Přepis řeči - Multiple Background Models for Speaker Verification

0:00:07	i have a
0:00:08	no i mcgill or talk about that
0:00:11	not all background model
0:00:13	well speaker verification
0:00:14	and the author of this paper
0:00:17	they are
0:00:18	we found the assumption and the downhill
0:00:21	and uh and a friend of which are down
0:00:23	and the assumption
0:00:25	my name is on the other hand i you would
0:00:27	talk
0:00:28	uh
0:00:28	that
0:00:29	this
0:00:30	um
0:00:31	the idea i mean these people are used to run into a single i think all for you
0:00:36	and
0:00:37	then
0:00:37	yeah
0:00:38	clean up the only
0:00:39	no
0:00:40	being
0:00:41	yeah this is the one that is having
0:00:43	cable
0:00:43	first introduction
0:00:45	in this introduction
0:00:46	i will
0:00:47	uh in
0:00:48	why we propose gift ideas and bases also our motivation
0:00:53	second
0:00:54	we really nice okay that vocal tract yeah school to speaker recognition
0:00:59	and then to to uh idea
0:01:01	when we do some
0:01:03	and then
0:01:03	and this
0:01:05	we can fix
0:01:06	parenthood
0:01:07	that ah
0:01:08	finally
0:01:09	we can raise the ticket
0:01:10	and and the multiple background models are proposed
0:01:14	and then
0:01:15	no
0:01:15	i have a cable components
0:01:19	first
0:01:21	got a mixture model
0:01:23	mixture model when was so background model
0:01:25	is that because speaker or occasions this term
0:01:28	is that way
0:01:30	forty
0:01:30	quality of that case
0:01:32	the
0:01:32	they all thought
0:01:33	this term
0:01:34	such as
0:01:35	but i do not see it
0:01:37	and the whole thing
0:01:37	attribute projection
0:01:39	is based on
0:01:40	um
0:01:41	uh
0:01:42	but that but that the
0:01:43	the
0:01:44	at that
0:01:45	gmm
0:01:46	you'll be an ad is basic
0:01:48	structure
0:01:49	and the
0:01:50	the most important
0:01:52	rolling in this
0:01:53	in this this
0:01:54	this
0:01:54	basics but
0:01:55	strong is the
0:01:56	you'll be an
0:01:57	and that we think are a complete ubm is supposed to
0:02:02	right yeah that the speaker independent feature distribution
0:02:06	and uh there are no man starts to counting the quantity of ubm
0:02:10	first
0:02:11	which we are
0:02:12	and misc data
0:02:14	and the means we pay for all the data better
0:02:16	two three but when it was so yeah
0:02:18	as a second
0:02:19	right right
0:02:20	gender
0:02:21	oh channel
0:02:22	and then
0:02:23	but
0:02:23	ubm
0:02:25	but uh
0:02:26	there may be
0:02:27	uh
0:02:28	there
0:02:28	there
0:02:29	there are other approaches
0:02:32	first that
0:02:33	okay
0:02:34	well okay yeah
0:02:35	the speaker our unity expensive i might have back
0:02:40	such that
0:02:41	speech rate
0:02:42	speech what we'll
0:02:43	emotion
0:02:44	well collector and the song
0:02:46	but the major differences between the speaker
0:02:48	is due to the difference
0:02:50	between their average week yeah
0:02:52	so in speech recognition
0:02:54	well we'll check yes no medication
0:02:56	is
0:02:57	is also used that to obtain
0:02:59	speaker independence insurance
0:03:04	now here is like
0:03:06	uh
0:03:07	you're only kills the frequency warping function
0:03:10	it's the crew though
0:03:11	this is the original frequency and bases there
0:03:15	what the frequency
0:03:16	and this is what you that current
0:03:18	we want
0:03:19	and that is what vector
0:03:21	is that when we want to
0:03:23	they want to get
0:03:25	but unfortunately
0:03:26	there is no closed expression for days but still
0:03:32	we use
0:03:32	this
0:03:33	this
0:03:34	great
0:03:34	to get
0:03:36	okay yes
0:03:38	and that this is what the speech
0:03:40	what what the features
0:03:41	is that what models
0:03:43	and then the rate of this value
0:03:45	is that they are
0:03:47	zero point
0:03:48	eight
0:03:48	eighty two
0:03:49	one point yeah
