Speech Transcript - Neural i-vectors

0:00:14	hello my name is to be addressed model
0:00:16	and in this video i describe our work in that narrow i-vectors
0:00:22	this work was go out for me by can likely and the only thing on
0:00:26	it
0:00:27	we don't we are from the university of east and be learned and going i
0:00:31	was the time of writing any
0:00:37	our study proposes a new way of combining gaussian mixture model based gender the i-vector
0:00:43	models discriminatively train the enhanced exact speaker and endings for speaker verification task
0:00:51	our aim is to improve upon existing i-vector systems
0:00:56	and we also hope to gain some insight is what causes the performance differences i
0:01:02	mean
0:01:03	in a speaker and the things and discriminant in the in speaker and that is
0:01:09	our study also is that is stronger convex and the gaussian mixture models and some
0:01:15	of the existing
0:01:17	ian and holding layers
0:01:21	as a background for how we look for different i are considered
0:01:26	the last three this can start can start suggested here
0:01:31	are combining ideas from all i-vectors and the intense
0:01:36	we a special events and the jurors universal background models and i-vector extractors all these
0:01:44	constants
0:01:46	let's to be the standard i-vector
0:01:50	so
0:01:51	key components here the two gender models are there
0:01:56	gaussian mixture model based universal background model and
0:02:00	i i-vector extractor
0:02:04	so even and
0:02:05	is used together it
0:02:07	initial easy readers to compute the supposition statistics for that
0:02:12	i-vector extractor
0:02:14	extract i-vectors
0:02:18	so we know the features are rule based and the rest of the components are
0:02:23	gender strange
0:02:28	then the nn i-vectors in this construct the universal background model is replaced by
0:02:36	these the and dates acoustic features as an input and reduce the senone posteriors as
0:02:42	an hour
0:02:44	these posteriors are used together and of c and it's easy features can be sufficient
0:02:50	statistics for the i-vector extractor
0:02:55	so this clustering differs from the standard i-vector and the universal background models discriminatively trained
0:03:03	one of the audience
0:03:06	third system descending an i-vector system
0:03:09	the system combines three modules one i-vectors is then the one neural network
0:03:16	these manuals are features statistics
0:03:19	when you
0:03:20	statistics i-vectors when you and that are module
0:03:23	is responsible
0:03:25	score and errors of i-vectors
0:03:29	training of these
0:03:31	kind of network goes as follows
0:03:34	so that are then used to train these individual modules
0:03:38	shortly these can benefit from the
0:03:43	i guess the a wrong
0:03:45	corresponding generative models
0:03:49	after these modules have been trained separately and they can be combined or and then
0:03:55	train
0:03:59	so this course there
0:04:02	you do less is generally models in the initialization stays
0:04:07	well i and i will use discriminative training the whole network
0:04:14	therefore and the last background construct this guy
0:04:19	is using and the nn with a mixture factor analysis fourteen year
0:04:23	in this for the authors used to estimate based you know texture you start speaker
0:04:29	and things
0:04:31	what is special about this
0:04:33	is that they use their own in a fourteen layer
0:04:37	these
0:04:38	for the error is basically an i-vector extractor implemented inside the in
0:04:45	is the m f a is based on after calling may or no must be
0:04:49	learned dictionary hangover
0:04:52	i think used right learned dictionary encoder right the wrong in this alliance
0:05:00	so we get all the components of these last construct our discriminately discriminatively trained with
0:05:07	speaker targets
0:05:12	okay
0:05:12	next we belong to the proposed neural i-vectors
0:05:18	before explaining the cluster itself
0:05:22	we will need to do
0:05:24	prerequisites for our model
0:05:27	and these are the know that and
0:05:30	the only layers and describe these two only layers by some and how they relate
0:05:36	to the standard c n
0:05:38	so then next initialize will be quite match for most
0:05:45	so first then it that
0:05:48	and we will study
0:05:50	the posterior combination formula or a standard gmm
0:05:55	we can see how we get the
0:05:58	note that formalize and all of their your question from here
0:06:04	so okay here we have
0:06:07	that was the number of gaussian components and its constant component power
0:06:12	covariance matrix mean vector and the associated right
