Speech Transcript - Commercial applications of speaker and language recognition

0:00:15	hello everyone and as you know i'm from the speech technology it sounds and today
0:00:22	i'm going two presents some information about c company
0:00:29	well
0:00:32	the speech technology its answer was found it in a nineteen ninety
0:00:38	and it was funded by the roll call for young scientist
0:00:44	and today
0:00:47	we
0:00:51	the company has been are working on the innovation technologies during twenty four more than
0:00:59	twenty five years
0:01:01	and today we have the
0:01:04	why stuff more than a three hundred and fifty a people who are working on
0:01:12	a different here we're on the speech technologists
0:01:18	the most of them all the scientists and engineers
0:01:23	our company has
0:01:25	twenty eight p h d degree
0:01:29	the people and here
0:01:33	the mean are
0:01:35	or well
0:01:40	a d r o all about set to yell so
0:01:44	well and the about the here we have of our research
0:01:49	and zero
0:01:51	our
0:01:53	solutions for the industry
0:01:55	we are working on speech recognition speech since it is always and basically a matrix
0:02:01	and speech and that in a six
0:02:04	well
0:02:07	the rap a range or for our interest and the of course we are very
0:02:13	interested in developing a good solutions for the
0:02:19	voice back down by an entry
0:02:22	a well you can see there
0:02:26	many of our customers
0:02:29	in our in
0:02:31	in the side in the rock try or international companies
0:02:37	l so we
0:02:39	collaborate with the assembly nowhere stays for example
0:02:45	well no well no one it's more university we has
0:02:50	our department there and
0:02:54	we
0:02:57	ditch our students and say
0:03:00	p h is you dance
0:03:03	well
0:03:05	also or
0:03:07	i
0:03:08	as i and all of the you know in the right this cocoa project started
0:03:13	a near a few years ago
0:03:16	and we participated the up to
0:03:21	well this to see manage to build just well irene tedium a in the industry
0:03:29	it is the one of the leading parent it in the wall
0:03:32	or was then the one hundred and twenty a technical specialists they are
0:03:38	i have mentioned it
0:03:40	well about the minimum energy management standards
0:03:45	we are fully certified
0:03:47	i so well nine
0:03:50	ninety or one company
0:03:54	liz i would it'd
0:03:56	financials and the
0:03:58	i can say that we have well balanced mentioned to him
0:04:03	and
0:04:05	this it's
0:04:07	about the company just in fewer
0:04:12	tension
0:04:20	okay given okay
0:04:28	thank you mean
0:04:29	so now i don't know cuddles
0:04:32	it is he's going to present something about zero one week
0:04:44	so with modeling every my thank you for coming so i'm gonna talk a little
0:04:48	bit about
0:04:50	zero is like by and mainly at i think the main purpose of these of
0:04:53	this session is to do hear from you so
0:04:57	so i had this open question is what that it's what to the role of
0:05:02	in this to be in this advancing the indy scientific a community so that's right
0:05:09	my performance by
0:05:12	trying to
0:05:14	set
0:05:15	so about zeros like
0:05:18	so a lot of
0:05:20	maybe some of you know this is a you know supplier of audio and volumes
0:05:27	and that was and there it should wait and somewhere solutions so we'd about habitat
0:05:31	somewhere
0:05:33	in terms of what we morgan a whole chain for all you change from the
0:05:38	microphone from these and microphone say at combat pairs dsp to the process of the
0:05:45	application process i'm back to the lower speaker and we also what software so it's
0:05:51	is
0:05:53	it is interesting for us which is
0:05:55	no need for signal processing not suppress the on and what in my profession and
0:06:00	of course
0:06:01	speaker recognition
0:06:03	and we focus some of my solo c rs optimal by the time and a
0:06:07	can summarily applications and in this feels in speaker recognition is mostly needed for authentication
0:06:13	so we you know already we don't focus on politics it's
0:06:18	okay or a feel is mostly authentication it's
0:06:22	we domingo size components from one house and place it is
0:06:27	all spread around the world and us a euro s
0:06:31	and it is one thing i want to mention which is i think it's good
0:06:34	for this is an american
0:06:38	so the court or is very important so we have this kind of slow as