0:03:51	well to waste that's that's
0:03:53	zero
0:03:54	point
0:03:54	zero eight
0:03:55	zero two
0:04:01	now we look at
0:04:02	i would think that
0:04:03	her mental state at the paper
0:04:05	okay
0:04:05	parents were having a a
0:04:08	yeah
0:04:08	i thought i was on a six
0:04:10	corpora encode past
0:04:11	condition
0:04:12	and in crosschannel conditions
0:04:15	that you'll be answering data were selected
0:04:17	from use
0:04:18	S I to solve
0:04:19	four
0:04:20	one side
0:04:21	there are about
0:04:22	sixties
0:04:23	and what it and the sixteen afternoon
0:04:25	and that
0:04:26	as i see two thousand and three at
0:04:28	as i tucson and two corpora
0:04:30	yeah about five hundred utterances
0:04:34	notice the feature where you mean you
0:04:36	and then a fifty
0:04:37	sexual mean subtraction feature what you know they're accepted
0:04:41	acceleration
0:04:42	and that right there
0:04:43	i use the
0:04:44	so we come back yeah
0:04:46	the feature weights that if you choose dimension
0:04:48	then hlda is used
0:04:51	uh the final
0:04:52	dimension of the feature you starting now
0:04:57	oh
0:04:58	this the finger of readout distribution
0:05:01	we present this encourage us
0:05:03	not want to
0:05:04	industry that the difference between male and female
0:05:07	we want to
0:05:09	focus on the S
0:05:11	it's not as bad
0:05:12	uh
0:05:13	the wedding is
0:05:15	uh he's
0:05:16	is that
0:05:17	uh the value range from this
0:05:18	if we used this value to it wide
0:05:21	paper
0:05:22	two three will be a
0:05:23	may be
0:05:24	maybe we can get that arnold yeah
0:05:29	so
0:05:31	that it has that much attention
0:05:33	we might need ubm training they turned into a to use pointed it here
0:05:37	says holding to the warping factor for example
0:05:41	uh database
0:05:42	first the
0:05:43	the walking factories
0:05:45	zero point eight eight
0:05:47	we have one hundred and the
0:05:49	it's it's three utterances
0:05:54	no this is the
0:05:55	whole structure of our proposed multiple background model
0:06:00	uh i think that this structure is that right
0:06:04	gmm you yeah
0:06:06	uh
0:06:06	i'm the think different these days
0:06:09	new gmm ubm
0:06:11	there is only one ubm
0:06:13	and then
0:06:14	in this way how ubm who
0:06:17	you're you
0:06:18	map adaptation
0:06:20	each ubm is adapted to generate
0:06:23	uh unique
0:06:24	a speaker model
0:06:26	and they
0:06:27	ubm and the speaker models have warmed up here
0:06:30	the only in the test the framework
0:06:32	and i is used for
0:06:35	yeah ubm and the speaker
0:06:37	G G M and to
0:06:39	table
0:06:39	the
0:06:40	but phonetically
0:06:41	we shall score
0:06:46	that
0:06:46	well that's what kinda stuff
0:06:48	results
0:06:49	baseline baseline performance
0:06:51	and that
0:06:52	uh
0:06:53	wait you gender independent yeah and ubm sister
0:06:58	uh
0:06:59	the eer for the forecast conditions are about ten percent
0:07:07	and then
0:07:08	wait while the
0:07:09	and by the data
0:07:10	you show and the
0:07:12	where you gender dependent
0:07:14	ubm
0:07:19	the results
0:07:20	if
0:07:21	if the gender all pool ubm an agenda of confusion unmatched
0:07:26	then
0:07:27	the performance
0:07:28	i'm not be improved
0:07:30	but
0:07:30	if the cross gender confusion
0:07:33	just contact dave
0:07:34	and the days
0:07:35	we can
0:07:36	how were bad
0:07:37	without
0:07:39	now this is that we have
0:07:42	dependent ubm
0:07:44	problem
0:07:44	it's table we can see that
0:07:46	four female condition
0:07:48	ubm to game
0:07:49	the bastards are
0:07:52	and a four male can you should ubm six
0:07:54	scale the best result
0:08:02	not have a good some performance
0:08:04	comparing that you'll be able to resolve for female conditions
0:08:08	and the ubm six results
0:08:10	for many conditions admits that it's not
