Speech Transcript - Analysis of Teager Energy Profiles for Spoof Speech Detection

0:00:16	why don't my presentation
0:00:19	and representing a or b i don and i this is often get energy profiles
0:00:26	also for speech detection
0:00:28	and a distinctly score for my all those
0:00:37	i'm not only
0:00:38	i can
0:00:39	well and myself
0:00:41	two additional okay
0:00:45	and a little or no one by matrix and the importance that issue
0:00:51	use of interest
0:00:52	and the key subject matter of this paper is development of cartilage or to alleviate
0:00:58	these phenomena
0:00:59	this can be that energy based features which is really for with different
0:01:04	a speaker and session position an attribute assisi
0:01:08	and we present experimental results on strong standard
0:01:12	but he wouldn't it as a namely is used for fifteen
0:01:18	was a big hassle
0:01:20	and in some ways
0:01:23	so introduction a speech signal okay
0:01:26	well with his that was of information and then applications well
0:01:30	but is unable
0:01:31	the linguistic information and banning was too much
0:01:36	the weighting function can be useful for speech recognition
0:01:40	language recognition but linguistic information can be useful for speaker recognition in all simulations this
0:01:45	estimation
0:01:50	there really is the us officials what into
0:01:54	automatic speaker verification which is the art speaker detection
0:01:59	so they are is there are in this paper we focus on speaker verification and
0:02:03	are also
0:02:05	so the decision in is equal be
0:02:08	enrolment in like background noise
0:02:11	and a noise channel mismatch
0:02:13	but it was because there is the data lines
0:02:17	the different be speaker in particular speaker vanity
0:02:21	because it should be noted that you when you three eighty speakers voice basically we
0:02:26	will speaker verification or speaker recognition
0:02:29	however
0:02:31	and dimension invariant also be clean speech communication channel
0:02:35	it is a speaker's voice well which is a
0:02:39	in the challenge
0:02:40	because of a really
0:02:42	in the speaker
0:02:43	if you have
0:02:44	with the whole middle residuals like
0:02:47	inter session variability
0:02:49	speaking style and pronunciation duration
0:02:53	features physical conditions gender
0:02:56	no
0:02:57	exactly
0:02:58	and their ability or a business activities in speaker recognition we call spoofing and things
0:03:05	with the means to create a lot of one of them
0:03:08	will be able to venice system effectively
0:03:11	and everything system
0:03:13	and the second system is to develop an icon colleges to be able to
0:03:17	and even something like that moment system
0:03:22	the and we'll have at all
0:03:25	something using model will be able to test the robustness of a system where is
0:03:29	an additional smoothing is important
0:03:31	could be able to design unfriendly a good you know all speaker system
0:03:39	well i secure by mixture
0:03:43	but the final types of the next well while the remaining speaker verification system
0:03:50	one is based on
0:03:52	we didn't or acoustics that his impersonation on the initial estimate astronomy
0:03:58	no recognition results acoustics
0:04:02	there is identical twins
0:04:03	and they're double the same animals do technologies such as speech synthesis and wise convolution
0:04:11	note or
0:04:12	news from charities based on this s and was in the
0:04:19	and found where there
0:04:21	there are in a very is and just activities as it would be in an
0:04:25	analysis of speech
0:04:26	that is
0:04:28	which is
0:04:29	very difficult to a very easy to model a doesn't wire technical knowledge
0:04:35	most loving where plate
0:04:37	in this story being the
0:04:39	on the order of speech or y
0:04:42	and so on but it is very difficult because the speech is from b
0:04:46	within the only
0:04:48	speech
0:04:49	well
0:04:50	there are in various in order for slightly a science and such things
0:04:56	one
0:04:57	a possibly case
0:04:59	the only four so it will be introduced on a relational some special sessions or
0:05:04	the sensitivities on
0:05:07	by image analysis of a false with a single remote control which will be next
0:05:13	however the major for star only
0:05:17	you know internationally will challenge timebin organized in please okay that's score a seasonal quickly
0:05:25	challenge to was
0:05:28	and the database was really was previously alright generalized for a
0:05:39	different in a little the content analytical
