Embed code
Note: this content has been automatically generated.
00:00:00
current experiments
00:00:02
can get results so so basically this is like a case study
00:00:09
so i i basically it's talk very briefly about speaker discussion pens and what they are
00:00:14
uh how the random data and then also the weather like detection is and um what attacks
00:00:20
after the speaker traditional parameter speak about metrics and how you fuse those two together
00:00:26
i was like i'm going to church right now i'm so stupid he is because occasions
00:00:31
system is basically a new in the role a speech sample from some people
00:00:37
you extract some features you model speaker and then you when you
00:00:41
then create a model so you speakers uh in this in the system and then when a test somebody wants
00:00:46
to actually check that uh uh looking into the system of the cape may ask it by the system
00:00:53
he speak into the microphone and uh if you sample those
00:00:56
six weeks the features extracted from the sample and uh
00:01:01
it's compared to uh with the model of the speakers in the database and then
00:01:06
uh it the checks if they're actually the speaker is uh is is user is the the one that isn't that the based on that
00:01:13
if he's matching uh refuse actually matching that the person who's pretending to be
00:01:18
and have like a decision score or whether this is a a person is verified remote
00:01:23
so that the big o. is out all the experiments uh four speakers occasion uh something
00:01:29
like this like a like a plot that do so only give faces like uh
00:01:33
a face recognition but also for speak recognition similar you have on the left side you have uh
00:01:38
uh scores that belong to a zero of the process the call
00:01:42
that basically known jane uses the users that uh that
00:01:45
that that are not those who claim they yeah basically different user send them blue scores of those that uh
00:01:52
uh actually jane user so the point of the system the good ones to separate these two
00:01:56
scores as far as possible so they don't overlap that is great system that means
00:02:01
it's works really well with the glasses speakers were well they're overlapping
00:02:06
too much the the speaker recognition system is not very good
00:02:10
so um how do you do this type of expense and back to so uh as
00:02:15
a as already mentioned binder uh uh as well so you you need to database
00:02:21
and you need to have some kind of a protocol that as a subset and that the basic the
00:02:26
full training you train like a general how model of
00:02:30
what is speech speaker in general is speech
00:02:34
and then uh um you train uh also in intro people specifically
00:02:39
and then uh you also passed on the different speakers
00:02:44
and uh i decided the base unit software and hardware which is quite important and um
00:02:51
you need specialised software that can do these uh tasks
00:02:54
and uh uh extract features generic models compute metrics
00:03:00
and uh at all so you need to have a environment around this of the on
00:03:04
and the hardware that is actually possible uh to use like we store big databases
00:03:09
do you is the g. p. use something there's a lot of neural networks now that require
00:03:14
a lot of people searching for example um so how
00:03:17
do run experiments so uh well one way mm
00:03:24
it's popular way you do take some typical sure software like uh something special is it is quite probable
00:03:29
and you know there's a lot of libraries a lot of things that the return for you
00:03:34
but there's no infrastructure set up for easy to
00:03:38
run an easy to reproduce anything even baseline
00:03:41
so this is an example that you can you have to do is review young researcher
00:03:45
this is a typical software some people might recognise with and um you basically have this type of weather long
00:03:54
very long file like bay somebody doesn't like go and this is basically the state of the art thing
00:04:00
and then have to figure out understand all this and there's a lot of hard coded like bash script here
00:04:06
that have some ah things and some environmental variables in
00:04:10
the linux system that you don't know actually you
00:04:13
what that system is what the what does the required to run that by scrape
00:04:18
on i mean which basher and who knows it's a lot of things here
00:04:22
and uh you like copy folders of data uh from one folder
00:04:26
to another to run your feature extraction so if you have
00:04:29
for different features tractors you have to have four different corpus of your data which can better by it
00:04:34
you have to have four provides disk space and then there's a lot of things
00:04:39
so it's quite a quite difficult situation here at this to me so
00:04:45
it's very difficult to reproduce later so if you say you give descriptor somebody else's
00:04:49
guy say or you have a great model together to stock in a
00:04:51
round the same thing yeah you can that's great good luck to him and
00:04:54
then you never play his email again pretend you don't exist basically
00:05:00
and then so there's another way so as standard short we have this bob framework that we try to
00:05:06
have um a good back to so for a approachable to sharing the cotton the results and uh
00:05:12
uh papers uh uh and also we can generate figures for the
00:05:16
papers uh in a in a repeatable manner uh so
00:05:21
both framework what it does it uh it's actually was built
00:05:24
is a possibility my that is important to exhibit design
00:05:29
uh and uh 'cause some genetic to change for experiments that the useful uh like a threshold
00:05:34
to the face recognition experiment but as we have the same the same tool chain
00:05:37
applies for speak recognition the only thing you need to change is a small blocks and
00:05:41
sidetracked with everything else is actually the same that framework is exactly the same
00:05:44
when recognition well it's different at the base and the different feature extractor so that's about it everything else assume it's easy
00:05:52
so it is the complexity of accessing dating stalling analysts any running things
00:05:57
so that there are some problems there that that is possible to the user should still you know