0:08:12	we can find that i
0:08:13	are you yeah maze
0:08:15	for an aspect and will so that get the training data
0:08:18	that are contained in the back the performance
0:08:21	in the ubm with
0:08:22	all the training before
0:08:26	now that's reached into his finger again
0:08:28	uh
0:08:30	we can get or
0:08:31	a lot of space
0:08:33	but there is wise enough to have his if you will
0:08:40	for a test utterance
0:08:41	which we are hampered you also but
0:08:43	okay
0:08:44	racial or just connect the ends
0:08:47	and B M can obtain a score vector
0:08:50	we can use
0:08:51	coffee remastered
0:08:52	to obtain the final results
0:08:54	when we talk about you and and the contributions of it all in singapore
0:08:59	vocal
0:08:59	and that's really
0:09:01	uh powerful tools
0:09:02	back to
0:09:04	but here we just want to
0:09:06	some simple
0:09:08	and the
0:09:09	a simple
0:09:10	simple
0:09:11	uh
0:09:12	buster
0:09:13	first
0:09:15	have a mismatch right
0:09:17	uh we just
0:09:18	you have a very
0:09:19	but uh
0:09:20	you look at this thing right
0:09:22	the results
0:09:23	it's not very good
0:09:27	and then we'll back
0:09:29	maximum
0:09:30	exactly what the master
0:09:32	and we use
0:09:33	the ubm
0:09:34	which
0:09:35	do not report is the max
0:09:37	as the final score
0:09:39	but
0:09:40	the
0:09:40	yeah we can discuss this also
0:09:45	and that's the way your minimum that you were the racial master
0:09:48	and the you know gives the best results
0:09:51	model based remastered
0:10:01	and then there is the question arises why the minimum
0:10:04	yeah the racial matters K with the best result
0:10:07	unfortunately
0:10:09	wait i'm
0:10:09	know the exact reason
0:10:11	uh
0:10:13	in
0:10:13	intuitively that
0:10:14	peak
0:10:15	we we we we try to do uh
0:10:18	ah
0:10:18	combination
0:10:19	in
0:10:20	jokingly that speaker
0:10:22	yeah and i'm actually
0:10:23	for the and the the you yeah you would wear both
0:10:26	increase if what match
0:10:28	has utterance is in court
0:10:30	and this is just which transmits meetings bother with no the reason
0:10:34	uh
0:10:35	we can make it
0:10:37	uh
0:10:38	the
0:10:39	means and the standard
0:10:41	every iteration of that
0:10:42	that
0:10:43	oh
0:10:43	that we would rituals
0:10:45	oh
0:10:45	well as i tucson has six
0:10:47	which each
0:10:48	yeah
0:10:49	and we put a thinker
0:10:51	just like this
0:10:53	uh i have to say that
0:10:55	it's finger is not
0:10:56	the reason of this
0:10:58	send
0:10:59	of this intense
0:11:00	uh
0:11:02	we just want to know why
0:11:09	now
0:11:09	we you know the components
0:11:11	ah in this paper with was to investigate here
0:11:14	the week yeah that's the right term
0:11:17	for ubm training the interesting action
0:11:20	experiment
0:11:21	short time
0:11:22	that you'll be actually them is about you
0:11:24	new media data with battered in the ubm trend with all that they are
0:11:29	based on this finding
0:11:30	we further propose a multiple background model system
0:11:34	yeah right
0:11:35	you take multiple speaker gmm and ubm yeah
0:11:38	for speaker recognition
0:11:41	uh through minimum
0:11:43	now we we shall feel with the proposed master
0:11:46	and improve the performance
0:11:48	i'm used to be
0:11:50	but yeah
0:11:51	you're right
0:11:51	open questions
0:11:53	what the minimum that we would we show master gave the best results
0:11:56	it's just locally experience
0:11:58	uh we will be posting the slow but
0:12:02	the property is under investigation
0:12:04	well techniques to improve the state of
0:12:08	uh standard of that
0:12:09	this term
0:12:10	uh for example if you
0:12:12	the yeah the system and the
0:12:15	yeah system
0:12:16	you are you yeah
0:12:17	the performance
0:12:18	yeah
0:12:19	cool
0:12:20	uh we know and the way people
