Přepis řeči - Evaluation of the Vulnerability of Speaker Verification to Synthetic Speech

0:00:06	um i haven't unique challenges you i'm
0:00:09	it in that case
0:00:10	um and we can see i'm sad also
0:00:12	but that's my reading
0:00:14	or
0:00:15	and then
0:00:15	i was
0:00:16	so
0:00:17	i have to prevent this
0:00:18	instead of
0:00:19	fast
0:00:19	the colours
0:00:21	but
0:00:21	to begin
0:00:22	which
0:00:23	a two
0:00:25	some D V C R
0:00:27	which mentioned simple code you know it's
0:00:32	oh
0:00:35	oh
0:00:37	um this that is
0:00:38	on that
0:00:42	oh
0:00:43	it could mean human speech
0:00:44	so topic
0:00:46	so what can we
0:00:47	all copyrights
0:00:48	good
0:00:49	and use it
0:00:49	in in
0:00:50	two
0:00:51	yeah i'm
0:00:52	fine
0:00:53	fig
0:00:53	in the future you may be possible meaning i'm one sport
0:00:56	that
0:00:57	uh but
0:00:58	equation
0:00:58	just one
0:00:59	yes
0:01:00	well
0:01:01	that are taken as you both
0:01:03	just
0:01:03	oh
0:01:04	problem is
0:01:05	okay well
0:01:05	the big one speak
0:01:07	search
0:01:08	and then
0:01:10	yep
0:01:10	this problem
0:01:11	cool
0:01:13	and
0:01:13	because then if you're still
0:01:15	yeah
0:01:15	two
0:01:17	so we in this
0:01:18	talk
0:01:18	we evaluate
0:01:20	how to secure
0:01:21	the speaker verification systems uh
0:01:23	okay
0:01:24	fig speech
0:01:25	yeah cool
0:01:26	someone
0:01:27	speech voices using just and sent
0:01:30	but once and
0:01:31	but we can call
0:01:32	and a speaker's voice from ten sentences
0:01:37	but this is a content
0:01:38	my talk
0:01:39	i to talk about
0:01:40	some
0:01:41	yeah now introductions
0:01:42	um i'm wrong
0:01:44	and then we
0:01:45	there's a lot
0:01:46	recognition ideals
0:01:48	the S U N
0:01:49	then
0:01:50	i we show some of its work
0:01:52	which
0:01:53	right then
0:01:55	i cats
0:01:56	this year
0:01:57	and then i think
0:01:58	panes of a speaker verification systems
0:02:01	for by
0:02:02	i think the system
0:02:03	useful
0:02:04	yeah
0:02:05	and then i mean streak payment conditions
0:02:07	um
0:02:08	uh
0:02:09	and then
0:02:10	i wish also
0:02:11	or some
0:02:12	quiet
0:02:13	to detect
0:02:14	synthetic speech
0:02:16	speaker verification
0:02:17	using
0:02:17	i mean so
0:02:18	cool
0:02:19	oh yeah yeah yeah
0:02:20	and is that what it right
0:02:22	and they are somewhat item i
0:02:28	so do you know about that
0:02:31	but it's not a you know
0:02:32	no my kids will be
0:02:34	because we can assist them
0:02:36	how some
0:02:37	you know
0:02:37	but i know we used to
0:02:39	tts systems that used to be
0:02:41	and
0:02:42	in the conventional we thought
0:02:44	conventional scenarios
0:02:45	pdf is then that's true
0:02:47	ah
0:02:47	it's great you need selection tts system is
0:02:50	right
0:02:51	so or what combos on technique
0:02:53	unit selection if the just
0:02:54	and peso
0:02:56	equated with ones
0:02:57	and he and then transform
0:02:59	so someone's voice
0:03:00	to target speaker with
0:03:02	using your joint probability of gmm
0:03:05	trained on all right
0:03:06	right
0:03:07	um
0:03:08	three
0:03:09	in any of that
0:03:10	six
0:03:11	well beatification
0:03:12	of course
0:03:13	can be things i
0:03:14	from only you can see but
0:03:16	um also since uh
0:03:18	oh
0:03:19	with a fair fight
0:03:20	five
0:03:21	can be transformed into
0:03:23	basically how good
0:03:24	ha ha
0:03:25	a voice
0:03:26	you think this document
0:03:27	but this
0:03:29	combination
0:03:30	oh
0:03:30	problem
0:03:31	speaker
0:03:31	or something
0:03:32	i think it's probably
0:03:33	speaker verification system
0:03:36	but
0:03:36	all the is distance
0:03:38	is it
0:03:39	this
0:03:40	yeah fictional
0:03:41	our tts systems uh
0:03:43	it can be
0:03:44	speech synthesis
0:03:45	cross speaker adaptation
0:03:47	such as embedded uh
0:03:49	it
0:03:50	um
0:03:50	this
0:03:51	just a
0:03:52	also
0:03:53	what's the problem
0:03:54	speaker vacations because speaker adaptation scan possible
