Speech Transcript - Subband Modeling for Spoofing Detection in Automatic Speaker Verification

0:00:16	hello everyone
0:00:18	my name is what's industry and i'm gonna present a people
0:00:22	i feel
0:00:23	so and modeling or splitting detection
0:00:25	you know make speaker verification
0:00:29	so in the next few flavours
0:00:32	of like to do with five from one
0:00:35	speaker
0:00:37	can you three
0:00:38	and compromises
0:00:41	so and automatic speaker verification system
0:00:45	instead of verifying the identity of james
0:00:48	so in this paper
0:00:50	options
0:00:52	a more distant from various it's a
0:00:57	and this speaker
0:01:00	you know the fine with the whole speech comes from the same speaker that the
0:01:04	case
0:01:04	the system accepts the test speech that can speaker
0:01:08	otherwise the system
0:01:11	as an enforced
0:01:12	so common application of is very good
0:01:15	for user authentication for example in by classification
0:01:23	recent studies have shown
0:01:25	the systems are not
0:01:28	a lot of the listings of
0:01:31	which can be
0:01:33	from various research and
0:01:37	one of the using synthesis has
0:01:40	and synthetic thing
0:01:42	speaker
0:01:44	second with involves and voice conversion have
0:01:50	to generate the speech of speaker and the colour from the track involved
0:01:58	replay attacks which involves link back
0:02:01	are created by means of a speaker
0:02:04	and for on the back in one
0:02:07	if question at the voice of target or not
0:02:11	a given x is then used system
0:02:14	so in this paper
0:02:16	or focuses on
0:02:18	replay attacks
0:02:20	or two main reasons
0:02:21	one
0:02:22	because of its simplicity
0:02:25	this formal that doesn't require any studies in seen any
0:02:30	a signal processing more speech data
0:02:33	second
0:02:34	the formal that is
0:02:36	i difficult to detect from the results because we are actually feeling that real speech
0:02:43	so it doesn't use any
0:02:44	a complete the algorithm the general speech
0:02:47	so this and ones for all target speaker
0:02:54	so
0:02:55	spoofing on to missus are primarily developed
0:02:58	two
0:02:59	based on dsps
0:03:01	so it really very fine control
0:03:04	video
0:03:07	so and i four think problem it can be viewed as a binary classifier which
0:03:12	means that discriminately with the input speech
0:03:16	is a non speech
0:03:18	or is supposed speech
0:03:21	so in our case when we talk about this posting on the major
0:03:25	so what exactly so there's a them in the
0:03:29	so but not in five
0:03:31	what we factors interest you know
0:03:33	independently back speech
0:03:36	so essentially you know case doesn't be
0:03:40	different channel characteristics
0:03:41	in this during playback and we recording speech
0:03:45	and also
0:03:46	background noise
0:03:49	involve doing this you are
0:03:51	parameters are essentially that you that are we expect model exploiting for playback speech detection
0:03:59	before i start
0:04:01	you know holes sub band modeling framework
0:04:04	for
0:04:07	like of all a sum of common commonly used a approaches towards designing us
0:04:13	supposing images
0:04:14	which since also
0:04:16	exploiting information of course on
0:04:20	so what we call is from one of them is a
0:04:23	for the on here is the
0:04:25	is an example of four
0:04:27	optimal or
0:04:29	gaussian mixture model which is a generative model
0:04:32	as can be seen
0:04:34	of given from the speech prince common spectrum is extracted
0:04:38	on its extractor
0:04:39	acoustic features
0:04:40	for example to use is you know is very low baseline feature
0:04:45	and on this speaker she's data
0:04:48	basically
0:04:49	a gmm models one for school and one on my
0:04:52	utterances are train
0:04:54	you not of the model the distribution of this
0:04:57	this process
0:04:58	so this is an example of commonly used gmm model it has shown promising results
0:05:03	in the st score evaluations
0:05:07	category all a total ban on committees that you're gonna study
0:05:11	is
0:05:12	discriminatively trained unit it's
0:05:15	which on very
0:05:17	problems on and in the thing
0:05:21	nineteen this peaceful nations
0:05:24	so as can be seen so this model takes for quantified in speech or its
0:05:29	and basically going on this isn't wrong three
0:05:33	between the one
0:05:35	so this actually motivates our current work
0:05:38	and in fact that means just question that motivates our work is
0:05:44	do we really need
0:05:45	all the speech
0:05:47	all the informational costs all the three best friends
0:05:51	already split into action
0:05:52	so or intuition is that may be doing
0:05:55	maybe they next speech to use might be somewhere in high frequency regions or maybe
0:06:00	you know very low frequency regions
0:06:02	so in this framework of what we what defined with different understand the importance of
0:06:07	different frequency bands for sporting detection
0:06:11	and also trying to come up of s o s and combination that improved if
0:06:16	the columns