00:06:01
you still have to have uh the uh uh the good stuff hardware to run this thing
00:06:06
there's some good friends are hard to run on their laptop so you need to maintain
00:06:11
this constantly you need to design experiment as well i mean but that's that's
00:06:16
you always get if you want to do except new experiment you have to design it and actually run it
00:06:20
and then you need to learn how to report unsure results it's as additional thing framework does not really provided
00:06:28
so therefore of the next step revolutions that as
00:06:32
people from that if you just so you'll
00:06:36
basically need for what is important is to understand in tactics is is possible
00:06:40
to the user would you still need to do when using b.
00:06:43
you still need if you have a new database you want to use you still have to ported into plot for which is the process you
00:06:48
have to go through like a great interface support for when stand whether they teach them how to do with it then you have to
00:06:55
to design the tool chain like actually like up your specific framework for
00:07:00
your experiment like what the building blocks what do they do like
00:07:03
which date uh how they connected with what algorithms and how they're get on the last which sometimes can be
00:07:10
uh and work effort to to have to put in the beginning and allocate experiment in
00:07:14
and around and then running is the easy part so the setting up a bit
00:07:18
can be her so he is the um the the actual practical experiment that
00:07:23
that uh i did for them one of the papers mm so i'll
00:07:32
are are basically start from the report so this is the public this is the
00:07:36
is the link that i have here the slides will be available can click late you don't need to that group that long uh uh at
00:07:44
so this is a public report that uh that is basically linked in the into the paper that was published
00:07:49
as a footnote and then person can go here see documentation that like above description what exactly this thing does
00:07:56
uh i have the grabs a the like the ticker so which is kind of like a particular
00:08:02
some uh uh expenses up like uh the the scholars unlike metrics
00:08:06
of uh and then you can click on links experiment itself
00:08:10
and so it's quite a different from them um
00:08:15
this the simplest experiment for was taken
00:08:18
faces is it much more blocks we can see there's a large list of
00:08:22
to uh to uh like uh the blocks that you need to go
00:08:25
through to complete this thing i will complicate feature structure model building
00:08:30
a lot of uh you know big an ally lighting block scorers scoring
00:08:34
the you know all those on the ledge database and also the
00:08:38
timing can notice it to tend to run on our farm out for
00:08:41
the training for example but it takes a couple thousand seconds
00:08:46
so which is what like a a so you can see like algae members uh basically you want to go get it looks like
00:08:54
so this is basically german or so then important but also be that the a
00:08:59
lot of things are hidden here in the sense is provided by people from
00:09:03
so they they so there is a simple i wrote was pretty simple right here
00:09:08
because they're already there's a lot of underlined call that is available
00:09:11
like bob actually use installed indeed so sort of demand
00:09:15
c. plus plus implementations that you just call functions here for your actually specific experiments which quickly uh_huh
00:09:23
okay so this is the experiment the uh for the different verification
00:09:30
uh there are like a different types experiences different databases that they that they use
00:09:36
because i want to see how are from different uh different types of things uh uh so yeah
00:09:46
so far look at so i can also like i'm just show you how you can for can create a new one so i can just uh
00:09:52
to be longer than the other one i'm not good run anything because it takes like a
00:09:57
a few hours to actually finish but you can change the the basis and the there's a selection of gifts
00:10:03
you 'cause you little protocol so that the basis you can try and different things of parameters much more
00:10:07
complex you can have a lot of you know uh uh the good the good thing about the is that when you select a lot
00:10:14
if if it doesn't match and vitamin or something it doesn't appear in the list because you have to look at least here
00:10:20
of a different uh things you can run like a like let's say feature extractor i
00:10:26
have i can have different uh it it it you know my this one that
00:10:33
mm kay so uh this is basically speaker together so you you go to some
00:10:40
just the way the um what it is so well act detection is um
00:10:45
it's uh related type of experiment but instead of a refines because you figure out if they're
00:10:50
trying to spoil the system basically the trying to pretend uh actively yes she died actually
00:10:58
page that an interesting uh so uh you trying to figure out
00:11:02
uh if somebody's actually trying to be pretend somebody he's not
00:11:06
so the typical way so this is like a point of attacking the system that that
00:11:10
most to because uh this type of attack like a one which call here
00:11:14
is when somebody record your your steely recording or download for me to be recording use of your voice
00:11:21
and played back to their education system in trying to pretend that you you uh she's you basically so
00:11:31
i think okay so they're basically that x. can fully occasion system so that access elaborate this reply back
00:11:39
because you you play back just it with all the changes a basic another speaker instead of
00:11:44
person speaking in microphone directly there's a speaker playing back the voice into
00:11:47
the microphone that's only difference really in that is that what that
00:11:51
so that tax here that overlap with the g. with the actual jane scores that means
00:11:56
system much on the stand that that is actually real person a and person little
00:12:03
uh authentic it'd be trying to be instigated by the system that means
00:12:06
