to compute i totally uh i i need to compute the oh the they are a function that is also the and how objective function to find out which a max seem might the uh objective function this they also semester i didn't this that in this paper we use the uh a final my sir two uh i i T be computed the and are objective function but now we whole some simple or mess for example to use the on a the relay a a of the amount of adaptation data you at the experiment results don't them what we want to see that's are in the second experiment the mandarin uh and and voice uh way say H the search data so um the and this is about i uh in this test data the sure that the is is better short uh here is that the oh a six seconds or ten seconds all a can only a question it's use here only a questions that it's use here so after is it matters for so we use a traditional i have a i are are the wer is fifteen point two per but bell be turn to the body in model uh the absolute wer the decrease by a one point five percent okay a a a a a a a a conclusion that the uh to our conclusions but but but in a models can fact that incorporates the as to be sat in in that before for the prior information and the lack pretty read the number of parameters space and the filter work is the and the first is to select G is to select the single that the number of single or it was J in a single right and the second is uh you our work we have or seen by the linear models this that the speaker or information and uh for example or the environment information but actually we have been know the speaker information be them know the environment information so the second filter work is uh we can was not one if we compare snow the speaker information and the new bound of information here it is that the class dependent to information so oh was he can do further to you you but uh to increase the performance of this bilinear model oh and the this so one is the how to control the speaker number still uh that's and uh that's our kids work so if you have a do have a fast race a a quick question okay so if you want to not some details about this work please right to i have john has sinned dot I B M calm and you very much you