0:06:18	so we test this hypothesis this idea of old remote on two benchmarks
0:06:26	this is feasible thing to seventeen and
0:06:28	is useful information
0:06:34	so this texas also proposed methodology
0:06:37	can be expected to us
0:06:39	in the first one with we basically
0:06:42	if or input into
0:06:44	i
0:06:46	spectrogram of different
0:06:49	and train and independency inance on top of the frame discriminative
0:06:56	and likewise what we obtain is we have been and dependency in that are trained
0:07:00	on and sum of spectrograms
0:07:02	so
0:07:03	this allows of actually look at
0:07:05	or indicate and sometimes
0:07:07	and understand which frequency regions are more discriminant e
0:07:10	of course for detection on this data sets
0:07:14	so
0:07:15	so this can also be viewed as a employing and independent workers
0:07:19	just focusing on
0:07:21	small sub band information and trying to and trying to exploit discriminative information for incorrectly
0:07:26	and a one speech
0:07:28	rather than having once you in which
0:07:30	it hasn't is the baseline once you know that is at least having to look
0:07:36	all the information and also
0:07:38	how to
0:07:39	if the bulletin of are on by means an introduction and while retaining the
0:07:44	most discriminative information what we do we begin by the fast basically be a we
0:07:48	kind of
0:07:51	allow independence into focus on one-pass
0:07:54	and then
0:07:54	we into a
0:07:56	improve performance by doing so
0:07:58	so without in the second step what
0:08:01	we and
0:08:02	basically think this prevent models and combine this alone features
0:08:06	and train or another classifier on top of so many all this
0:08:11	different features and then jointly update the weights of entire framework
0:08:16	so in it what we do
0:08:18	we are also making use of the cold information but and
0:08:22	not using only once you
0:08:23	you in
0:08:25	in
0:08:28	any given in c n n's and then we give this and independent features on
0:08:32	cognitive and then than if we train
0:08:35	the classifier is trained on
0:08:38	so this is our proposed method which we test on the two it's really seventeen
0:08:43	and twenty nineteen is possible data
0:08:48	so
0:08:49	let me let me talk about the experimental results well start with the baseline
0:08:55	baseline gmm and the ceiling
0:08:57	so the gmm baseline is trained on stick uses it is
0:09:02	and
0:09:03	the c n is trained on spectrogram and spectrogram
0:09:07	condition from the experimental results
0:09:09	we find that the discriminatively trained union
0:09:14	in performance
0:09:16	two
0:09:17	sufficiency handcrafted gmm baseline
0:09:21	on the data six
0:09:23	and one thing the of the on the training dataset to increase equal error rate
0:09:28	on gmm is given the fact that we apply
0:09:32	preprocessing step
0:09:34	on the audio signals
0:09:36	which involves discarding the cost of p zero value of silence
0:09:40	from evidence only not prior work in interspeech trained and
0:09:45	and that's the reason why our baseline security gmm is quite
0:09:50	different and then the
0:09:52	baseline and the for this effect
0:09:56	so no let me talk about our first experiment is okay
0:10:00	where we
0:10:01	spectral input into uniform sometimes
0:10:04	and one thing to more here is that
0:10:06	in this paper we have adopted a very simple to extract is your are uniformly
0:10:11	segmenting the input and have non-overlapping
0:10:16	so we just in the in between two
0:10:19	six
0:10:20	we
0:10:21	we
0:10:22	frame independence in and one of two in
0:10:26	and then having frame that model was intensely from we combine them to train a
0:10:30	concatenated model or a joint model
0:10:34	so i would be calling disjoint framework task at in the experimental results
0:10:39	which basically is that
0:10:41	the framework of this except that we trained
0:10:44	pictures and classifier is trained one
0:10:48	so let's look at the experimental results
0:10:52	as can be seen from this
0:10:55	right distribution
0:10:57	on the extended data set we find that the commission in high frequency reasons seem
0:11:03	to be more likely in contrast to if you regions
0:11:07	on the contrary
0:11:09	in between the same reasoning can be the same define objects
0:11:13	we see people
0:11:14	in c reasons be more discriminative in contrast the high frequency region
0:11:20	nonetheless on what the datasets our proposed
0:11:24	of frame more difficult compare
0:11:26	the on off between models anymore
0:11:29	since all four improve performance on what the data set
0:11:33	so in this section in the second experimental setup what we do not is you
0:11:37	know for that is there or input
0:11:39	into four
0:11:41	uniform segments each of the two sub bands
0:11:44	no with i
0:11:45	we can look at
0:11:47	more detail on discriminative information rather than just having a
0:11:52	for us so i
0:11:55	so it is our rows or and experimental setup so we have or independency anything
0:12:02	on two khz and a nation
0:12:05	and then there's models are placed on mine
