Speech Transcript - Advances in Speaker Recognition for Telephone and Audio-Visual Data: the JHU-MIT Submission for NIST SRE19

0:00:13	and a causes b well i think we're interest and therefore above is the speaker
0:00:17	recognition for telephone number of is one data
0:00:20	usually my these submission form is design a
0:00:24	this is a during war from distances on the human language than only standard estimators
0:00:29	and standard orders
0:00:30	it's and language processing from the two in my feeling
0:00:36	these assigning the income tax
0:00:38	c d is like we telephone speech intonational i mean
0:00:42	and the audio visual like composed of them on the internet deviates from the bus
0:00:46	core
0:00:47	there is one that have a speaker recognition on face recognition only working model of
0:00:52	this
0:00:53	all also used and what database or some formula one and very well why don't
0:00:58	you lda or cosine scoring okay
0:01:01	in the other side the key points for women for still you what do you
0:01:05	really
0:01:06	not is gonna the nn are businessmen
0:01:10	still be lots kind of a place the nn vectors
0:01:13	also
0:01:14	but cannot they still using in the melee but mostly from any estimator fine tuning
0:01:19	to in domain data also
0:01:22	one will be assigned the key points where usage of rain is that are based
0:01:28	bodies
0:01:29	okay we use cosine a score in several areas these to combine it is and
0:01:33	variance from different be there
0:01:38	again we use this on the i
0:01:40	what i will face acoustic features that similar well for be overcome based detection
0:01:48	problem that we just a speaker
0:01:50	or face images
0:01:52	and do not isn't based on but when there walking and we kind of in
0:01:56	this course
0:02:00	we start describing the oracle systems
0:02:04	so we're starving different acoustic features we use and this is used for units vectors
0:02:09	and build a lattice for rest of this vectors
0:02:12	it's be we use community vad or sixty s and you don't and v for
0:02:17	really
0:02:18	in video we constantly system
0:02:23	so what we'll from there is a sin was clustering or a be lda gmm
0:02:27	that a single speaker factors in speaker labels posteriors
0:02:33	we used to estimate of labels
0:02:35	based on similar you know and as the best one double will make me but
0:02:39	is not very sure would be is generally
0:02:42	also on responsiveness might consider are less money is
0:02:48	this and that was one based on god i
0:02:53	we got some improvement will is but i during
0:02:55	i seriously we're finding the for the n and then what we in domain data
0:03:00	just finding the leslie using four letter words in this way embodies becomes a sinus
0:03:06	or
0:03:07	and we call this
0:03:10	besides discriminant percent
0:03:15	so we have seven that is that or architectures
0:03:19	we have
0:03:21	i was gonna be and then but since
0:03:23	three five basis
0:03:25	than better since what we're gonna since the is the same that we use of
0:03:29	and sre
0:03:31	the contains translators from new domain
0:03:35	with a linear size of
0:03:36	one thousand four
0:03:39	alright is an utterance
0:03:42	we unknown
0:03:45	and therefore based on find a we have regulators five miles away two thousand forty
0:03:52	eight
0:03:53	are very agreements
0:03:56	we also several possible ways that five questions
0:04:00	they're having less than wireless the inverse there wasn't one the and that's always been
0:04:05	feeding
0:04:08	this is this one of the datasets used for training or not the inspectors
0:04:14	so it's in serious condition
0:04:17	zero use switchboard was designed for okay
0:04:21	r c of this work
0:04:23	it's
0:04:24	there isn't or is something the in work we use all the data set someone
0:04:29	one their completion
0:04:31	a is evident in a we use the same but with the model
0:04:36	we remove the so systems i one microphone
0:04:41	lincoln labs the still use businesses
0:04:45	microphone
0:04:48	confrontation or this though
0:04:50	we used as i e one
0:04:52	and i'm gonna is this study
0:04:55	and you state
0:04:57	or they are all from being the one d c and we just use the
0:05:01	most of the thing in this
0:05:06	we have a for like principal equations
0:05:09	c l is the only one last use the first configuration that's the line of
0:05:13	the you're
0:05:16	let's say that we have some all domain and some in domain
0:05:21	first we and that the out-of-domain in domain using their or a little
0:05:26	and they're all in an out-of-domain data in
0:05:30	then we use a different thing that in for in domain
0:05:35	although mean data
0:05:37	we use common whitening
0:05:38	then the my face
0:05:41	the other two in domain data
0:05:45	are then at the score normalization was and in domain data a calibrated
0:05:51	but for steely and have a three by conventions
0:05:55	something that and use for that all lda
0:05:58	and the use yes everyday the lda for a swear and very nice thing what's
0:06:04	almost instantly
0:06:05	are also in the scoring or
0:06:09	we also the lda for cases where
0:06:13	and then it is then we only the model in salt
0:06:16	or
0:06:20	so this is a this what are the values something the markets
0:06:24	that's a small difference between sites
0:06:27	but as forces us ordinance yuri a on
