i have a no i mcgill or talk about that not all background model well speaker verification and the author of this paper they are we found the assumption and the downhill and uh and a friend of which are down and the assumption my name is on the other hand i you would talk uh that this um the idea i mean these people are used to run into a single i think all for you and then yeah clean up the only no being yeah this is the one that is having cable first introduction in this introduction i will uh in why we propose gift ideas and bases also our motivation second we really nice okay that vocal tract yeah school to speaker recognition and then to to uh idea when we do some and then and this we can fix parenthood that ah finally we can raise the ticket and and the multiple background models are proposed and then no i have a cable components first got a mixture model mixture model when was so background model is that because speaker or occasions this term is that way forty quality of that case the they all thought this term such as but i do not see it and the whole thing attribute projection is based on um uh but that but that the the at that gmm you'll be an ad is basic structure and the the most important rolling in this in this this this basics but strong is the you'll be an and that we think are a complete ubm is supposed to right yeah that the speaker independent feature distribution and uh there are no man starts to counting the quantity of ubm first which we are and misc data and the means we pay for all the data better two three but when it was so yeah as a second right right gender oh channel and then but ubm but uh there may be uh there there there are other approaches first that okay well okay yeah the speaker our unity expensive i might have back such that speech rate speech what we'll emotion well collector and the song but the major differences between the speaker is due to the difference between their average week yeah so in speech recognition well we'll check yes no medication is is also used that to obtain speaker independence insurance now here is like uh you're only kills the frequency warping function it's the crew though this is the original frequency and bases there what the frequency and this is what you that current we want and that is what vector is that when we want to they want to get but unfortunately there is no closed expression for days but still we use this this great to get okay yes and that this is what the speech what what the features is that what models and then the rate of this value is that they are zero point eight eighty two one point yeah well to waste that's that's zero point zero eight zero two now we look at i would think that her mental state at the paper okay parents were having a a yeah i thought i was on a six corpora encode past condition and in crosschannel conditions that you'll be answering data were selected from use S I to solve four one side there are about sixties and what it and the sixteen afternoon and that as i see two thousand and three at as i tucson and two corpora yeah about five hundred utterances notice the feature where you mean you and then a fifty sexual mean subtraction feature what you know they're accepted acceleration and that right there i use the so we come back yeah the feature weights that if you choose dimension then hlda is used uh the final dimension of the feature you starting now oh this the finger of readout distribution we present this encourage us not want to industry that the difference between male and female we want to focus on the S it's not as bad uh the wedding is uh he's is that uh the value range from this if we used this value to it wide paper two three will be a may be maybe we can get that arnold yeah so that it has that much attention we might need ubm training they turned into a to use pointed it here says holding to the warping factor for example uh database first the the walking factories zero point eight eight we have one hundred and the it's it's three utterances no this is the whole structure of our proposed multiple background model uh i think that this structure is that right gmm you yeah uh i'm the think different these days new gmm ubm there is only one ubm and then in this way how ubm who you're you map adaptation each ubm is adapted to generate uh unique a speaker model and they ubm and the speaker models have warmed up here the only in the test the framework and i is used for yeah ubm and the speaker G G M and to table the but phonetically we shall score that well that's what kinda stuff results baseline baseline performance and that uh wait you gender independent yeah and ubm sister uh the eer for the forecast conditions are about ten percent and then wait while the and by the data you show and the where you gender dependent ubm the results if if the gender all pool ubm an agenda of confusion unmatched then the performance i'm not be improved but if the cross gender confusion just contact dave and the days we can how were bad without now this is that we have dependent ubm problem it's table we can see that four female condition ubm to game the bastards are and a four male can you should ubm six scale the best result not have a good some performance comparing that you'll be able to resolve for female conditions and the ubm six results for many conditions admits that it's not we can find that i are you yeah maze for an aspect and will so that get the training data that are contained in the back the performance in the ubm with all the training before now that's reached into his finger again uh we can get or a lot of space but there is wise enough to have his if you will for a test utterance which we are hampered you also but okay racial or just connect the ends and B M can obtain a score vector we can use coffee remastered to obtain the final results when we talk about you and and the contributions of it all in singapore vocal and that's really uh powerful tools back to but here we just want to some simple and the a simple simple uh buster first have a mismatch right uh we just you have a very but uh you look at this thing right the results it's not very good and then we'll back maximum exactly what the master and we use the ubm which do not report is the max as the final score but the yeah we can discuss this also and that's the way your minimum that you were the racial master and the you know gives the best results model based remastered and then there is the question arises why the minimum yeah the racial matters K with the best result unfortunately wait i'm know the exact reason uh in intuitively that peak we we we we try to do uh ah combination in jokingly that speaker yeah and i'm actually for the and the the you yeah you would wear both increase if what match has utterance is in court and this is just which transmits meetings bother with no the reason uh we can make it uh the means and the standard every iteration of that that oh that we would rituals oh well as i tucson has six which each yeah and we put a thinker just like this uh i have to say that it's finger is not the reason of this send of this intense uh we just want to know why now we you know the components ah in this paper with was to investigate here the week yeah that's the right term for ubm training the interesting action experiment short time that you'll be actually them is about you new media data with battered in the ubm trend with all that they are based on this finding we further propose a multiple background model system yeah right you take multiple speaker gmm and ubm yeah for speaker recognition uh through minimum now we we shall feel with the proposed master and improve the performance i'm used to be but yeah you're right open questions what the minimum that we would we show master gave the best results it's just locally experience uh we will be posting the slow but the property is under investigation well techniques to improve the state of uh standard of that this term uh for example if you the yeah the system and the yeah system you are you yeah the performance yeah cool uh we know and the way people to the expression experiment how about that computational cost the and the bases another sure finally i just talking about you have plenty of time for questions there were so hmmm oh i'm sorry it wasn't clear to me exactly how you choosing the you have multiple unions and we'll selecting what within each ubm you mean the two semester you just you have multiple units which are built a little different datasets to uh address how do you decide what when each year uh yeah it's cigarettes the what factors is different and the we use the if the hmmm for example that if the uh warping factor is that the airport mine then this is the with some back this paper to train the ubm when there is a website well it does that and this and that it is yeah what during the enrolment and the test the and the the uh we can have oh the tree data and test data and not extracted they are just the scroll back a two uh ubm and the speaker gmm pattern so i'm just going to be a a synthesis using the all the user density and the just to use they were easy to combine this discourse is used two to school too if i noticed like usual as well you you remote questions uh hmmm uh you know where you use yeah did you for speaker no i mean you mean how many of them is maybe if you assume that you are using you you you did you yeah you know yes i know that's question yeah in fact there are really many mass transit to uh so get data and the i think of like yeah it's just a way of them which has just well then and they are ah true yeah me either right it would be and that's where one last question you yeah hello i just one question um you just use the vocal tract length normalisation to select the population for the building the different buttons models and you can also use the acoustically to to select appropriate calls for normalisation looking for speakers which uh more close to the actually the speaker how do you own it is pretty maybe comparing using different but remotes or different population for normalisation with we propose from it i can't i think it's a i'm i'm just the asking these uh it's do you uh do you have this is also the vocal tract length normalisation to select a cohort six of the speaker for normalisation of the school yeah you mean i don't know uh i don't know what you good right the same that would be probably okay yeah she's didn't uh uh we don't want to just the framing of will or will not too close and uh if you yeah very nice did you do that for for you would you only yeah so you remotes and we move the twos that's because figure