0:03:57	speaker independent agent
0:03:59	which
0:04:00	which
0:04:01	which
0:04:02	um
0:04:03	which are cool but this was more distinctly dsp use
0:04:06	into the target
0:04:08	okay
0:04:08	a voice using small amount of data
0:04:11	and then
0:04:12	uh that
0:04:13	any i don't use for verification
0:04:15	can
0:04:16	insights from
0:04:16	and update more
0:04:18	so
0:04:19	this
0:04:19	justin also
0:04:21	probably
0:04:22	speaker recognition
0:04:23	but
0:04:24	we'll be the justice system
0:04:25	it's more
0:04:27	probably
0:04:28	more
0:04:29	interest
0:04:30	it's combinations i needed
0:04:31	i think
0:04:32	basically
0:04:33	that's
0:04:33	and this problem was fast
0:04:35	reported by my scores
0:04:37	and you have a go
0:04:39	so why do we need
0:04:40	why do we need
0:04:41	this is used
0:04:43	there are several times
0:04:44	um
0:04:46	the positive and its performance
0:04:48	it is it and basically
0:04:50	thanks
0:04:50	the whole month of its fear
0:04:52	it can be this way
0:04:53	that's right
0:04:54	quite different ways
0:04:56	in power
0:04:56	the quality of a ten base
0:04:58	yeah it's no problem
0:05:00	with
0:05:00	well detection systems
0:05:02	and
0:05:03	and then
0:05:04	well
0:05:04	it
0:05:05	it in disobedience
0:05:06	of holding elections
0:05:08	more specifically in basically field agent based
0:05:12	it's
0:05:12	same as human
0:05:15	uh under
0:05:15	speaker adaptation techniques
0:05:17	what speech is a hot
0:05:19	what
0:05:19	maybe
0:05:20	cochlea
0:05:21	it
0:05:21	well yes
0:05:22	we can do
0:05:23	speaker adaptation
0:05:24	unsupervised
0:05:25	but uh
0:05:26	like
0:05:27	is that
0:05:28	which add up to a much past
0:05:31	we got a job which is
0:05:33	and also we need
0:05:34	be able to use
0:05:35	we can use
0:05:36	when the us
0:05:37	in part
0:05:38	clean speech data
0:05:39	uh
0:05:40	fig
0:05:40	adaptation data
0:05:43	so
0:05:45	taken together
0:05:46	it is now possible
0:05:48	automatically create how did it
0:05:49	because tts voices from any at all
0:05:52	what about it
0:05:54	which
0:05:55	i thought that
0:05:55	right
0:05:56	no
0:05:57	which means
0:05:58	what do by
0:05:59	oh
0:06:00	available ones of it
0:06:01	can be used
0:06:02	oh
0:06:02	affecting
0:06:04	speaker but it was just
0:06:07	so i think should not you
0:06:08	yes
0:06:09	not you
0:06:10	he
0:06:12	that's fine
0:06:13	speech data
0:06:13	five a quite well
0:06:15	well the gas
0:06:17	well look at
0:06:19	or texture
0:06:20	like this
0:06:21	you know we can record my speech
0:06:23	i think
0:06:24	my speech might be like
0:06:26	right
0:06:26	um anyway
0:06:27	so
0:06:28	we can
0:06:29	why a speech one board
0:06:31	all the cows
0:06:32	well cast
0:06:32	because jazz
0:06:33	well it
0:06:35	then using this
0:06:36	well
0:06:36	p2p does
0:06:37	that that uh
0:06:39	yeah about it can be a speech
0:06:40	systems
0:06:41	right
0:06:42	the other ones
0:06:43	then
0:06:44	you think about
0:06:45	what is
0:06:45	but
0:06:46	yeah
0:06:46	right speech
0:06:47	or beatification
0:06:48	useful
0:06:50	because it gives
0:06:51	it is
0:06:51	and then
0:06:52	we prepared
0:06:54	accept samples
0:06:56	um
0:06:56	which
0:06:57	much
0:06:58	to the scenarios
0:06:59	so we really is terrific speech from yeah
0:07:02	does it have a year
0:07:03	which
0:07:04	but cats
0:07:05	and also
0:07:06	clean it up
0:07:08	they can
0:07:08	well i guess
0:07:10	i go
0:07:10	you know
0:07:12	it's got
0:07:13	and there
0:07:14	um see fig speech
0:07:15	is
0:07:16	you know
0:07:16	i pray
0:07:17	couple samples on this in six speech samples
0:07:20	oh okay from a genocide that
0:07:22	yeah
0:07:23	it's put together and how
0:07:25	yes
0:07:25	so what's up with
0:07:26	george bush
0:07:29	yeah
0:07:42	so he's adapted with this meeting we keep a T S
0:07:45	one george bush
0:07:46	or not
0:07:47	and then
0:07:49	clean it up
0:07:50	fig
0:07:57	yeah
0:07:59	right
0:08:02	can you identify how
0:08:04	oh
0:08:05	oh meeting people communicate
0:08:08	yeah