0:06:42	like silver from san the you know like that in your when you're right that
0:06:45	and you spend a boss dataset and all that stuff so it's just say okay
0:06:51	that's then school and no
0:06:54	i think within about this is that they are really the really get about the
0:06:57	employee so they and this everything for as a selling receptionist because
0:07:02	i mean
0:07:03	divided we motivated that do in your
0:07:06	you or your so if you
0:07:09	if you need to be in the scientific community in you need to collaborate with
0:07:13	other teams and you need to you know us error and the defence we do
0:07:16	here they understand that and they say okay that's it that's the way you work
0:07:20	in this way so they don't aids i mean it's
0:07:23	it's a matter of you know keeping the ideas of the of the
0:07:29	the researchers have otherwise in one
0:07:31	the one of the
0:07:33	let's see how we set thing in and that's important for this from a combination
0:07:39	where cultural is very important
0:07:41	so
0:07:43	that's all i wanted to say
0:07:44	so that all zeros lower one or several power
0:07:48	presentation is about the remote of interest in the scientific community so how eight
0:07:54	what's are you and i want to select we you and maybe
0:07:57	get some feedback any directories
0:08:01	let us know how
0:08:03	come we had but can we do so
0:08:06	so in our previous update in the industry have assessed themes and processed things
0:08:11	i idea is that they should have one foot on the on the scientific community
0:08:15	and of course of emitting state of the company and in the end to combine
0:08:19	is the one can you
0:08:22	this of what you're right
0:08:26	but egg
0:08:27	the but it's we have a so in the industry and let's say that we
0:08:33	know
0:08:34	that we are more aware of the challenges of the we use cases that a
0:08:43	magi day the reset so actually data you select two sources of innovation one is
0:08:48	the and what we call technology both and you're one is my
0:08:53	so technology post what we generate here so
0:08:56	if you're serving bs and we get this is this new technology it's gonna change
0:09:01	speaker might be it's gonna be better than before but is not really oriented to
0:09:06	any use case or any porpoise but we have also the other which is the
0:09:09	market was so maybe it's important to face noise phase reverberation i don't know what's
0:09:15	but in s use case what's gonna be
0:09:18	how is gonna be the morning five year so that's
0:09:20	that's the thing that companies may be may know more about that because they do
0:09:25	some complex and they have a marketing change from the inside looking to that so
0:09:29	this kind of information
0:09:31	and a help us define the challenge s and i think that's that maybe or
0:09:36	are our best contribution yes help to define it telling us how things to be
0:09:43	evaluate how and
0:09:47	once what have a customer a suspect team while a how the future is gonna
0:09:52	be how do we see some to set and maybe a
0:09:56	contribute to that so that's two main points and
0:10:00	something here so
0:10:01	so in this case of hadley finding the time yes
0:10:04	but also we have a full in the scientific communities i think
0:10:10	it's important then we keep it very competitive correlation so if we define very long
0:10:15	but long term goals a
0:10:18	it several okay framework to collaborate and seventy s a
0:10:22	are not very close to the problem so does
0:10:27	so that's
0:10:28	our bill
0:10:29	i want to know
0:10:30	so that's what
0:10:31	so the role of the industry being the scientific community
0:10:35	okay that's all i want to see
0:10:43	as follows now it's time and then you will have a few minutes
0:10:47	i know you can do it
0:11:02	okay
0:11:04	i'm gonna be very brief what'll people the top because i think that was your
0:11:07	you know idea that we just
0:11:09	talking twenty minutes but unfortunately the forensic people be than the low us to detect
0:11:13	so much time but
0:11:15	so the main the main goal of my speech now it's gonna be just give
0:11:19	you a brief introduction recognition because i don't know how many people here is aware
0:11:24	of we are we are they and you see we are like the tiny little
0:11:27	company between these two other guys we rst like
0:11:31	i like this evening like c as like it is small power of x techniques
0:11:35	you know that we are resisting to
0:11:37	to be guys