0:06:18	okay in that assumes covariance matrices
0:06:23	or gaussian components
0:06:27	we will okay this four-mora in the is for
0:06:30	by expanding the normal distributions
0:06:38	then
0:06:39	but no
0:06:41	this inverse covariance times minima there will
0:06:45	my god
0:06:48	and
0:06:48	then the slow terms minus the other there we see
0:06:55	we get
0:06:56	this
0:06:59	and this happens to be exactly
0:07:02	formally used in note that
0:07:04	paper from two thousand and sixty
0:07:09	so
0:07:10	we have basically some on the last two means their covariance matrices
0:07:15	and the gmms we get that
0:07:19	same formalize and as in that
0:07:23	okay in it but i
0:07:25	illinois or learnable parameters there
0:07:29	note that there are
0:07:31	this form of grass
0:07:33	she's and news
0:07:36	and estimating these forming class and z is has to do not what i mean
0:07:41	there
0:07:43	we see from the posterior combination formula that doesn't depend and
0:07:49	from the mean vectors it is quite interesting signal can and the
0:07:54	standard gmms
0:07:57	but anyway
0:07:59	there you can compute the posteriors
0:08:03	or is there
0:08:06	input feature vectors
0:08:09	we can compare the component wise
0:08:13	what's
0:08:14	or in it but layer
0:08:18	formalize zone
0:08:19	on the right side the screen
0:08:22	and then
0:08:23	well right there
0:08:26	we have the first order centre so some statistics
0:08:32	the denominator just length normalized is then
0:08:37	so for each gaussian component we get one
0:08:41	vector
0:08:42	and finally
0:08:44	is no but they are male
0:08:45	concatenate is
0:08:48	component lifestyle closed form a supervector
0:08:53	so this is very similar to a
0:08:56	standard
0:08:57	c gmm supervectors
0:08:59	and how they are form
0:09:05	okay next
0:09:07	do the same for the learned dictionary encoder
0:09:10	only layer
0:09:12	so we start be there
0:09:13	gmm posterior combination formula
0:09:18	okay this time we you know we then
0:09:21	is
0:09:23	once colour term
0:09:27	do we get this
0:09:28	by expanding the normal distributions
0:09:34	okay
0:09:35	no if we assume
0:09:38	isotropic
0:09:39	or spherical covariance matrices
0:09:44	this formula
0:09:45	we simply by
0:09:47	this four
0:09:51	and
0:09:52	this is the
0:09:53	for music in that kind of prediction reading over all in layer
0:09:59	although in the
0:10:00	original publication or is and t is
0:10:05	be the term was not included but it was added later on by other authors
0:10:14	so the key point here must the
0:10:16	by assuming isotropic covariance matrices the l d
0:10:22	formulation from the standard gmm performance
0:10:28	then learnable parameters of this energy will are
0:10:33	i is
0:10:34	it's on the scaling factors for covariances then the mean vectors and is
0:10:40	i the terms
0:10:44	similarly as in that we can then going to the component was a rules or
0:10:49	is there
0:10:51	so again then we write directly have the first order some for some statistics
0:10:58	well okay in the standard and denominator is will be different so it is a
0:11:02	sample
0:11:04	posteriors for its
0:11:07	each component
0:11:09	so this is model i and it's the traditional maximum likelihood ratio on the east
0:11:16	is on one and vice outputs
0:11:21	and then the
0:11:22	on the nist and form a supervector
0:11:29	okay
0:11:31	so now we have the necessary can start you
0:11:35	constructs explained extend the proposed neural i-vectors
0:11:41	so we start with
0:11:43	and standard
0:11:46	extractor architecture
0:11:50	and we replace that
0:11:52	and are willing layer
0:11:54	we either that or l d coordinator
0:12:00	and as its or from the previous bias we can use
0:12:04	each polling layers the extra stuff isn't statistics
0:12:09	so we do that
0:12:11	and by using this present study is this weekend frame
0:12:16	regular i-vector extractor and you can also then extract i-vectors from these statistics
0:12:25	so that's the idea
0:12:32	so now we can completely stable
0:12:37	so how our how our proposed functions are dressed differs from there
0:12:45	able in their roles is that the
0:12:48	i-vector extractor is generally
0:12:52	otherwise the cluster is the same
0:12:55	if we compare our proposed neural i-vectors we then the in and i-vectors
0:13:01	we can see that the
0:13:02	i-vector what is the same but that
0:13:05	users and you the in verse
0:13:08	no one ever then restrained speaker utterance
0:13:13	and also the features are obtained from a
0:13:17	last layer before the one in there
0:13:22	next
0:13:23	that's model and the experiments and results