0:05:43	organise for speech synthesis
0:05:44	and what is a difficult question
0:05:47	well as those of you know there those on b
0:05:51	changeable elements which
0:05:55	exactly
0:05:56	last year lamenting janice was no
0:06:02	and i just the listeners and
0:06:05	and there was based on was used converting a play detection then the real speech
0:06:10	detection
0:06:11	and also
0:06:12	we well configuration of the physical or a system
0:06:16	and in a similar systems and so on addition
0:06:20	this is useful right
0:06:25	only the really comparison of the
0:06:28	i really wasn't the risk analyses also been or someone looks for
0:06:32	various kinds of been applied in also mention also be a d o meetings and
0:06:38	with
0:06:38	so since the i know the
0:06:42	the statistically meaningful car was for things and is a the
0:06:46	impersonation is not available
0:06:48	and therefore the risk is unknown in a problem at least
0:06:53	but as in industry in
0:06:58	lattice based on an additional reason is
0:07:02	and you know no is that is a very high
0:07:05	well models when we don't
0:07:09	otherwise that i this data in industry so on t
0:07:13	the latter detection training without a nice
0:07:18	it is what we mistral content based recording what and then us smartphone and in
0:07:24	control of the conditions
0:07:26	so the available gender errors could be you know
0:07:30	right behind mobile where the risk is very high that's for sure
0:07:35	an individual nineteen combine the enforce it is unity okay
0:07:39	in the different my from some additional which
0:07:45	so
0:07:46	forcible the problem here we can consider a l stand alone today the
0:07:51	which can be considered as a
0:07:53	you in without in berkeley systems contain natural speech
0:07:57	and there are four for speech detection
0:08:01	and the something that's things you go to be here
0:08:05	can be done i know that the we can be gone it was given that
0:08:09	we're getting microphone point on it is really and transmission channel
0:08:16	a sufficient wind
0:08:18	i'm really and in the literature so it is also based on just one there
0:08:25	so in this thing but we consider three times a small signal x based on
0:08:30	this reason why someone
0:08:32	is visible unit database
0:08:34	and the latter at s
0:08:39	so we finished isn't are from a speech
0:08:43	and convolutional useful analysis system designer and initial speech synthesis systems sort of speech a
0:08:50	natural language text
0:08:51	in was very little mostly extending steganalysis
0:08:55	then we need to find healing was and possibly a speech in addition we accent
0:08:58	speech
0:08:59	and this is at application that actually
0:09:02	how we can use different that is system to
0:09:06	communication
0:09:09	one thousand
0:09:10	like it is linguistically conversational wise samples for speaker who is one there is something
0:09:15	maybe once or speaker so this is in the intervals between the most
0:09:23	and basically kind of or
0:09:26	by considering the impostor basically and speaker
0:09:29	and that actually that would be that you from this is the same why second
0:09:33	one is
0:09:34	later in this context or
0:09:37	and things with something s is used
0:09:42	and so is a really useful
0:09:44	that's for eliciting model as convolutional on the actual speech we the
0:09:49	in both as follows actually plus the acoustic my
0:09:53	and i and i so i'm the impostors realistic will be on relational or i
0:09:59	mean the convolutional was response of the microphone
0:10:02	you recording idealise speaker
0:10:06	in the multimedia speaker and acoustic
0:10:09	so here the problem is to be able to understand the if we wanted to
0:10:13	the acoustics
0:10:14	which means you can detect some of the characteristics of equality
0:10:19	was legally or maybe condition
0:10:21	because you noting you wanna do speech coding speech
0:10:25	the parameters and
0:10:27	to build you
0:10:29	really independent acoustic something both channels according to my acoustic and one
0:10:35	we will understand whether the speech community and a genuine are indeed
0:10:40	and does not
0:10:42	so in this paper the one to exploit be okay spatial the initial so anything
0:10:50	that in your you get as you it is really the in energy off basically
0:10:59	i think is similar well energy or something that so for example in the traditional
0:11:04	signal also literature we learned online in that you where