0:11:38	so like means a sanitised company that was created in two thousand and four and
0:11:43	it was a as an a startup of id sc-ubm grouping method
0:11:48	in the university and buddy technical university matrix actually it's funny "'cause" now twelve years
0:11:54	ago we where publicly presented in a novel use a congress the last one that
0:11:59	was organized here is pain and still to this one about back and of an
0:12:03	important little
0:12:05	so
0:12:06	so it's been a long time now we have we have three main locations married
0:12:12	of course where the heck what is are we have and of using in washington
0:12:16	where we're supporting the us business or whether and capetown where or
0:12:21	so that we can guys are helping us to solve problems like you collaborate and
0:12:26	so on
0:12:27	in total we are around twenty five people
0:12:31	as on a small company we are still growing and in fact that we lost
0:12:37	some people not so long ago move into two zero is because we decided to
0:12:43	just so sailor or
0:12:47	they mobile business of the company to zeros and now all these guys are
0:12:53	are working with them
0:12:55	and we are probably the most doesn't think that agnitio has is that we are
0:13:00	probably the only company out there that is fall fully focus on automatic speaker recognition
0:13:05	that's all we do we don't do a speech
0:13:08	we don't do any other language-related a technology we will be bored i in both
0:13:15	paramedics so we train would try to be as good as good as we can
0:13:19	for that
0:13:21	and today a or b guest book pretty nice is still in the in the
0:13:25	global market because of it with born we porting the forensic community because actually the
0:13:30	very large enough the company's a joint program between the ubm the university madrid and
0:13:36	what if you need in this pain that they just a to invest some money
0:13:39	in the group and so on and that's
0:13:41	the a priori enough of the goal but actually our forensic two
0:13:45	it was the very first problem we developed for them
0:13:49	actually i do using previously
0:13:52	that they the people talk about the level for quite awhile
0:13:58	during the presentations today according to the last interpol survey that the backend lightyear
0:14:07	we are around roughly fifty percent of the for in c class that are doing
0:14:13	the automatic speaker recognition out there
0:14:15	so
0:14:16	in the case then we are finally we
0:14:18	we got some traction and some
0:14:21	some success can be seen people that we are good
0:14:24	and i case and i don't want to extend much more you are interesting the
0:14:30	problem we're gonna be layer in the
0:14:32	in the silent so until the session so more than happy to show you what
0:14:37	we control because we have problems about also institutional but for those that
0:14:41	that we can we will
0:14:43	and
0:14:44	and as and in fact
0:14:47	we are
0:14:49	rate full of well being here
0:14:51	the embedded because
0:14:52	actually our view it's pretty close to what kind of the explain that we are
0:14:58	truly believers of the these become data give a community and we try to share
0:15:05	as much as we can then we try to the us open as
0:15:08	as we can because we really think that is the only way to make these
0:15:11	technology a evolve and the good thing is that we can provide
0:15:16	some insights from reality that sometimes is important because sometimes you guys need some ground
0:15:22	to two types and to understand what are the real problem so there and not
0:15:25	only just reduce that point one in the equal error rate
0:15:30	because maybe you are you're just reducing point one and equal error rate and you
0:15:34	are multiplied by ten the memory that you're using so maybe that's important also to
0:15:38	have that that'll to be that is okay and it works great but i wanted
0:15:42	to work in a laptop or
0:15:44	or like these guys of y two working now mobile phone that is even more
0:15:48	difficult you know
0:15:50	so
0:15:51	so and that's why we keep we keep a very close relation with many with
0:15:56	many people actually around and
0:15:59	and
0:16:00	and that's pretty much it and i hope that
0:16:05	you tell us because i want this to be not for some very long but
0:16:10	i wanted to be here we meet interactive then we would love to understand what
0:16:14	do you guys the spectrum as also because the of course we get advantage of
