Speech Transcript - A Hybrid Factor Analysis and Probabilistic PCA-based system for Dictionary Learning and Encoding for Robust Speaker Recognition

0:00:15	i
0:00:16	and B
0:00:20	i
0:00:21	i this is a probabilistic pca based system for
0:00:24	dictionary learning and encoding for speaker
0:00:47	oh
0:00:49	i'm second i'll go into instructing limit of the bic student
0:00:54	at present my paper isa
0:00:57	because pointed out it's data
0:00:59	oh it's a hybrid factor analysis system
0:01:03	and it uses ppca
0:01:05	oh basically to simplify the compute intensive computation intensive parts of the factor analysis system
0:01:12	as we have seen from the previous
0:01:14	talks
0:01:15	oh the fact that the system is
0:01:18	what competition intensive and a this work is still any about how to simplify some
0:01:25	parts of the system
0:01:26	us so that
0:01:29	oh be again some advantages at the same time but not a having some state
0:01:33	of wanting performance process but
0:01:36	so basically all of the four
0:01:40	and that as we basically explaining why such a
0:01:47	such a thing is possible for so far and what perspective of this factor analysis
0:01:51	system that enables us to simplify
0:01:53	a those parts that i'm gonna talk about especially the hyperparameter estimation technique which is
0:01:58	basically the T matrix in the subspace model
0:02:01	a spastic total variability space
0:02:05	so and the end will be looking at how the performance of the system is
0:02:10	and two is a very modest a representation of the entire vector
0:02:14	spain but
0:02:15	so basically you have supervectors only to fix the damage the representations of speech utterances
0:02:22	a that are converted to a low dimensional
0:02:26	i-vector representations
0:02:28	so this a second time you press
0:02:30	i D W is so basically the representation used in this paper
0:02:35	and i have i'm going to fix the same station
0:02:37	this paper is that
0:02:40	so oh to just
0:02:43	for the sake of completeness i just get i
0:02:45	most of us my most been doing what is
0:02:47	oh there was a light of the system but
0:02:50	but still thank you
0:02:52	the
0:02:53	i just explain what is happening and
0:02:55	then we know the perspective that is very important for the test
0:02:59	and this is that the that is
0:03:02	for from a speech utterance
0:03:04	we consisting of feature vectors at
0:03:06	and basically we use the G a gmm parameters
0:03:11	gmm ubm parameters
0:03:13	to basically for the supervector
0:03:15	and what was once we have a supervector be from the development data we used
0:03:20	to train the
0:03:22	subspace motivated videos is met
0:03:25	and then we try to extract the feature i-vectors that are quite of the sent
0:03:29	data
0:03:30	and get low dimensional representation to presentation
0:03:33	okay
0:03:34	speech utterance
0:03:36	so once we have
0:03:37	a testing phase we try to find acoustic distance between the target speaker and the
0:03:43	brain patterns
0:03:44	and this is how the agenda the general framework of speaker recognition system
0:03:49	and such a system can actually be viewed as
0:03:54	oh consisting of two
0:03:58	and encoding
0:03:59	so here once you are the development data this product development data you like you
0:04:05	oh estimate a subspace in which all these people with the state is the total
0:04:08	variability space and this can be done to speech and it in
0:04:11	and i also wonder and one of that's this paper
0:04:15	this is an overcomplete dictionary
0:04:18	and
0:04:20	so once the subset you matrix is lower
0:04:23	we try to encode the data based on the supervector it has been observed
0:04:27	okay so this is
0:04:29	audio video frame but
0:04:31	that there is used in this paper
0:04:33	so we'll see how decoding these stages of the and rate if the system is
0:04:39	that by the us
0:04:40	the basic motivation behind this is
0:04:45	a variational importance of encoding and importance of decoding this entire system
0:04:51	oh as i in two phases which is the dictionary learning
0:04:54	sure but you
0:04:57	so basically the end of the
0:05:00	but must be done on a sparse encoding procedures better for example if
0:05:06	you take a orthogonal matching pursuit
0:05:09	algorithm a bit okay sparse vectors
0:05:12	and you train a dictionary using those using that algorithm they also that using in
0:05:18	your encoding algorithm
0:05:20	but not work
0:05:21	so they have the that some of the encoding of buttons what better than the
0:05:24	others in any does not is necessary had to be the optimal matching pursuit algorithm
0:05:30	set
0:05:31	so i'll for example this they have that the soft thresholding scheme works better
0:05:37	a way back to the speaker