0:11:08	however in the actual speech production the and belongs not sufficient which the energy requires
0:11:14	a statistical because
0:11:16	in that using whatever statistically hundred dollars acoustic signal
0:11:20	is there are more or less then
0:11:22	then as it were removed and doors a sticks in the context of acoustic signal
0:11:28	x
0:11:29	in the physical environment like simple and emotion and along a low dimensional the probability
0:11:34	like single emotion
0:11:36	automatically i dunno systems to describe why
0:11:40	and efficient which i was solution it was gonna silently and the ministry that's implement
0:11:46	motionless agenda
0:11:47	estimation of cornish energy bussgang energy which each year
0:11:52	two
0:11:53	but and frequency which is there a five cent signal energy is not only functional
0:12:01	roles signal that all the time and frequency and not
0:12:06	which is completely ignoring the actual and long
0:12:10	well at a in the energy so much
0:12:13	and of course a bic is nothing that you colour synergy
0:12:18	because in a sense that easy and square
0:12:22	so but c the speed of light in vacuum
0:12:25	and even a smaller cost
0:12:27	given that and then they rely on which
0:12:31	so
0:12:33	the binary that it may not here
0:12:35	is that the energy is not only depend on only one
0:12:42	and that is ignored in the conditions these approaches
0:12:46	so what we do we consider distributional is the channel and by considering these are
0:12:50	just a
0:12:52	and the ones a single emotion we consider speech portion of it should be speech
0:12:58	recognition the
0:13:03	so the solution is this is in business in the final say
0:13:06	which is presented in this really so well before they're a little briefly mention that
0:13:13	these features are just an initial estimate a model
0:13:16	the thing to their is you know which include pitch an electrician sufficient condition
0:13:22	metaphysics fusion and features
0:13:24	i'm late
0:13:26	these features based on systems and what i
0:13:28	and that the energy based features that capture if a parent and the resulting in
0:13:35	and the features more variability
0:13:38	didn't seem as mentioned in section which was then you know yes
0:13:43	you know constant
0:13:46	so be in that an active speech production facilities where s is i was really
0:13:51	you know in the union model questioned by
0:13:55	no not limited of each motion is more speech
0:14:00	and this is a little investigation which means as shown that it will lie
0:14:06	this is a mission the line and didn't you
0:14:09	well with a mean basically a
0:14:11	the revolution unit that is one that
0:14:15	and in one year
0:14:16	and you show that different phase maybe do this
0:14:21	and
0:14:23	is the house italy having to improvements in the only one sound an instructional
0:14:32	so this is basically a sound signal just and yet and you also mean and
0:14:37	you see that was and with the total number between one managed to make a
0:14:42	basically
0:14:44	something clustering based on and minus twenty which is coming up
0:14:47	and false was less value omega
0:14:50	we begin basically
0:14:52	so i'm innocent civilians form it is clear luminosities nothing but a square and minus
0:14:57	itself in boston based on this is nothing but this is gonna ministry rather than
0:15:02	the previously well as you get a functional these we will call is based
0:15:07	given that is good anything the
0:15:10	is generally the this profile which is given by exploiting minus one
0:15:15	plus one into it and minus one
0:15:17	in order to the difference between the simple elements one
0:15:22	okay so
0:15:24	in this there will be used to you is really a limitation for
0:15:29	these things are a lot and silence a minute do you we use the energy
0:15:35	because
0:15:36	there is a ducking under in using their so that and in his and you
0:15:40	can imagine also not using
0:15:43	this is a signal so you will be superior i'm writing
0:15:47	you know
0:15:50	so well
0:15:52	no she can see that are you really
0:15:57	the view point explicitly of speech
0:16:01	this is that the of speech
0:16:02	and this is the view point
0:16:05	so as to what these are the a new values ten split speech
0:16:10	as we can see that the audio file is maximum you
0:16:14	indicating that
0:16:15	and it is a high snr and using both linear prediction inverse like iteration well
0:16:22	very high energy and but being able to the energy so use high energy as
0:16:28	well