0:16:19	you working on these speaker we learn and then that's also always useful
0:16:24	but we also want to learn from you know and that's why i think be
0:16:28	that would opportunity to be here and discuss
0:16:31	thank you
0:16:38	that sounds and i think that basically the view that follows the is more or
0:16:44	less
0:16:45	a lot of components are
0:16:47	unless he responded so basically i think that we have to the nominees
0:16:52	to listen you and
0:16:54	or to listen that cutting
0:16:56	so the first question or the first topic we can we can talk about not
0:17:01	a long time
0:17:02	e c and
0:17:04	how in the how i can only see the relation between economy and just and
0:17:09	then we can discuss some of that if you used in these interesting
0:17:13	so
0:17:15	anyone wants to talk
0:17:22	okay
0:17:24	for lid
0:17:32	yes it is easy to explain what academic a expecting from you it's morning
0:17:38	so you know that but
0:17:41	to be a little bit more precise we need more may to do open research
0:17:46	we have good relation was industrie for applied research we all missing all of this
0:17:52	morning for
0:17:55	new innovation you know ten years we have movies unimodal because whole one of the
0:18:01	morning we have is that the two
0:18:04	close them on a project so if you want to help of you academy just
0:18:09	you to give or information b s p l c formation was maybe no more
0:18:15	than four pitch you can be your
0:18:17	and just if we if we to use the remote this he asked to for
0:18:22	a competition we would propose
0:18:25	the topic and you will to give a paid for the kids t
0:18:30	in this case we could have really open topics
0:18:34	and the we could make the market for you in ten or twenty use but
0:18:39	so
0:18:44	okay any of you want to
0:18:46	the player to just comment
0:18:53	okay that's
0:18:55	that's a little bit
0:18:57	that money is always select the tickets over some four
0:18:59	compensate i mean they are they are therefore probably we don't forget that so
0:19:07	i we can do that we can do that but it's the subways i mean
0:19:14	when you
0:19:15	i mean when accompanist a what you money somewhere
0:19:19	he they expect some
0:19:21	some redone so we have some collaboration with big universities a
0:19:26	going on which is really based on subcontract in and
0:19:31	leading then and go ahead go ahead with research and so on
0:19:35	but it's
0:19:36	it's actually unless it could be it could be possible that you
0:19:41	several companies get together and agreed to find something like
0:19:44	but you're mentioning its i don't see why couldn't be done i agree with that
0:19:50	but if there is only a few of then it's
0:19:53	it's a little bit controversial because you're funding for the comparison as well so that's
0:20:01	that's always complicated doesn't obligated is so components is only one
0:20:07	a star finally in some projects with for the academia but there are
0:20:12	involve a small groups maybe some
0:20:15	not is not quite open
0:20:17	but it could be the case that we could get together and five something bigger
0:20:22	not would be and i'm not directly but
0:20:25	they are real that
0:20:30	and i would like to add a due to that
0:20:34	that
0:20:36	strongly i strongly agree with you at that we need to give you long term
0:20:41	because i think that is t and i think that these myself i do and
0:20:45	i pretty for the disguise also have pretty clear where we would like to go
0:20:49	and ten years
0:20:52	is true that
0:20:53	that as a scholar said we need to find a balance between them between investing
0:20:58	in research open investigation and the party that any company wants to wants to want
0:21:06	to get right so that but i think that i and actually i think that
0:21:10	different companies over here
0:21:12	we are kind of all band and we try to parties actually which we try
0:21:15	to but was as much as possible that sometimes believe me is not so easy
0:21:18	now company just to get the also relation just to get popping whatever you are
0:21:23	getting right so
0:21:24	but i think it's all about the find a balance because the at some point
0:21:28	we want to help and then in the other and the other hand
0:21:33	the owners of our companies want to get some profit for from that right so
0:21:37	that's a key i think that find that balance between these two things