0:06:32	the use this study for then i x values to some well on this study
0:06:36	in u one
0:06:38	for the dc one
0:06:41	as you use the is something at you by
0:06:45	we also and since the only problem we also use the unlabeled
0:06:52	that it really by doing clustering
0:06:56	or other score normalization we use the only really
0:07:00	i'm use the sre seen that for
0:07:05	or maybe a we just think that can almost the latter
0:07:08	this is a very good speakers in the white honestly demos data
0:07:13	score by bayesian also us
0:07:16	the i have to be also provided us an significant improvement
0:07:21	a value will use this i think bias you one for calibration
0:07:29	that's you know this used the silence
0:07:32	first we analyze the us also that five million and
0:07:37	romana something the we use
0:07:40	where a source false or misleading there
0:07:44	on the on the lower a sliding i b d one all the
0:07:49	the base then system used unsupervised really in a bayesian with this study only
0:07:56	then in the signal were we is that in the u one okay
0:08:01	provides a very nice
0:08:04	then we i we are noise segmentation lately
0:08:09	that improves the convince your in the u
0:08:13	then we have that the a spectrum and also
0:08:16	and the in domain be i get some room and you the by a small
0:08:20	improvement
0:08:21	all in one
0:08:23	i think that if we change that sure or then run your that's where we
0:08:28	made the grade on our way we
0:08:31	getting some
0:08:32	implementing that you well limbaugh an improvement in the
0:08:39	also analysis on this you by also versa before rest
0:08:44	the bayesian network use a risk of a system for based silence mean versus evaluation
0:08:49	will also must present a unique
0:08:54	then we alignments unless something dusty the data
0:08:58	provides a nice improvement in the u and it again
0:09:01	then we a the we got a number of channels in the network and that
0:09:06	provides a small role
0:09:09	not remote really okay and we define the never will always unusable sinus fourteen
0:09:15	so on without use of us more ergonomically baseline but in there about their grace
0:09:22	and they always fits to the or something or thirteen data
0:09:30	and that's was in those identity
0:09:35	these are also all four to all the single system
0:09:40	the based system is your five better results before was one of the database sinus
0:09:45	ability have okay
0:09:48	so we're very close to be easily affected formal system for which channels
0:09:51	a personal one of the
0:09:56	and
0:09:57	for this part of the nn with the
0:10:00	will be the training set
0:10:03	in all cases you was greater than this method was i
0:10:12	or we apply several
0:10:15	medals for the fusion we have there
0:10:19	but it's a you don't use of in it was used in calibration and yes
0:10:23	is for a basis for
0:10:26	an efficient v
0:10:28	once you so in the real assisting calibration a one when you mean and another
0:10:33	is that it is not the union that i mean and
0:10:37	the scores
0:10:40	a quality with a where we can see that is consistent when interviews with a
0:10:45	very high or station
0:10:49	are you sure we got everything we on over and over
0:10:55	so the based system for us your proposal by in address the source for calibration
0:11:01	i
0:11:03	i think five series systems with but like plus three system is not possible
0:11:12	or
0:11:13	usually might need them
0:11:16	we have the fusion of existence
0:11:20	and the basic progress is a thing with fusion be but obviously once she
0:11:29	the best results that they want you can see that are the system also
0:11:35	the present problems phones your feature
0:11:42	no it's either a your problem of your results
0:11:47	was also an analysis of our last for the nn are where lunges it was
0:11:52	also for delay of advanced
0:11:54	or the u s
0:11:58	the first figure analyze this problem i phase you're
0:12:02	so and we can see that score normalization provides more meetings in a savvy the
0:12:07	in domain sre an eighteen
0:12:09	also we can see that i mean by handle this problem i faced is that
0:12:14	why
0:12:15	provide some a similar guy
0:12:18	great
0:12:19	the second year so the was also a v i
0:12:24	right and that we will one between their usage
0:12:27	so the decision rule
0:12:30	the relative improvement in bic studies
0:12:32	log in this i mean idea of illness i in
0:12:35	so systems and it is easier to the utterance
0:12:40	besides the results of the signal system that we used in all submissions
0:12:46	we can see that there is anything about christmas is to have that is that
0:12:50	e d u
0:12:52	these is too small
0:12:54	so you systems for the reestimation by a significant
0:12:59	all by n c l is be part of the nn a waitress
0:13:03	there is no right in assigning from using y for a given in a network
0:13:08	for this
0:13:12	we use a real efficient is the input shows the system for fusion
0:13:16	we just reading writing i
0:13:19	includes your we still is involved in an a small step
0:13:24	so you're right value is yes one system
0:13:27	you'd reminding contrast to estimate ubm
0:13:31	the misuse
0:13:33	have a very similar a million this year use women right i have the base
0:13:38	a once you
0:13:42	now see the face recognition systems
0:13:47	this is there may be a front end