0:16:29	secondly we use
0:16:31	a lot in this way speech just convolutional
0:16:35	you was responsible a automatic systems only do not be an impulse response of there
0:16:41	are some interest
0:16:42	kind of all places for a moment for isn't it was also sponsors
0:16:47	will be you know in both senses so therefore
0:16:51	i system is already a this represent one
0:16:55	and the display the impulse response of one
0:16:58	therefore if we consider with the v by giving a basically
0:17:04	so for you in this explicitly mechanism where only here
0:17:08	then you will remember only the data streams are not institutions these fluctuations are basically
0:17:14	all
0:17:15	most so lately
0:17:17	we wanna speech or something and so otherwise huge estimation of impulses also otherwise
0:17:23	we really
0:17:24	we in both systems
0:17:26	so fourteen recorded in boston signals are relocation only there is an excellent in the
0:17:32	u i
0:17:33	and blue or within the one on the gas
0:17:37	this is for models a star
0:17:39	the impulse response is considered that and also for all right and so do you
0:17:45	whatever sometimes was the gate functions
0:17:48	however control the explanation involved only more using function you can see that all this
0:17:54	stuff to that real
0:17:56	and negating their
0:17:58	these speech distortion cannot consider only because of being
0:18:03	so
0:18:04	v high on what we
0:18:07	actually in this work we
0:18:10	consider this observation on the ball in the meeting to do not constitute this innovation
0:18:17	and costly for our miss and false illusion meetings
0:18:23	we consider this is added to this for signal which is easy
0:18:27	the final season is in the next we think that and we are going back
0:18:32	to that used and the
0:18:33	of speech for example
0:18:36	the these and the yearly was that i here
0:18:40	other than in an actual speech corresponding to you guys also very
0:18:45	and that's
0:18:46	but it wasn't there is needed is one as constant consistently so the overall while
0:18:51	also give you get a constant high on the energy
0:18:56	and here llr fluctuations in the next speech and in the next speech for the
0:19:00	one extracted via
0:19:02	the model such
0:19:03	no fluctuations almost or the rest of the homes of this work
0:19:08	socialisation investigation well sure well why that
0:19:13	basically
0:19:15	this
0:19:16	models will be also a for
0:19:20	a small degradation and we also this one on the spectral features
0:19:25	we also this one that we that addition a
0:19:28	we found that in an actual speech
0:19:30	comparison of lr
0:19:32	the initial with features new was really matter
0:19:35	got additional getting the
0:19:38	but haven't in capturing the performance chosen separately distribution in an action was itching speech
0:19:44	we also have the same thing basically on the bus one sixteen or database for
0:19:48	the natural
0:19:50	and e
0:19:50	but it is in each condition is therefore anyone important easily speech just from the
0:19:55	missing the native speech
0:19:57	we also that
0:19:59	the buttons on go
0:20:01	this one you again and you which is a screen
0:20:05	and decision that was basically features which is a on
0:20:09	yes there's recognition and signal
0:20:13	we passing through the band filters
0:20:15	and are
0:20:17	on the this thing and mel filterbank
0:20:19	we use an explicit about
0:20:21	and then
0:20:21	this filters out a little sub band signal that again as you face
0:20:26	and then this can now investigate nearly one of those that's why we model
0:20:30	then we move a mean and averaging all those in that you and the non
0:20:34	dct comparing the
0:20:36	energy research which uses the assumption contribution in the time
0:20:42	we standard it is generally is this problem can database and because of the database
0:20:47	and we use this is that it is feature dimension low dimensional feature does not
0:20:52	model so that it does not want using gmm and the frequencies
0:20:58	i think is used to use lost in the mfccs and then linear sequences elements
0:21:03	yes mel frequency
0:21:05	and we employ union
0:21:07	so the nine dimensional feature vector is more or less one twenty one twenty four
0:21:13	of this increases the finances is thirty nine and it is commonly used in addition
0:21:19	to capture
0:21:20	and six
0:21:23	so far results online in the master
0:21:26	we also there the results for the proposed features as i