0:21:46	i would i would like to say well of course have to agree completely with
0:21:51	some possible to apply a dream for us but anyway a there's another
0:21:58	option that day starts to be almost mandatory it want to get funding from the
0:22:03	european union of these that most more involvement
0:22:06	on the core of the companies only thing
0:22:10	initiatives
0:22:10	because with a i'm talking about you from a europe is pretty perspective
0:22:16	with f p seven framework programme seven of the european union
0:22:20	a the big projects where mostly the well it rate it will escort when huge
0:22:27	enrollment all companies was present with a rice and twenty the new paradigm for your
0:22:34	parents funding
0:22:35	the involvement of industry have become has that's even increased
0:22:39	and the many of the proposals have rejected because they are not space that the
0:22:45	that those places this a but they are asking for companies to lead the proposed
0:22:49	accordingly the proposals to get more involvement in the proposal
0:22:53	not just as integrators or facilitators but just four
0:22:58	doing the research themselves
0:23:01	and generate a products that are
0:23:04	i the they're going to be integrated and on a high-level at entertainment as it
0:23:09	a little bit
0:23:10	so
0:23:11	perhaps a good the mum
0:23:15	middle point between something or research
0:23:19	totally and a doing what we're doing currently which is collaborating as a subcontract would
0:23:25	be to get the mobile then in this kind of initiatives with which of course
0:23:28	there risque because sometimes they are not found that many times the other funny there's
0:23:32	truly competitive
0:23:34	at least in my i'm at the environment around
0:23:37	most of the time i see a big project grant it is because the companies
0:23:43	cody at eight we have this speaker identification integrated project now from f p seven
0:23:48	that is correlated by parent
0:23:51	that's a good example that's when companies gets a huge involvement of this then we
0:23:56	can expect to have a lot the funding for lots appear the students and whatever
0:24:00	else but it eight in it it's a for in my opinion it
0:24:03	it makes sense that the companies assume that risk that i would like to hear
0:24:08	the pin the of the companies
0:24:11	yes then you have to code in the template so induce the academy and government
0:24:16	or european okay
0:24:19	so many common about how government or you'll open that case
0:24:24	so that's
0:24:25	i mean i fully agree i in fact
0:24:29	assets the company recognition is a
0:24:31	we think that we are we're getting to h twenty that we didn't we didn't
0:24:34	do enough if be sure that and i'm gonna talk now from a comedy or
0:24:39	perspective not from being just in general because i'm i don't know many others but
0:24:44	for is huge afford only because as a twenty five people company and
0:24:51	be part of the combination of a project like these that you know how we
0:24:55	these that usually both like thirteen different companies from seven different countries you know that
0:25:01	it's creating only that discrete in like point emails per day and
0:25:06	and let a mum is it's tough stuff but this is so that we want
0:25:11	to be there and we're moving in that direction but is to
0:25:16	i don't know if this so much about
0:25:19	have been a common a big corporation the leaving the a the initiative it's more
0:25:24	about
0:25:25	been able to percent rio the usability for whatever you but we i think that
0:25:31	that's exactly that i think that the european union is able eighteen very possibly his
0:25:36	that
0:25:36	the capability that you have
0:25:38	to put whatever you the below
0:25:40	i'll they're doing things
0:25:43	use useful things right and i think i think that of course for that and
0:25:47	industry should be should be also uses there are some in users also or here
0:25:53	and i think that it's also important that they get involved as well in the
0:25:58	case of for in six the forensic labs in the case of
0:26:00	some other and billions the corpus is the government agency in the case they of
0:26:09	i identification of all of citizens the proper government brian so i think it's also
0:26:17	three party thing is not only about the you guys researchers we guys the company
0:26:23	but also the in users that also be about the about proposed to the project
0:26:31	i am also