0:08:09	maybe
0:08:09	and and of course
0:08:11	yeah it's inside
0:08:12	speech
0:08:13	yeah i know
0:08:36	so
0:08:36	the
0:08:37	with this
0:08:37	just
0:08:38	octaves
0:08:39	have also
0:08:40	but uh
0:08:42	oh
0:08:45	size
0:08:45	so is this poses
0:08:47	yeah times
0:08:47	with
0:08:48	fig
0:08:51	so
0:08:52	um
0:08:52	yeah
0:08:53	let's go back to
0:08:54	sorry
0:08:56	um
0:08:57	but
0:08:59	okay
0:09:02	so
0:09:03	i hope you understand that
0:09:05	the security issues of this
0:09:07	and then we use um
0:09:09	explain
0:09:09	i'm sure that is uh
0:09:10	yeah i guess
0:09:11	two thousand and
0:09:12	that's
0:09:13	so we
0:09:14	we use it in this
0:09:15	databases
0:09:16	which
0:09:17	ah
0:09:18	i agree
0:09:19	but
0:09:19	speech
0:09:20	uh why
0:09:21	when you and john
0:09:22	because
0:09:23	and then we you really really simple speaker verification system is in place because
0:09:27	um but we well i yeah i know
0:09:30	this then
0:09:30	but yeah so what standard gmm ubm
0:09:33	and also you know gaussian
0:09:34	but but if it's at the end
0:09:36	which
0:09:37	you know some people
0:09:38	this
0:09:38	yeah
0:09:39	use
0:09:40	right now
0:09:40	um you know
0:09:41	so the with
0:09:42	score normalisation feature
0:09:44	normalisation
0:09:45	but when he is
0:09:47	there's no significant device
0:09:49	this from a point of views
0:09:51	because in
0:09:52	most cases
0:09:53	the speaker verification system
0:09:55	a tape
0:09:56	green
0:09:56	fig speech
0:09:57	voice
0:09:58	you know i think
0:10:00	um
0:10:00	so
0:10:01	in the store
0:10:02	i
0:10:02	it was one indian
0:10:04	you'd be in it
0:10:06	but
0:10:06	what we have used
0:10:08	but
0:10:09	um which are basically the same
0:10:12	so this is the design
0:10:14	previous
0:10:15	so it's
0:10:16	oh
0:10:17	oh
0:10:17	well
0:10:18	what distributions one ten german speakers
0:10:21	um
0:10:22	this
0:10:22	we do not
0:10:23	sure
0:10:24	school
0:10:24	what human speech
0:10:26	target
0:10:27	because
0:10:28	um
0:10:28	the
0:10:29	human
0:10:29	that
0:10:30	ha
0:10:30	sure the human
0:10:32	each
0:10:32	well this was just
0:10:34	and
0:10:35	this is a
0:10:35	i see fig speech about
0:10:37	impostors
0:10:38	it is not
0:10:39	with a button
0:10:42	uh
0:10:42	you did
0:10:43	and then this is a
0:10:44	fig speech about
0:10:45	oh i guess
0:10:47	and
0:10:47	green one
0:10:48	really
0:10:50	right
0:10:50	these figures
0:10:51	sure
0:10:52	scene six speech will again
0:10:55	that you can see
0:10:56	these qualities previews on
0:10:58	for human size
0:10:59	speech
0:11:00	for both
0:11:01	postures
0:11:02	and and also
0:11:03	what to do green
0:11:04	i need
0:11:05	i think
0:11:07	yeah that was
0:11:08	it can
0:11:10	okay
0:11:11	it's not
0:11:11	and
0:11:12	but was yes
0:11:12	but the problem is you know pretty
0:11:14	payment
0:11:15	because number of speakers is
0:11:17	yeah
0:11:18	yeah
0:11:19	too small
0:11:20	and then
0:11:21	the speech data use
0:11:22	was a
0:11:23	read speech tagged as
0:11:26	but you know oh i think it's not you speech data
0:11:29	why
0:11:30	be you know
0:11:31	it's assumed to be not a
0:11:33	clean
0:11:35	so
0:11:37	in this
0:11:38	cool
0:11:40	in this new book
0:11:41	so we use three hundred speakers
0:11:43	included
0:11:43	was the channel zero
0:11:45	i would say to
0:11:46	eight
0:11:46	so what
0:11:47	right
0:11:47	they were
0:11:49	this this
0:11:50	oh
0:11:51	much of it that some tts corpora because
0:11:53	yes
0:11:54	this is
0:11:55	yeah i agree
0:11:56	you know it's not perfect
0:11:58	three
0:11:59	you know and vitamin
0:12:00	and stuff
0:12:02	is
0:12:02	because
0:12:03	and p2p
0:12:04	uh what is the point or something else
0:12:06	i think snotty
0:12:08	and also we therefore happiness to you
0:12:10	formation on missile could detect
0:12:11	fig
0:12:12	speech
0:12:13	because it cation systems
0:12:15	what sample sample
0:12:16	it with
0:12:17	sup sup
0:12:18	i thought it was a mess
0:12:20	fig
0:12:20	speech