0:12:09	we have been a feature is it into this classifier and the whole framework is
0:12:14	trained
0:12:15	well over the so
0:12:18	you know if all the with this
0:12:22	so
0:12:23	yes take a look at the experimental results so well as can be seen on
0:12:29	different this we find that
0:12:32	the
0:12:33	information in between
0:12:36	two khz to six khz in to be not comedy
0:12:40	in contrast to
0:12:42	information present in the last two killers sub band and we first
0:12:48	so in contrast on the training dataset we find that the first two khz sub
0:12:55	band is more informative
0:12:58	how valuable of the safety
0:13:00	as in previous case
0:13:01	our on a model
0:13:03	some of the best results
0:13:05	in what the data sets
0:13:06	so
0:13:07	the next experiment a set of what we do
0:13:10	is we know
0:13:12	but this
0:13:13	a input for their into
0:13:16	eight subbands so it's of one khz
0:13:20	so essentially what we do if we see it independency and everything on one delivered
0:13:26	formation
0:13:28	and is a this frequency entirely online again to
0:13:31	previous
0:13:33	a framework it is it's data it units all data
0:13:39	do
0:13:41	you can improve the
0:13:42	do so okay performance
0:13:46	so this is the experimental results for one lower sub-bands units
0:13:51	so this distribution across different sometimes allows us to actually
0:13:57	understand the impact of different bands
0:14:01	so only twenty seven dataset what this
0:14:04	is
0:14:06	one khz information seem to be the most informative
0:14:10	as we in and
0:14:12	eer of a two point one which is
0:14:15	e r in contrast or the frequency bands and interesting about defined is you've on
0:14:20	this isn't between
0:14:22	of different this and we just one two khz and
0:14:26	and it is seven killers is a system that informative as can be seen high
0:14:31	eer
0:14:34	and these second informative frequency for fixed expensive the be the first one khz
0:14:39	and of course will be or compact model operating on the
0:14:44	all h
0:14:45	and seem to give with comments but then we also right in just the last
0:14:52	a seven eight khz band and the first one dollars
0:14:55	which seem to give us the best performance in one for here
0:14:59	on the financial data see what we found is we found that the first one
0:15:03	is most informative
0:15:05	in contrast of the fifty s
0:15:08	so as i mentioned earlier so this is due to the fact that the twenty
0:15:12	seventeen and twenty nine dataset completely different so the fink intention is a simulated data
0:15:18	while twenty seven dataset is
0:15:21	is the real data that was
0:15:24	recorded and it back
0:15:26	using a
0:15:28	speaker verification it has it all right
0:15:31	so that this kind of explains the
0:15:34	difference a mismatch in the behavior
0:15:39	the final set of experiment we performed in this study is in terms of prostate
0:15:44	a simple ones
0:15:45	or with we with some of the best
0:15:49	models
0:15:50	not mentioned in dataset and original dataset and test it
0:15:56	comment on is visible twenty nineteen real be tested
0:16:00	we have used it is a very small essay
0:16:03	of
0:16:04	thousand utterances that was
0:16:07	instinctively conditions like the is organized as
0:16:11	and we want to see how this models
0:16:15	performed on realistic s conditions can be seen from the high
0:16:20	error rate distributions for all our models
0:16:25	this solicitous that the cutting us holding datasets training model doesn't actually
0:16:32	much on the realistic or if the conditions
0:16:35	so this thing that we might have to think about a few design or
0:16:39	training and validation sets for or standing s
0:16:46	so to improvement all
0:16:50	in this paper work with the we will be basically
0:16:54	but at all events and in
0:16:58	by discriminatively training independent seen in on
0:17:03	and
0:17:04	so
0:17:07	if variable a figment
0:17:09	and then there's a lady the later on the combined
0:17:13	and
0:17:14	and independent possible is trained on top of that
0:17:17	using the proposed methodology we found people performance on will be twenty seventy three datasets
0:17:23	and an interesting observation but not for or it which is a language that some
0:17:29	of the for this war is that
0:17:31	under twenty seventy dataset
0:17:33	e
0:17:34	seventy eight khz frequency formations in to be more informative
0:17:38	with however doesn't hold true on the training dataset
0:17:42	between ti din dataset the first one khz information seems to be more formally
0:17:47	and we also found that
0:17:51	the this wanna do not generalize a real on the on the realistically if conditions
0:17:57	with so this
0:17:59	that it is still room or for kids from
0:18:03	designing and validate in
0:18:05	this dataset for training effective
0:18:08	are replaced with an addiction models
0:18:11	so that i would like to control my
0:18:14	and you very much

Subband Modeling for Spoofing Detection in Automatic Speaker Verification

Spoofing and Countermeasure 2

Bhusan Chettri, Tomi Kinnunen, Emmanouil Benetos