0:12:22	to the expression
0:12:23	experiment
0:12:24	how about that
0:12:26	computational cost the and the bases another
0:12:29	sure
0:12:31	finally
0:12:32	i just talking about
0:12:40	you have plenty of time for questions
0:12:45	there were so
0:12:59	hmmm
0:13:00	oh
0:13:01	i'm sorry it wasn't clear to me exactly how you choosing the
0:13:05	you have multiple unions and we'll selecting what
0:13:08	within each ubm
0:13:10	you mean the
0:13:11	two semester
0:13:13	you just you have multiple units which are built a little different datasets to uh address
0:13:18	how do you decide what
0:13:20	when
0:13:20	each year
0:13:22	uh
0:13:23	yeah
0:13:26	it's cigarettes
0:13:27	the what factors is different
0:13:30	and the
0:13:31	we use the
0:13:32	if the
0:13:34	hmmm
0:13:35	for example that if the
0:13:37	uh warping factor is that the airport mine then
0:13:41	this is the
0:13:42	with some back this paper
0:13:44	to train the ubm
0:13:46	when there is a website
0:13:47	well it does that and this and that it is
0:13:50	yeah
0:13:51	what during the enrolment and the test the
0:13:53	and the
0:13:54	the uh
0:13:55	we can have
0:13:56	oh the tree data and test data and not extracted
0:14:00	they are just the scroll back
0:14:02	a two
0:14:02	uh
0:14:03	ubm and the speaker gmm pattern
0:14:11	so i'm just going to be a a synthesis using the all the user density and the
0:14:16	just to use
0:14:17	they were easy to combine this discourse is used
0:14:22	two
0:14:22	to school too
0:14:23	if i noticed like usual as well
0:14:26	you you
0:14:29	remote questions
0:14:39	uh
0:14:40	hmmm
0:14:41	uh you know where you use
0:14:43	yeah
0:14:44	did you
0:14:45	for speaker
0:14:46	no
0:14:47	i mean you mean how many of them is
0:14:50	maybe if you assume that you are using
0:14:54	you you
0:14:55	you did
0:14:56	you
0:14:58	yeah
0:15:00	you know
0:15:01	yes
0:15:02	i know that's
0:15:03	question
0:15:03	yeah in fact there are really many mass transit to
0:15:07	uh so get data and the
0:15:09	i think of like yeah it's just a way of them
0:15:12	which has
0:15:13	just well then
0:15:14	and they are
0:15:15	ah
0:15:16	true
0:15:16	yeah me either
0:15:18	right
0:15:19	it would be
0:15:28	and that's where one last question you
0:15:32	yeah
0:15:32	hello
0:15:33	i just one question um
0:15:36	you just use the vocal tract length normalisation to select
0:15:40	the population for the building the different buttons models
0:15:43	and you can also use the acoustically
0:15:46	to to select
0:15:48	appropriate calls for normalisation
0:15:50	looking for speakers
0:15:52	which uh
0:15:53	more close to the
0:15:55	actually the speaker
0:15:56	how do you own it is pretty maybe comparing
0:15:58	using different but remotes or different population for
0:16:02	normalisation with
0:16:03	we propose from it
0:16:07	i can't
0:16:08	i think
0:16:09	it's a
0:16:09	i'm i'm just
0:16:10	the asking these uh
0:16:12	it's do you uh do you have
0:16:14	this is also the vocal tract length normalisation
0:16:17	to select
0:16:18	a cohort
0:16:19	six
0:16:20	of the speaker
0:16:21	for normalisation of the school
0:16:24	yeah
0:16:25	you mean
0:16:27	i don't know
0:16:28	uh
0:16:29	i don't know what you
0:16:30	good
0:16:31	right
0:16:31	the same
0:16:32	that would be
0:16:33	probably
0:16:35	okay yeah she's didn't uh uh we don't want to just the framing of will or will not too close
0:16:41	and uh if you
0:16:42	yeah
0:16:42	very nice
0:16:43	did you do that
0:16:44	for for you
0:16:45	would you only
0:16:46	yeah so you remotes and we move the twos
0:16:49	that's because figure

Multiple Background Models for Speaker Verification

SESSION 3: Background modeling in Speaker recognition, Forensics

Přidáno: 14. 7. 2010 11:08, Autor: Wei-Qiang Zhang, Yuxiang Shan, Jia Liu (Tsinghua University), Délka: 0:16:58