0:12:21	in speaker verification
0:12:23	wow
0:12:24	but again
0:12:25	which is you
0:12:26	speech becomes much better
0:12:28	someone
0:12:28	so we have a body
0:12:30	dismissal
0:12:31	obvious
0:12:32	certainly
0:12:33	um
0:12:33	probably more
0:12:34	this impostors
0:12:36	a lot it's and it's
0:12:37	hmmm
0:12:41	um
0:12:42	histology about your name
0:12:44	ubm
0:12:45	guns
0:12:46	i think right
0:12:46	but
0:12:47	you
0:12:48	uh the way you want it
0:12:50	uh we use
0:12:51	if the end of the
0:12:52	the the stuff
0:12:53	no energy on it
0:12:55	data
0:12:56	um we a bright future
0:12:57	one thing
0:12:58	right
0:12:59	robustness
0:12:59	proposed
0:13:00	by then
0:13:01	uh we had that is
0:13:03	G and then you'd mark
0:13:04	adaptation
0:13:06	um
0:13:06	in addition to
0:13:07	what janet was we evaluate it
0:13:10	yeah but you didn't system
0:13:12	you'd be used for the whole process
0:13:14	which
0:13:14	we have a
0:13:15	but
0:13:15	because
0:13:16	and uh
0:13:17	okay right
0:13:17	what
0:13:19	right okay about
0:13:21	right
0:13:22	and
0:13:23	which is
0:13:24	level or more
0:13:26	is that it
0:13:27	right
0:13:28	so
0:13:29	probably
0:13:29	this
0:13:30	she's
0:13:30	be
0:13:31	and uh
0:13:33	um
0:13:34	this is the
0:13:35	quite well but
0:13:36	that over the whole
0:13:38	i don't speak about it
0:13:39	it's
0:13:40	so quite because it's in this piece of this
0:13:42	it's the complex it's
0:13:44	that's that's
0:13:45	in march possibility
0:13:46	i really want in speech right
0:13:48	but
0:13:49	no
0:13:49	speaking
0:13:50	so we use
0:13:51	this guy same technique uh
0:13:53	it starts
0:13:54	training
0:13:54	average for some of this
0:13:56	which is
0:13:57	basically
0:13:57	yeah
0:13:58	i did
0:13:59	ubm
0:14:00	or
0:14:01	speaker independent agenda
0:14:02	so we use
0:14:04	because of it i mean
0:14:05	yes it is hot
0:14:06	it is with some of the
0:14:08	yeah we
0:14:09	uh
0:14:09	uh what is
0:14:10	you think
0:14:11	adidas
0:14:12	functional like houdini regulations
0:14:14	well you know pulse train and made it off or
0:14:17	it's not about
0:14:18	see in the data
0:14:20	yeah
0:14:20	small amount of because
0:14:21	okay
0:14:22	be
0:14:23	then
0:14:24	we generate
0:14:25	acoustic on that
0:14:27	such as
0:14:28	but
0:14:28	um uh so
0:14:29	each duration so some
0:14:31	noise
0:14:32	for me
0:14:32	citations from the side of it and then
0:14:35	you mean maximum likelihood
0:14:36	on occasion as well i
0:14:38	proposed by
0:14:39	with a ninety five
0:14:41	for this taken out
0:14:42	can you it's
0:14:42	yeah
0:14:43	how much someone says
0:14:45	and then
0:14:46	and then
0:14:47	you think it is generated
0:14:48	acoustic um it does
0:14:49	we run
0:14:50	and i would be
0:14:53	with the whole
0:14:54	right
0:14:55	proposed by colour
0:14:59	and then
0:15:00	this is about patience
0:15:01	so we can create
0:15:02	new
0:15:04	tts voice
0:15:05	from
0:15:07	um
0:15:08	senior
0:15:09	just
0:15:10	that's from three minutes of speech data
0:15:12	was
0:15:13	if
0:15:14	speech database
0:15:15	a bit of more quickly becomes bit
0:15:17	but
0:15:17	minimum
0:15:19	the meeting
0:15:20	if
0:15:20	where am i
0:15:21	yeah
0:15:21	i think i'm leery ha
0:15:24	of them with this or that
0:15:26	and this
0:15:27	small
0:15:28	sure
0:15:29	at that
0:15:29	individual speakers
0:15:31	and then they
0:15:32	well actually the
0:15:34	a female speakers
0:15:35	and in this
0:15:36	remark
0:15:37	sure the male speaker
0:15:38	other people will
0:15:40	uh as you can see
0:15:42	this paper
0:15:42	how about
0:15:43	his point
0:15:45	and also china
0:15:46	and so on
0:15:47	um
0:15:47	you
0:15:48	and that's it
0:15:50	and sounds
0:15:51	which one
0:15:54	and that was my question
0:15:56	how many voices available in
0:15:58	mark
0:15:58	can be
0:16:00	who the speaker verification systems
0:16:06	so again
0:16:07	our scenario
0:16:08	it's not building tts system
0:16:10	on speaker verification databases
0:16:12	it is no money you don't narrow band
