so that and give you a a a uh a um all of you up the whole to locate and just going to you give a brief description of how the a to model various acts to model classes or organise just to give you a flavour of file what is meant by the court is modular and parts that don't need to know about each of the north um um so just tool re rate um uh the thing that we support currently it's it's mainly the the standard max in the cute training of acoustic models together with a gmms and in the kind of max that cute framework um we have the usual in your transforms like lda to and S T C um we also support speaker adaptation currently if a are is a we have tested it in the recipes mllr lower court is there it's you mean tested um this still any to right so somebody needs to write the the cable and and on them um so mllr is not in the recipe almost done um and well uh and uh and leather obviously has but it it's with which trees then if and lower has to variations of one it's it's just a global transform or with which trees a uh yeah and i this is the point which and once uh then can mention that that we had some discussion whether two a sub for um things like uh do you known type systems are be take models where uh and uh for now uh things are fairly simple um we decided not to do it now maybe if the need is felt in feature and sometimes P also for the course of this development a a couple of times a part my be good to have a system like that but currently when a gmm it's it's uh a very specific thing with means and covariances uh and i'm going to just be few also see how the gmms are implemented um and yeah the sims in the thing with is gmms we also have the uh if from lower adaptation court phrase gmms uh and a little bit uh um that there are few results we had previously published which are still lot in this new code base but there uh going to be added so this is this is already been talked about we have a gmm class and uh it knows really in about nothing else other than and what what it contains uh that is the parameters and there is that acoustic stick model class which is just a vector of gmms and for implementation reason or pointers but not that uh interesting uh a thing but uh uh a the green of in this slides would uh signify this technical term called knows about where hit which is and it it could be a so it's so we have a did um as much of inheritance has because uh so um most of the time things are not uh inherited things if uh uh uh object needs to cheap uh track of another object it's either by keeping a once preference uh it that's that case otherwise yeah specific fake uh that will take just pointers and modified that um so knows was about is in that sense that you can think that you know if you have to write the code you have to be to the head or four this on the thing right um uh so so so the gmms are parametrized um using the natural parameters which is a which a natural parameters in the sense of um the that's of parameters of an mention distribution where uh if you right of the like your got you get um this too i think that the uh them the there is a uh the mean time the inverse of the covariance and the inverse of the covariance of the natural parameters of few M and the reason for doing that is then you can do the like your calculation using just two matrix vector multiplication locations because it or if you have diagonal covariance system you have your and you have the mean times in this covariance is the vector and say you five components are i mean i components and you have your data vector and you just do this to make exact vector but and there are last ratings for doing that obviously yeah a to blast is yeah not the most optimize thing but i mean it's still uh a nice um uh we of doing things so um so uh uh uh a graphical uh overview of uh what dan has already said that uh uh we have this as to model class but when it in to the decoder it contracts with this decodable uh object and uh the decoder knows only about uh this the court of an interface and for each type of acoustic model we need to implement the project us as with the able uh interface uh for that model right and the decodable uh object is the one which all some about features and um just that isn't you'd of the like computation and this is exactly how the decoder interface looks like so so but when i be avoid yeah using uh in here dense this is the only exception which would be uh when V have interfaces which we have a you for features for portable and a few of the things uh and these are actually pure interfaces uh so that what B a a a that's only case where we hate um so as you can see it's a simple E the main function is that like you good combination and uh the decoder can know that but there at there no more frames and yeah how many states essentially you have so a for every other model type you then in heard from this end uh in not so um that was the decoding for training we similarly have a object for spring that matters and uh for the gmms and uh in in the same way that the acoustic model is just a vector of gmms the uh the acoustic model trainer is just a vector of uh objects with screen that you and uh yeah yeah okay yes sure this this yeah that my slides are not compatible yeah so um yeah ah um and and and the red arrow means that uh this classes with modified those classes obviously modifies it implies it also knows about and typically modification it doesn't keep any or an object up the other class pictures it has a method which will um take that object and do the modification um so how do you adaptation adaptation for that say uh for feature space mllr um and so it's if it's global it's implemented as as as a simple matrix uh and the matrix doesn't need to know what it as like a a it's it's only the estimation which makes it that from the ladder so the estimator knows about acoustic model nodes about revision too if you're using the version three and if you're using regression P the timber object has just multiple transform um and similarly to so that it from another object then however doesn't know about uh regression feed this concept it just has a bunch of transforms it's a decodable object which nose hoping to read this thing a similarly with mllr uh obviously that has to know that "'cause" model and them a lower uh can either uh you can it can acoustic model and tell it give me an adapted models are to just a all the means and give you and you model uh a i it can do it lazy so that every you can um um so the decodable the decoder will as the D portable to get the lack you'd from an out of date model the the decodable will quite either the M other object which then we'll see fit has already completed this i mean it catches the mean if not then will uh a the mean from the acoustic model and i weekly see that then convert it right so which which is how you would use it can practical uh situation there's gmms have very similar structure again yeah there is that the able uh on the is gmm oh it that should say S jim and the gmm class um it the is gmm model it has this you switch um that's why needs to know about the gmm classes as well right and just for yeah the convenience of coding there's gmm up for the gmm classes that can lead to send out dating class is the same for is you rooms they different because there many uh a big method used in is yeah and things sort nets so am and uh so so the first bullet point there from lower basis for for you miss already published like know to your own work on most uh it's in the old code base new we need to put it in the new one um partially actually done um then a couple of is back then present the symmetric extension of is gmms um so at you can people keep an asking what's summit at means uh um uh uh uh so so that that's also partially done um and then has then mention that we of reading for um that generation to finished and we can out of the this thing things um yes there but parts and discussions and debates and this um and on supporting multiple feature transforms currently you only have global transform send their just put into one chain a regression class yeah i i you can have regression classes for M F and alarms but then you can compose it with any other transform which has multiple john some as well so yeah so so that when i say no yeah no so that's the thing with that but would feature transforms and okay that is to multiple here first of for for each type there are multiple transforms and then my that's types composed of good and i don't know for the roof feel the need for a but when me to the need for a will think about four to do this i and probably will be handled in something like a decodable uh object level because nothing else needs to know about uh how the compose so that's the end of we would be you of a models i i