0:13:50	the bible any something will be different for enrollment and test
0:13:53	but elsewhere well
0:13:55	phase of that still
0:13:57	then enrollment
0:13:59	we use the reference mumbles and you the test phase
0:14:04	but overlap with the telephone calls
0:14:07	in this will yes all the faces with it
0:14:11	then we used the final
0:14:13	modeling more on the original on a small line ungrounded phase and then we use
0:14:17	that are facing varies
0:14:20	we use briefly visited those and invariance
0:14:23	you just be used every now and a snack implementations or within a face on
0:14:30	our face unless you use the one d by the implementation
0:14:34	we examine the task as a c n
0:14:39	the video but since what are based on percent is for
0:14:43	series system doesn't use score normalization for enrollment the average the enrollment and variance
0:14:50	and the test set the new animated clustering with a twenty one clusters
0:14:56	unless you listen we have several and robustness the
0:15:00	but based methods also indicated in table we have
0:15:06	you mean and variance
0:15:07	averaged and variance the median of a multi clustering so turns you form an alliance
0:15:14	you
0:15:15	maybe also balanced young ones used for in somewhere in the media we go
0:15:20	similar to his twitter that's they will i know fine inventing which is then weighted
0:15:25	average
0:15:26	all the meetings rooms
0:15:28	in the total attention we obtain a single invading for this with a weighted average
0:15:33	all the testing babies
0:15:37	but also
0:15:38	and enrollment set
0:15:41	no see the this problem model
0:15:46	we have analysis the csp markets for this experiment we used in save face first
0:15:51	one hundred and very
0:15:52	the best figure is without is not understand your is it is not
0:15:57	is not improve the low in the guns you one and it's a need in
0:16:01	the
0:16:01	well rules less in this study night in
0:16:05	you one
0:16:07	and the baseline and in the about is the
0:16:10	made in enrollment bonuses are limited clustering in the that is
0:16:15	well as in the other datasets
0:16:17	the baseline peons overall only once the contents of attention
0:16:22	there are more steam or impostors are statistics
0:16:29	we compare the different and variance improve work of the us you by the question
0:16:34	and now there was as follows we have
0:16:38	the questions all the inside phase
0:16:43	printing models
0:16:45	we use the whole or can we use a already some enrollment and omit the
0:16:51	last three test
0:16:55	area so we can see that the white gaussian is better than a form the
0:17:00	exact reason but is there a lot of in the network a very significant we
0:17:05	can see that doesn't work on my personal
0:17:14	this of the submission process
0:17:17	then used primarily
0:17:20	is a really use general
0:17:22	the only last three assumes be systems on the taste of is a well this
0:17:26	year
0:17:27	this using a system is close to the right
0:17:30	using a system is worse a posteriori because we're and based on we were or
0:17:38	generally but one best so that no one
0:17:42	analysis
0:17:44	against a
0:17:46	based on the equal error rate
0:17:49	well no that's impossible
0:17:56	this was also than one model
0:17:59	in addition
0:18:01	so for the fusion we assume that independent within that we live video these so
0:18:06	we assume this calls
0:18:08	in the figure we have a combination of more than useful single all those used
0:18:13	in
0:18:14	the additional value systems
0:18:16	single videos used in a fisherman previous nist and finally in one more
0:18:21	we can see that
0:18:22	we can get yours implement all eighty percent exactly
0:18:27	when we will from a single of assistant
0:18:29	who but it would be more efficient
0:18:35	okay
0:18:36	the key will results was using be data
0:18:40	the no more than one the one used
0:18:44	well cts less money loss
0:18:46	probably provide some woman we're got significant improvement of that a spectrum that for some
0:18:51	backends we
0:18:52	small liberal in domain
0:18:55	they can perform better than listening
0:18:58	what a probability of the screen but it was saying performance where
0:19:02	without the need for every
0:19:05	the results difference between as i the n-best and instantly in obvious that we wonder
0:19:10	why is that the is fitting work
0:19:14	so it is also studied in it has led with the transform it is because
0:19:18	the italians or entity that or
0:19:22	i mean doesn't in there is no
0:19:25	so we won't remember a city bus always focus on the same the on the
0:19:30	other side exactly you have already body was also incredibly or we don't want to
0:19:35	solve problem
0:19:37	we're really on all levels
0:19:39	i mean and variance
0:19:41	and organs performing very well
0:19:43	i mean it is obvious what is only obviously modalities are when
0:19:49	in the unimodal this so we will maybe that's came are used
0:19:55	that's all from my say thank you for

Advances in Speaker Recognition for Telephone and Audio-Visual Data: the JHU-MIT Submission for NIST SRE19

Evaluation and Benchmarking

Jesus Antonio Villalba Lopez, Daniel Garcia-Romero, Nanxin Chen, Gregory Sell, Jonas Borgstrom, Alan McCree, Leibny Paola Garcia Perera, Saurabh Kataria, Phani Sankar Nidadavolu, Pedro Torres-Carrasquiilo, Najim Dehak