0:16:14	ooh
0:16:15	go to the noise
0:16:16	or maybe all five microphones
0:16:19	oh what can i do
0:16:20	is
0:16:20	you know
0:16:21	most of my nearest acquire speech because
0:16:24	um you know we
0:16:26	why you
0:16:27	crises
0:16:28	like this
0:16:29	they adapt
0:16:30	yeah
0:16:31	fine
0:16:31	so we use
0:16:33	okay i think we we use
0:16:34	also i don't know
0:16:35	um
0:16:36	data bases
0:16:37	sort of this
0:16:39	database
0:16:39	yes
0:16:40	um
0:16:41	two hundred eighty four speakers
0:16:42	uh we
0:16:43	weeks
0:16:44	once because
0:16:45	fig
0:16:46	can you got even
0:16:47	uh we use
0:16:48	and it's a speech
0:16:49	and then we buy
0:16:51	excited for it
0:16:53	speaker but you in to see it
0:16:55	if you see the old
0:16:56	and that it
0:16:57	it's for
0:16:58	training data source
0:17:00	tts
0:17:01	um in the set they retrain
0:17:03	average voice models
0:17:05	or by speaker adaptation
0:17:07	individual speakers
0:17:08	we use
0:17:09	she made it out was trained and data for the patient
0:17:12	and that be it
0:17:13	training data that's for speaker recognition systems
0:17:16	um
0:17:17	right
0:17:18	any buzz about that one
0:17:19	what is
0:17:20	in a moment
0:17:21	uh we have that
0:17:23	yeah that's what
0:17:24	but
0:17:25	set see it has been as
0:17:27	which have
0:17:28	these accounts
0:17:30	all speech data part
0:17:31	but also
0:17:31	that's because
0:17:32	and this
0:17:33	to be
0:17:35	if that's true
0:17:36	speech data
0:17:37	just from
0:17:38	useful cations
0:17:39	um i did for a couple of samples
0:17:42	um
0:17:43	data from this
0:17:44	yes
0:17:45	trained on this was original data
0:17:49	oh
0:17:57	come on
0:17:58	one
0:17:58	this policy
0:18:01	yeah
0:18:09	yeah
0:18:24	yeah
0:18:24	so
0:18:25	is this too long reverberation
0:18:27	you you
0:18:28	huh
0:18:28	this thing
0:18:30	is um
0:18:32	yeah
0:18:32	a big car
0:18:33	they show yeah
0:18:34	right
0:18:34	ready to
0:18:35	additionally the weight of a
0:18:37	oh
0:18:37	you must
0:18:38	not
0:18:39	um
0:18:40	and of the you know the
0:18:42	equal error rate
0:18:43	it
0:18:44	just
0:18:44	the point five
0:18:45	this is a
0:18:46	false alarm probabilities and
0:18:48	season
0:18:48	diction
0:18:50	um
0:18:51	so
0:18:52	we can
0:18:52	see
0:18:53	speaker verification
0:18:54	for human speech
0:18:56	so you don't know
0:18:57	yeah
0:18:59	but that's why you know we can say our speaker verification systems channel
0:19:04	they are
0:19:04	can't distinguish
0:19:06	because yeah speakers part
0:19:07	almost part
0:19:09	and the
0:19:09	this is that is that
0:19:11	human
0:19:12	speech
0:19:12	but
0:19:13	speech
0:19:14	um
0:19:15	if the score distributions
0:19:17	uh
0:19:19	similar to create
0:19:21	i mean this is the human speech
0:19:23	what are you
0:19:24	fig
0:19:24	because
0:19:25	um
0:19:26	this is the same sex
0:19:27	speech about
0:19:28	target because
0:19:29	um this is a human speech
0:19:31	input just
0:19:32	well this is
0:19:33	six
0:19:33	speech but also
0:19:34	just
0:19:36	and
0:19:37	the distribution
0:19:39	all this
0:19:40	was good
0:19:40	this for distribution
0:19:42	um no
0:19:43	i don't know anymore
0:19:44	but as you can
0:19:46	these
0:19:46	they uh
0:19:47	significant or whatever
0:19:49	in but
0:19:50	in march
0:19:50	claimant is
0:19:52	where
0:19:52	lies voice
0:19:53	okay
0:19:54	you know
0:19:54	maybe the extreme um hum
0:19:57	about
0:19:57	ninety percent
0:19:59	speech
0:19:59	but
0:20:00	it it
0:20:01	so
0:20:03	see
0:20:03	much
0:20:05	train
0:20:06	uh two hundred
0:20:07	sixty
0:20:08	was
0:20:10	oh
0:20:11	fig
0:20:11	two hundred six people
0:20:13	was actually
0:20:15	so someone is of course
0:20:17	but despite
0:20:18	excellent performance
0:20:19	because the case was this thing which
0:20:21	uh
0:20:21	one
0:20:22	why
0:20:22	it out i
0:20:23	all right
0:20:24	the speaker i didn't
0:20:25	speaker
0:20:26	his eyes
0:20:27	before
0:20:28	speaker out of it