0:21:30	and it really of their combat mfcc and you design a reversed is easily
0:21:36	well
0:21:37	little better results than people just
0:21:40	but when we use this the results where he can be used
0:21:48	and the six
0:21:50	these are the leading goals
0:21:52	we can see that the results for the on the development and the results for
0:21:55	the future statistically significantly better than
0:21:58	mfcc
0:22:00	four
0:22:01	a development set on it is useful not continuous
0:22:06	and then we also statistically
0:22:08	results is only mentioned it does it is s one s ten
0:22:12	and asked an is used in this is that is the highest
0:22:16	so it is very important role nor their
0:22:19	the equal error at least a relatively low
0:22:23	for the and was features are and this is because as you compare
0:22:27	the nazis is you an existing which was well when on the basis for almost
0:22:32	a phone a just a
0:22:34	in this work
0:22:36	well contingency
0:22:39	however one hundred and it would be here was you just you very large
0:22:44	well you think whether you listen for testing
0:22:46	and it is expected because
0:22:49	s ten is
0:22:51	based on tts four wheel which is the little based on a decision based on
0:22:55	that each
0:22:56	an active speech i is organized as we chose unity right
0:23:00	well in the model suggesting there are basically
0:23:03	the one thousand and then what is and using
0:23:06	created in the gmm based system
0:23:08	and the standard english more
0:23:11	you know
0:23:13	and then use the best performance on the
0:23:15	on the features it only has a very also
0:23:19	i was features on better than the existing to generate mfcc and sixty
0:23:26	what is it again mapping
0:23:27	and a few though features in windows uses the
0:23:30	and a serious a or b and disability related in these score distributions o d
0:23:37	development and the
0:23:39	you versions therefore the
0:23:40	mfcc
0:23:41	c is easy and easy gives you know it is the
0:23:45	system and english versions of one hundred and four not
0:23:49	be these features
0:23:52	and we also found this on the initial results on the eval set
0:23:56	we also that the proposed features to perform better than
0:23:59	miss consistent use based on spectral energy features and it is a little on the
0:24:04	mean an additional classifier
0:24:05	now and stands for the r m is just a
0:24:09	however unknown the that actually you know
0:24:12	the last one do what s
0:24:15	using that was just a matter
0:24:18	was recently small for the baseline systems and ms
0:24:22	indicating that the emotions are data needed
0:24:28	i think
0:24:28	this is the lead to go sure would be just as soon as it will
0:24:32	it is also their the on the development set the most features to a that
0:24:38	is shown as a and b
0:24:40	depending on the line
0:24:41	no not tonight
0:24:43	performs significantly better and then speakers in a nursing
0:24:46	and the fusion is a lot on a similar to the but with features and
0:24:51	there is this is just
0:24:53	well i
0:24:55	similar results are only versions are there
0:24:58	on the that was just system fourteen better than this and
0:25:03	and
0:25:05	feasible is not from phone
0:25:07	really doesn't bother to at most
0:25:10	combat or indian
0:25:13	finally or something but
0:25:15	in this thing but only exploited bouncing you try to form that was just an
0:25:19	addiction
0:25:20	he of is known about features are evaluated on the standard as well a system
0:25:25	been viewed as well as she was and only better than existing we just
0:25:29	this is i don't do not will for testing there is units isn't this just
0:25:33	and just which is based on okay show that is just understands exploits
0:25:38	it doesn't deal almost a single
0:25:41	and s is the problem is a really going to do so especially in
0:25:45	in addition to nist
0:25:49	the senator differences and only time
0:25:51	as a result be a sixty nine under
0:25:54	well i wouldn't look at compression and distance well
0:25:58	in speaker recognition for what in the score should the
0:26:01	it just one or more or fess
0:26:05	the organisers or
0:26:06	on is to go with marshall argument each a and of course urination nine just
0:26:12	basically
0:26:13	is it is possible to not contain challenge
0:26:15	i don't think industry
0:26:17	i mean and five shows

Analysis of Teager Energy Profiles for Spoof Speech Detection

Spoofing and Countermeasure 2

Madhu Kamble, Aditya Krishna Sai Pulikonda, Maddala Venkata Siva Krishna, Hemant Patil