0:05:39	oh
0:05:40	so yeah i is an opportunity to see if we can
0:05:44	we list a particular a base that is very computationally intensive to explain the observations
0:05:51	made in this work
0:05:55	just to
0:05:56	see this is cool for any improvement in terms of computational
0:06:01	a efficiencies
0:06:02	we look at the union
0:06:04	step i was thinking is taken from
0:06:07	the egg S P
0:06:08	you the e-step we want to
0:06:11	we started images with that them
0:06:14	please
0:06:14	and then accordingly anomalies the
0:06:18	columns
0:06:20	you
0:06:21	you get a i-vectors of development data and then keep the estimating a given okay
0:06:26	oh convergence
0:06:29	so this
0:06:32	this isn't about the and you see
0:06:35	okay is
0:06:37	machine
0:06:39	additionally density and i'll try to formalise it in terms of a big limitation data
0:06:46	so once this is done
0:06:48	i
0:06:50	to look at an alternative if they don't if it is the total variability space
0:06:55	model
0:06:56	which is a problems T V C
0:06:58	the
0:06:59	once is in this one is it is yes introspective in addition
0:07:05	oh they come up with a stick three parts of this estimation
0:07:09	and one of the important ones that they have in that is that is
0:07:16	it just a special case of i
0:07:17	that is a set of them badly covariance matrix is
0:07:21	for example
0:07:23	and one of the main
0:07:27	S but there is that how the computation of the covariance matrix in a probabilistic
0:07:32	we use L
0:07:34	it's less intensive in terms of four
0:07:37	computational complexity when it comes to
0:07:40	a very high dimensional data samples like this product of the P
0:07:45	so a lot of the reasons if it's but
0:07:47	but that's be observed the proposal ppca is not as good as the
0:07:53	oh that of the i-vector techniques
0:07:56	the conventional factor analysis techniques and we'll see how to
0:08:00	oh
0:08:01	complete all the observations that we need to know into a
0:08:05	these systems
0:08:07	so this just use
0:08:09	say i hear the unit and step
0:08:13	i
0:08:16	and a similar i saw analysis case
0:08:19	except a that as so the first a P
0:08:24	kenny mentions that the ppca does not the
0:08:27	necessarily assume that that's what it does come from a gmm
0:08:32	so this that the computation vol
0:08:35	yeah instead
0:08:38	so we can see that
0:08:42	and what steps we can see there's a huge
0:08:46	different
0:08:49	you
0:08:50	less intensive than that of the conventional technique
0:08:53	your C
0:08:56	i
0:08:56	images
0:08:57	ubm
0:08:58	yeah the dimensionality of the feature use
0:09:01	i say
0:09:05	i five
0:09:07	i don't understand percent in this case of development data sequence
0:09:10	we should be allowed
0:09:12	and that's why
0:09:14	they said to be
0:09:16	i
0:09:18	i
0:09:20	i'm you know his own
0:09:21	but what can be done
0:09:24	oh
0:09:26	is the
0:09:28	you say you know
0:09:30	this that these sources
0:09:34	you
0:09:35	is it possible
0:09:37	to consider only i mean you method using the ppca
0:09:43	in
0:09:45	using the conventional thing doesn't give any advantage of these sets of observations
0:09:51	to make it
0:09:52	and
0:09:55	so the proposed approach
0:09:57	and basically
0:09:59	the estimated you made using
0:10:01	here
0:10:03	i
0:10:05	okay
0:10:07	the conventional technique
0:10:09	bad
0:10:10	you make an assumption
0:10:12	that which makes an assumption that the supervector comes from a german
0:10:18	so what happens is
0:10:20	the i-vector that are encoded using
0:10:23	taking
0:10:24	oh
0:10:25	the rest of this
0:10:27	i was going this time constants which is
0:10:32	oh
0:10:32	okay
0:10:35	my presentation
0:10:37	i think
0:10:38	see without so suppose i
0:10:41	ppca estimate of T
0:10:44	the i-vectors are that are estimated so what was interesting here
0:10:48	that if i had to estimate that is using the conventional ppca technique
0:10:55	oh i is that you want to be
0:10:58	what happens is estimated using the
0:11:01	oh proposed approach
0:11:03	is that i think that information
0:11:06	as
0:11:08	speech are basically covariance matrices
0:11:10	and in the middle but that expression access to normalization
0:11:17	which seems to be really useful
0:11:19	which will be seeing in this
0:11:23	so
0:11:24	using
0:11:28	was that experiments on is that it is that
0:11:33	okay
0:11:34	i
0:11:36	the data development dataset used is right
0:11:41	minimalistic then compared to the box
0:11:44	oh
0:11:47	these are
0:11:48	is that
0:11:48	databases
0:11:51	me parts that are missing
0:11:53	there's
0:11:54	i
0:11:56	and the mfcc features extracted
0:11:59	basically the string to cepstral coefficients
0:12:02	and those are numbers in addition
0:12:04	consider the feature extraction is