0:20:29	it is because this is
0:20:31	hi
0:20:32	enough to allow the use
0:20:33	right
0:20:34	pause
0:20:35	to do human
0:20:35	right
0:20:36	going on
0:20:38	see what i keep up
0:20:40	well
0:20:40	what
0:20:41	yeah
0:20:42	um
0:20:43	because they have significant overlap
0:20:45	i just meant
0:20:46	decision
0:20:47	the shooting was
0:20:48	one of my vision
0:20:50	uh uh like
0:20:51	the head
0:20:53	so of course problem is how can we
0:20:55	this
0:20:57	yeah we are not
0:20:58	all right
0:20:58	right
0:20:59	it was
0:20:59	so yeah i
0:21:00	yeah i just
0:21:01	so we
0:21:02	why
0:21:03	yeah
0:21:03	extra missile
0:21:04	yeah the commission on it
0:21:06	which uh
0:21:07	nothing
0:21:08	if i see them like we do
0:21:10	what's your idea i propose but so what
0:21:13	and also we use
0:21:13	what is that what data rate
0:21:15	um we can from the us
0:21:18	curious
0:21:19	you know
0:21:19	oh no
0:21:21	um
0:21:22	a base
0:21:22	and define
0:21:23	right
0:21:24	it's pretty
0:21:25	both
0:21:26	it just
0:21:27	define sees the right kind of video thing
0:21:31	this is the like
0:21:33	right
0:21:33	right on
0:21:34	yeah
0:21:35	um
0:21:36	we
0:21:37	of it
0:21:39	but this is simple
0:21:40	but he was
0:21:41	useful
0:21:42	do they
0:21:42	six
0:21:42	speech
0:21:43	because
0:21:44	p2p anything from a challenge
0:21:49	and how
0:21:50	or was
0:21:50	this project we switch out
0:21:52	it's more of a spy
0:21:53	i
0:21:55	that's
0:21:56	and also things expedia the unit selection
0:21:58	and have
0:21:59	john
0:22:00	trajectories
0:22:01	uh
0:22:01	uh
0:22:02	change from point
0:22:03	which is that
0:22:04	data is
0:22:05	yeah i yeah
0:22:06	i
0:22:07	but
0:22:08	and tedious and it can be a speeding is
0:22:11	included
0:22:12	some global time but is from all this
0:22:14	by
0:22:15	the
0:22:16	with
0:22:16	kind of
0:22:17	for what some of them
0:22:18	effect
0:22:19	both
0:22:19	project for the
0:22:20	in fact
0:22:21	this
0:22:22	um this is that is that
0:22:24	average five year
0:22:25	we do right sure
0:22:26	human speech
0:22:28	i think it
0:22:29	a few months
0:22:30	um
0:22:31	the same one
0:22:31	well
0:22:32	speech
0:22:34	and it
0:22:35	if if angel
0:22:36	that's okay
0:22:37	speech
0:22:37	and you can be
0:22:39	they have
0:22:40	quite
0:22:40	all brought up
0:22:41	and therefore
0:22:43	this measure
0:22:44	no longer robust
0:22:45	you know
0:22:46	fig
0:22:47	speech
0:22:48	cool
0:22:48	they yeah
0:22:49	it ended
0:22:50	this
0:22:51	and
0:22:53	uh
0:22:54	because i
0:22:55	yes uh well you know it
0:22:57	because
0:22:58	in speech patterns to use
0:23:00	if the school or
0:23:02	six
0:23:02	speech
0:23:03	maybe
0:23:04	okay
0:23:05	fictions
0:23:06	uh based in p2p humour
0:23:09	speech
0:23:09	so we sort
0:23:11	it might be possible to save in p2p you
0:23:13	fig speech
0:23:14	yeah what they like
0:23:15	it's all
0:23:16	um
0:23:17	we
0:23:18	p2p up to a month
0:23:20	yeah marcy
0:23:21	it it E G
0:23:22	okay i'm up for it
0:23:23	yeah
0:23:24	oh
0:23:24	um evaluate
0:23:26	well be right
0:23:27	human speech
0:23:28	um
0:23:29	fig
0:23:29	it
0:23:30	um
0:23:31	this is the weather right
0:23:32	this is a
0:23:33	yeah there are a
0:23:35	as you can
0:23:36	the
0:23:37	we tested it
0:23:38	fig speech
0:23:39	was found to have
0:23:40	data where there are a few
0:23:43	or both
0:23:43	grammar
0:23:45	a few months while they're writing about
0:23:47	involved in
0:23:47	for the first six speech where they just say well
0:23:51	in it means that
0:23:52	if you go
0:23:52	grammar
0:23:54	yeah there are huge differences
0:23:56	uh
0:23:57	and then
0:23:58	this
0:23:58	it's too
0:23:59	even for the adaptation data is
0:24:01	just
0:24:02	one me
0:24:02	speech today
0:24:04	so
0:24:04	it is not i
0:24:06	you yeah
0:24:07	what they write
0:24:07	is that
0:24:08	fig
0:24:09	fig
0:24:15	um
0:24:15	i to summarise my talk
0:24:17	um
0:24:18	this but
0:24:19	the extent
0:24:20	almost
0:24:20	speaker verification
0:24:21	yeah