you can spoof it that's actually one of the main reason why this type of speaker recognition systems are not used in actually practised because they're afraid
00:12:12
but people just that's moving them not very reliable nonsense so therefore that a lot of issues of
00:12:17
how to detect the study but actually train a different type of model different type of system
00:12:22
to separate the attacks from their real scores you learn some other parameters of the signal to figure this out
00:12:28
and then when you should try that then you have to joint these together to systems you have attack detection system and
00:12:33
occasions just me join them together one after another altogether in
00:12:37
parallel or something so so for back for example
00:12:42
um
00:12:45
so we we have basically yeah i go go from ripple public report because just
00:12:49
a huge which is so if if you have a this type of uh
00:12:53
uh also g. m. m. m. f. c. c. baseball it's a like a quite a standard
00:12:58
standard way to do it so this is another tool chain you can see here
00:13:04
and uh i so now so this is this is basically similar stuff
00:13:09
just a bit different a different type of system and modelling
00:13:12
so now i want to just show the uh example when you join these two systems together so the goal is here
00:13:18
you have this is before you have attacked detection in education system your attacks and not being recognised as attacks
00:13:25
and now you add a affected actions is that together with the education system and now
00:13:29
you have clear separated your good users the the nice guys from the all the
00:13:35
not like space nicely so you have a separation please oppression in the score so you your your system is it
00:13:41
very fine people correctly and this isn't attacks so now getting bit by what do you
00:13:46
need to do not have two different systems right is that the show you
00:13:50
two different put chains so they're a little bit of a problem here then when you create a big
00:13:55
uh we should join this thing you have to draw something like this
00:14:00
because they have to mainly do it with well click and click but you know
00:14:05
double droll these blocks the comment this is like a circuitry
00:14:10
of the big go liberate a due process or something
00:14:16
it's like the g. or something
00:14:19
so this is a bit about sixty plus blocks and the you better not make mistakes
00:14:25
when you do that
00:14:27
because you have to correct them later so you see right so so you do that but
00:14:33
you well the good part here i mean a a scientific experiment requires effort right there's
00:14:39
no way you can escape that i mean if you want to get your p. g.
00:14:42
or get a grant or whatever or paper probably just you have to put effort
00:14:46
so i mean the good thing about these type of approach is you do it once and then you can
00:14:52
basically just the forget about it and talk for years to come and
00:14:55
it's going to be there for you to show and impress everybody
00:14:59
goes maybe also i dunno depends on the girls and basically and and it's verifiable so
00:15:05
people can actually take that explain now that i did once and use it
00:15:10
change the parameters change a little bit block a little blocks and see if it actually
00:15:15
change something databases like there's like five different databases you can put proprietor the
00:15:20
basin around the same topic spread is a state of the art experiment
00:15:23
as to the the uh things you use you can you can for this uh also
00:15:28
told changes huge but you can add small blocks there if you want an improve
00:15:33
the system and the the different type of things so it's it's requests uh it's it's
00:15:39
quite some effort but it's it's stages you for ever so which is very nice
00:15:43
um and you have a public reported that cannot nobody can delete and to you know yep stops paying for the for the costing
00:15:52
let's call that never happens until i'm still alive just um okay
00:15:57
so i i hope it gives you some type of
00:16:00
idea of uh you know how to do this type of experiments like that actual expensive the published can be problems
00:16:07
so you design so that once and but then you you basically just you know you're done right after that you
00:16:14
do that you can repeat you can surround different data can we defy you can change shy compare results
00:16:20
like a likely the boards and all those things if you people interested is not that hard and you can see how actually down like the each other
00:16:26
if you click you have the court exactly like what is going on so
00:16:30
we can really clearly see what what was happening it's very nice
00:16:34
so this is there's was like uh this uh paper that is actually basically it's uh
00:16:38
this results will present in the papers another think is a good for the researchers
00:16:43
when you have a paper in the conference is limited by certain number page like four five six ten right and then after
00:16:49
that it's not paying i mean even journal you can have sixteen page button to pay a two hundred bucks per page
00:16:53
which is kind of sad uh when you have a lot of things to share so these
00:16:58
actually allows you to sneaking then he give much more information than the paper pass
00:17:03
the format allows and you you win because this report that just look uh you know i social so we can can
00:17:09
protect you can put another five papers inside of text and the people can read all the details they want
00:17:14
and the code and everything so it's like much more each information than just the ten pages paper
00:17:20
and you extend your basically page size unlimited early if you want if you're like the right yep that's all

Share this talk: 


Conference program

Welcome
Sébastien Marcel, Senior Researcher, IDIAP, Director of the Swiss Center for Biometrics Research and Testing
24 March 2017 · 9:17 a.m.
Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:20 a.m.
Q&A: Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:54 a.m.
VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 11:35 a.m.
Q&A - VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 12:07 p.m.

Recommended talks

session IM2 Start-up Jamboree, Koemei
Temitope Ola, Koemei SA
1 Sept. 2011 · 6:04 p.m.