0:12:11	so that means we're and you matrix
0:12:15	so it is five hundred but with in what's techniques
0:12:20	and the standard hmm based is only that has usually in the in the one
0:12:25	we wccn
0:12:26	and the data that can be
0:12:28	a plate mental image
0:12:32	so the doctors directions are
0:12:34	leading
0:12:37	and just to say support for that you see it's
0:12:42	much faster than the conventional technique
0:12:46	this
0:12:48	the
0:12:49	although system was implemented
0:12:51	oh
0:12:52	yeah systems with that in matters fashion using
0:12:57	and the contents of speech recognition is and that makes
0:13:00	so
0:13:01	if you look at the context of
0:13:03	technique
0:13:04	vol
0:13:05	yeah
0:13:06	we see here for different
0:13:09	the whole that the difference in
0:13:11	many B E citation
0:13:15	and
0:13:16	with the final class
0:13:18	so
0:13:19	we wanted to
0:13:22	take advantage this of
0:13:26	this exactly
0:13:30	oh
0:13:31	so
0:13:32	one C
0:13:33	just
0:13:34	preliminary tests
0:13:35	we see that the i-vectors incorporate
0:13:39	in that way that we have
0:13:40	proposed
0:13:41	are
0:13:43	good enough for
0:13:45	well being used in speaker recognition
0:13:48	so this shows that interspeaker and intraspeaker is
0:13:51	is
0:13:54	i
0:13:57	i
0:13:58	in the performance to ensure that
0:14:01	it is
0:14:03	you
0:14:03	the degradation
0:14:05	yeah
0:14:11	conventional factor analysis system
0:14:13	on this
0:14:14	still
0:14:18	so one aspect of a
0:14:21	this work that is interesting is to find of the relationship between the two i-vectors
0:14:27	that are extracted in two different ways so just to look at the whole data
0:14:32	related with the relationship is linear we need to be used to
0:14:37	correlation analysis
0:14:38	canonical correlation analysis and
0:14:42	applicable
0:14:44	so basically cca is like mutual information greatly usually
0:14:49	estimation of mutual information and that all there is a large population
0:14:55	but when you have a basis so if
0:14:59	if you need to determine that the relationship is nonlinear that is
0:15:03	it is in here in a high dimensional space
0:15:05	all but try to use P C
0:15:09	so what we can see is that the convention and i think and ppca subspace
0:15:14	is not nearly but
0:15:16	that is what and suggest is that what is based on that
0:15:21	and then you look at a yeah
0:15:26	it should be
0:15:28	the i-vectors extracted from the
0:15:30	a conventional approach and the i-vector six extracted
0:15:35	and the proposed approach
0:15:36	the extent of the asian to its
0:15:40	oh basically a full kind of T V though in the space generated by a
0:15:45	point we can
0:15:46	and this is
0:15:49	what is a most interesting aspect
0:15:51	what it
0:15:52	gives is that
0:15:54	is an opportunity to look at different splitting procedures
0:15:57	oh
0:15:58	so that the performance of what systems can be much
0:16:03	so in this
0:16:05	a baseline system forty eight and C six is given
0:16:09	and you see you know
0:16:14	oh i have a problem is
0:16:16	the so that this
0:16:18	six C
0:16:22	is it is
0:16:23	and you can see in two days
0:16:26	what is it that it's a
0:16:29	yeah
0:16:32	or it's a ppca
0:16:36	so if you look at the ppca technique
0:16:39	and the proposed
0:16:40	though there is a clear improvement in terms of the
0:16:46	so in summary
0:16:47	oh the ppca
0:16:49	this is actually which is the total variability space
0:16:53	oh matrix
0:16:54	and in doing so we
0:16:57	oh
0:16:58	speaker for explosives or
0:17:00	and the performance
0:17:02	close to one
0:17:04	point
0:17:06	respect to the
0:17:08	the degradation was just one thirty nine point
0:17:13	yes
0:17:14	and
0:17:15	which that is also a baseline system
0:17:18	and
0:17:20	we are basically ppca system the improvement is
0:17:23	related to point a person
0:17:25	and the i-vectors
0:17:27	system
0:17:28	one important conclusion
0:17:30	i-vectors and the
0:17:32	proposed extracted using the proposed approach
0:17:35	is non-linearly related to
0:17:37	those and the baseline
0:17:55	i
0:17:58	yeah P
0:18:00	this
0:18:02	oh
0:18:05	that was a context as
0:18:07	cost "'cause" use
0:18:09	but you're based sparse structure is always
0:18:12	so you used
0:18:15	oh is the observation was that
0:18:17	i don't know the reason why the decoding of the
0:18:21	dictionary learning and encoding part and me is there is there is only observation right
0:18:26	number i digital data stream
0:18:29	and then some
0:18:34	so i'm not
0:18:38	oh
0:18:50	and the speaker again

A Hybrid Factor Analysis and Probabilistic PCA-based system for Dictionary Learning and Encoding for Robust Speaker Recognition

SESSION 01: Speaker Recognition - Compact Representation

Srikanth Madikeri