0:24:22	yeah
0:24:23	speaker age and it
0:24:24	because i didn't
0:24:26	speech
0:24:26	yeah
0:24:27	i got that
0:24:27	a channel
0:24:28	yeah
0:24:28	this
0:24:29	something
0:24:30	school it's tedious
0:24:31	it's high enough of these
0:24:33	inside was
0:24:34	possible
0:24:34	to the human right
0:24:36	this thing brought it
0:24:37	the speech data available
0:24:40	i guess
0:24:41	can be
0:24:42	you import
0:24:43	speaker verification
0:24:44	this can
0:24:44	in
0:24:45	i don't know how many
0:24:47	well i guess
0:24:47	but
0:24:49	or support because
0:24:51	oh
0:24:51	it is
0:24:52	impostors
0:24:53	okay
0:24:54	fig
0:24:55	yeah
0:24:55	and then i'll mention a missile
0:24:57	you think
0:24:59	uh commissioning but yes i hear it
0:25:01	or what they write
0:25:03	fig
0:25:04	fig
0:25:04	what
0:25:05	no
0:25:05	moreover
0:25:06	robust
0:25:07	no
0:25:09	but
0:25:10	yeah but it is
0:25:10	this you know security issues
0:25:12	we
0:25:13	and we like to do these
0:25:14	this
0:25:15	voice going
0:25:16	speaker adaptation
0:25:17	two
0:25:19	for free or on the way
0:25:21	right right
0:25:22	provides a base
0:25:23	what's going on
0:25:27	well but you don't know why
0:25:29	um
0:25:30	so
0:25:30	this technique
0:25:32	um
0:25:33	um
0:25:34	we have about them and
0:25:35	from all speakers
0:25:36	what
0:25:37	and
0:25:39	so
0:25:39	national
0:25:40	in it is not his fantasies you please
0:25:43	and uh i
0:25:44	and you like to
0:25:45	it's hard
0:25:47	that's you you can
0:25:49	because this technique
0:25:50	can cool
0:25:52	people's
0:25:52	has
0:25:53	yeah
0:25:54	talking some T Vs
0:25:56	cool
0:25:56	sample
0:25:57	and you want to use
0:25:59	we have
0:26:00	right
0:26:01	be
0:26:01	um just techniques
0:26:02	can
0:26:03	because
0:26:04	welcome
0:26:04	hoping
0:26:05	someone
0:26:06	that's just
0:26:06	because voiced and use the voice
0:26:09	um we can associate with
0:26:10	they are embedded devices
0:26:13	that's
0:26:14	voice
0:26:14	indication eight
0:26:17	so
0:26:18	yeah that was
0:26:18	we need
0:26:19	they do you future
0:26:21	but it is
0:26:22	the screen
0:26:23	voice
0:26:24	since it
0:26:24	voice
0:26:25	and he must
0:26:26	oh
0:26:26	oh
0:26:27	that um this
0:26:31	that's all
0:26:37	right
0:26:38	presentation
0:26:40	uh
0:26:40	we should
0:26:45	oh
0:26:46	so
0:26:49	oh
0:26:50	or
0:26:52	four
0:26:53	sure
0:26:57	oh
0:26:58	uh
0:27:01	with
0:27:02	oh you do
0:27:05	oh
0:27:06	right
0:27:07	which
0:27:09	hmmm
0:27:10	so
0:27:11	but
0:27:13	oh
0:27:14	replica guns working on speech transmission to
0:27:21	oh
0:27:25	your your
0:27:26	yes
0:27:29	which
0:27:31	i see
0:27:32	but i
0:27:34	ninety percent of the voices box that accent
0:27:37	so even speaker verification
0:27:39	cranes
0:27:40	to start with
0:27:41	sure uh
0:27:42	identical people
0:27:44	well i think of puzzles
0:27:45	we have to that
0:27:46	the uh
0:27:48	to speech
0:27:48	one
0:27:50	ooh
0:27:52	mark
0:27:55	oh
0:27:55	four
0:27:56	yeah
0:27:58	oh
0:28:01	hmmm
0:28:01	oh
0:28:03	for a moment
0:28:06	we
0:28:08	um
0:28:10	or
0:28:13	oh
0:28:14	or
0:28:15	but
0:28:18	oh
0:28:19	well when we were different circumstance
0:28:24	yeah
0:28:28	oh
0:28:28	oh
0:28:30	oh
0:28:31	sorry
0:28:32	sure
0:28:33	true
0:28:34	oh
0:28:36	oh
0:28:37	well if we
0:28:39	yeah
0:28:41	for the money
0:28:42	would be to model
0:28:45	i'm not like
0:28:47	uh_huh drawn from from uh yeah that too
0:28:52	well
0:28:52	right
0:28:53	one
0:28:54	actually going on
0:28:57	we we do
0:28:59	um
0:29:00	right
0:29:01	perhaps a big challenge
0:29:04	yeah i think that's that's the that's the crystal
0:29:06	this
0:29:07	see fig speech maybe they can variables in doing that
0:29:10	right
0:29:11	hmmm
0:29:18	uh_huh
0:29:20	okay
0:29:20	i am i
0:29:22	we have some similar work and
0:29:25	i
0:29:26	some
0:29:27	paper
0:29:27	so there we also
0:29:29	um
0:29:30	right
0:29:30	um
0:29:31	yeah then fine yeah
0:29:32	um see signs
0:29:34	oh
0:29:34	transform tonight