0:13:27	so we can say that speaker verification experiments on the speakers and one evaluation
0:13:33	first we compare our role as the results the other i-vector systems
0:13:39	the single fine from the literature and these are some of the best ones
0:13:46	on the line we have started in this easy i may or
0:13:51	and in the second one may have i-vector system that is isn't perceptual linear prediction
0:13:56	features together with the actual in the features
0:14:00	and this w the a is
0:14:03	dereverberation system
0:14:06	so we can see from this results the then all i-vectors performs the best
0:14:13	okay
0:14:14	so partial but let's next
0:14:18	compare our results
0:14:21	the nn speaker and endings
0:14:23	so we can use the same the nn sticks there either sufficient statistics for a
0:14:27	narrow i-vectors
0:14:30	or the can extend the speaker and endings directly from the audience
0:14:38	so
0:14:39	here are all these are our results so
0:14:45	in the first line we have a
0:14:48	the and we notable dictionary encoder whoever
0:14:53	be into one zero two equal error rate
0:14:57	but then the corresponding no i-vectors
0:15:00	okay that is that we use the same union the extended sufficient statistics
0:15:05	and then bending
0:15:06	then trained that generally
0:15:08	i-vector extractor so
0:15:11	the roles we can do one nine three
0:15:14	so no that's
0:15:17	okay the third level we have a modification of the learned dictionary encoder
0:15:23	so this uses
0:15:25	so i dunno
0:15:27	covariance matrices in instead of
0:15:29	isotropic covariance matrices
0:15:33	so we got been improvements by doing
0:15:36	these verification
0:15:38	the last two lines of their
0:15:41	results for then applied on there
0:15:47	so the interesting
0:15:49	they here
0:15:51	i wonder is a
0:15:54	what
0:15:55	what courses the performance difference between the
0:15:59	we generalize there's and then the in the things
0:16:02	because these are using the same the nn
0:16:05	but in
0:16:09	so there are two possible sources for this dress
0:16:14	so the first we used one
0:16:16	is the difference between the
0:16:20	generated by the model and they're
0:16:24	thereafter holding their
0:16:27	so
0:16:29	because of the holy where there is only one layer
0:16:33	even for the in the in here so
0:16:35	only this small part seems to really
0:16:40	well alarms
0:16:41	in terms in the equal error rate
0:16:45	so it seems that the discriminative
0:16:48	training objective is better
0:16:52	okay there is another
0:16:54	possible reason for this performance difference
0:17:00	so there is like mismatched how we trained the
0:17:04	b and n
0:17:06	one or we how we trained in the in holding linear and how
0:17:10	how we use it in the i-vector approach
0:17:14	can see that the and we explicitly form a supervector
0:17:20	and in there
0:17:21	i-vector
0:17:24	roles it is not a
0:17:26	i-vector of proteins the base adamantly so
0:17:30	is
0:17:32	like
0:17:33	console how many alignments how many frames are aligned itself the gaussian components
0:17:39	so this is missing from based supervector a row
0:17:44	so this is one of the
0:17:47	future works so
0:17:49	i used in the in owning layer is that it will resemble more there
0:17:55	i-vector approach
0:17:58	so this mismatch will be going on there
0:18:03	another
0:18:05	idea for the future work is
0:18:07	explain here
0:18:09	so
0:18:10	instead of substance that these three extra
0:18:14	the errors and the universal background model on there
0:18:19	the posteriors from this one in there
0:18:24	and
0:18:25	by using the is
0:18:27	we will then
0:18:30	have a neural gmm-ubm system we train based scoring
0:18:36	so this might be useful for some
0:18:39	special application why our for a welder a sore race and speaker verification
0:18:50	before i finished i have to related announcements first one is the program goals are
0:18:55	available
0:18:57	so we have i-vector extractor and providing their systems and in addition to speakers in
0:19:02	the mind you have also that there's
0:19:06	the goal or python and by those based
0:19:09	well we have or ugandans the on can be more research such
0:19:16	the second announcement is the this study was also included in my dissertation
0:19:23	and or is this tradition i have been extremely nice and coming residual and its
0:19:30	but weeks so
0:19:32	anyone who wants to jordan is pretty to design and we can be found
0:19:38	well
0:19:38	here
0:19:41	so you there

Neural i-vectors

Speaker and Language Recognition

Ville Vestman, Kong Aik Lee, Tomi Kinnunen