0:29:36	basically to intermediate
0:29:37	speech will be
0:29:38	back to the speaker identification system
0:29:41	so um
0:29:42	also i don't know
0:29:44	what street journal
0:29:45	in a nice
0:29:46	and then you so we can
0:29:49	and
0:29:50	and you had to to to type
0:29:53	and speaker identities instead
0:29:55	right
0:29:55	based on like
0:29:56	ubm agenda like using that low level
0:29:59	acoustic features
0:30:01	in the other one is
0:30:03	a novel speaker identification system
0:30:05	and
0:30:06	such as no phonetic
0:30:07	that
0:30:08	right
0:30:09	so what we are used and
0:30:10	that um
0:30:12	and they generate
0:30:13	generated
0:30:14	the
0:30:14	it is and
0:30:15	and
0:30:17	i think that now and a novel feature based speaker identification
0:30:21	hmmm
0:30:22	small
0:30:22	one double
0:30:23	hmmm
0:30:25	well
0:30:26	well whatever bottleneck and
0:30:28	and you really to be selected by now
0:30:30	generative
0:30:32	i
0:30:32	um but yeah
0:30:34	looks like a high level
0:30:35	yeah
0:30:36	speaker I D's
0:30:37	i didn't use instant
0:30:38	it's not
0:30:40	make
0:30:40	it's
0:30:40	not robust
0:30:42	okay
0:30:42	at low levels
0:30:44	so like
0:30:44	mm
0:30:46	they stand for
0:30:47	and just got in speech you
0:30:50	yes
0:30:50	right and then
0:30:51	it looks like and
0:30:53	hmmm
0:30:53	there you can
0:30:55	do you like
0:30:56	a i mean
0:30:57	i yeah
0:30:58	this
0:30:58	p2p
0:30:59	a speech
0:31:00	reason or
0:31:02	no it's not
0:31:02	yeah
0:31:03	so probably
0:31:04	um
0:31:06	and that's what you have done
0:31:08	yeah experiments also you see
0:31:10	and at that time
0:31:11	speaker verification system using
0:31:13	try using a novel
0:31:15	and features
0:31:16	yeah
0:31:16	yeah
0:31:17	temporal features mike
0:31:18	not only not long range and speed
0:31:21	make some characteristics
0:31:23	probably
0:31:23	and now be
0:31:25	more robust against that
0:31:26	that generated
0:31:28	speech
0:31:29	so basically
0:31:30	and
0:31:30	so
0:31:31	hmmm the speaker I D C
0:31:33	that was
0:31:34	transformation all three
0:31:36	yeah
0:31:37	can be too
0:31:39	yeah
0:31:41	we see each other
0:31:42	and then something to do with this
0:31:44	uh_huh
0:31:45	D C
0:31:46	so
0:31:47	yeah we can probably also borrow
0:31:49	yeah
0:31:50	symphonies
0:31:51	um speech since it's it's uh
0:31:53	jenny generation you know
0:31:56	yeah
0:31:56	try to make
0:31:57	speaker and
0:31:58	so
0:31:59	but
0:32:00	yes and and and probably
0:32:01	and now also on how expensive
0:32:03	where is it is fine to use
0:32:04	for the speech thing is is that you probably
0:32:07	two
0:32:08	okay normally
0:32:09	and
0:32:10	teachers are
0:32:12	that's
0:32:12	right
0:32:13	sure
0:32:14	yeah
0:32:14	yeah it is
0:32:17	yeah
0:32:18	okay
0:32:18	so no time i i just got my question
0:32:22	uh no no no no no
0:32:24	you you you you you you and uh yeah that should be used them on the same
0:32:31	what would happen if you change the
0:32:34	yeah
0:32:37	so uh i
0:32:38	questions um
0:32:39	we use you know
0:32:41	gmmubm systems and svm
0:32:45	with
0:32:45	you know caution
0:32:47	it's a contest
0:32:48	um but we haven't ones that in a long time
0:32:51	future
0:32:53	yeah it's real time values
0:32:55	um
0:32:57	but
0:32:57	um
0:33:00	um
0:33:00	but
0:33:02	uh_huh
0:33:07	we have a new features
0:33:09	um we have one
0:33:10	you huge
0:33:11	which is we
0:33:11	really
0:33:12	one
0:33:13	so i reassured that is uh
0:33:15	right next
0:33:16	i guess
0:33:17	next
0:33:18	with bonds and you
0:33:20	yeah
0:33:21	that's not a long time
0:33:22	yeah
0:33:24	right
0:33:26	right
0:33:27	yeah

Evaluation of the Vulnerability of Speaker Verification to Synthetic Speech

SESSION 7: Speaker and Language recognition - Evaluations and performance testing

Přidáno: 14. 7. 2010 11:08, Autor: Phillip DeLeon (New Mexico State University), Michael Pucher (Telecommunications Research Center (FTW)), Junichi Yamagishi (University of Edinburgh), Délka: 0:33:36