0:26:33	okay like to comment it's
0:26:36	i completely agree with you but it there's
0:26:39	one problem we also hired in this in discourse a it's
0:26:43	that is that it's a lot of overhead
0:26:46	yes for going to the court so that's
0:26:48	that's a big running takes a long time a needs
0:26:52	sometimes from the company
0:26:53	point of view c then someplace if it's even easier to identify
0:26:57	the rather than
0:26:59	you know that go ahead then try to spend all these maybe from nothing
0:27:03	the problem the problem you
0:27:05	you have they're also is
0:27:08	is that
0:27:10	and i think essential we have in a in our side in the industry here
0:27:13	we have to fix it which is a
0:27:16	we should look into this pre-computed discolouration more but in the industry side so i
0:27:22	i always see this okay this competition in this utopia riots where you have one
0:27:28	or two company leading someone prior than the other two competing and it's maybe for
0:27:32	this one or for the other one and they don't get together and they have
0:27:35	a different same thing so it's
0:27:38	so that something some we may need to fix on our side so far were
0:27:42	computed experiments but i
0:27:43	i four that we would you that that's
0:27:45	problem either graph to get
0:27:47	to get well all we want
0:27:50	we all
0:27:55	okay any question
0:27:57	or comments down
0:28:04	so i ask a question i'm not sure if you're grew up there probably you
0:28:08	be unhappy with his question but
0:28:10	if you're looking at maybe in the feeling forensics and so forth there are researchers
0:28:16	the mean i in this room especially in speaker recognition space that in focus more
0:28:22	on the sre
0:28:24	but i think if you look at bs very in you look at maybe some
0:28:27	of the work in the forensic space there there's some more at their connected but
0:28:31	disconnected as well and also think that the ability to get access to realistic data
0:28:37	in the field is very challenging i mean it's not easy to get her case
0:28:42	data
0:28:43	to be publically released so i'm wondering from a company perspective
0:28:48	the willingness of your organisations to be able to share your data even if it
0:28:53	means it's in a controlled setting maybe of some type of work
0:28:57	other user license that allows individual labs to get access to that i would be
0:29:03	very important because i think as dark pointed out you know data drives a lot
0:29:09	of our work and i still think that there is a somewhat of a voice
0:29:13	separation if you will between the technologies
0:29:16	and the forensic speech scientist and i think access to the data will actually help
0:29:20	people on the technology side released on the research side in a lot better way
0:29:29	we're going
0:29:32	so
0:29:33	a probably this is the politically incorrect to say here but i fully agree that
0:29:38	most of the research community for the last five years
0:29:43	in my opinion is to
0:29:46	sre by s
0:29:48	because the most of the algorithm that you find here are focus on optimize whatever
0:29:54	we sold you get in the whatever condition of whatever evaluation all mixed and
0:30:00	and is true that
0:30:02	reality differs a lot from that i mean it's a good
0:30:06	you try having everybody log data and anything that needs has been doing a huge
0:30:12	work to provide data to everybody
0:30:14	but of course that is only representing a small portion of the different conditions that
0:30:19	you can find out there
0:30:20	and even for as we struggle to get data like that and of course we
0:30:26	have our own our own databases we try to build them but
0:30:30	but sometimes
0:30:31	is not chasing because most of the in users and for the morning for in
0:30:35	six
0:30:36	they are not so easy to get that data to share everybody and for s
0:30:41	and it's true that they are like a few initiatives from the use of for
0:30:46	instance what inferior the created a database to really make a i think that they
0:30:51	the d the something similar
0:30:54	but is to that
0:30:55	the idea still it's like what in jail was present in it it's just elected
0:31:00	a small databases is more for and i know and you know that there is
0:31:04	tong update after real data out there sitting
0:31:08	sitting down and
0:31:10	and is true that we need to do something because it is true that the
0:31:13	end users
0:31:15	that's not help
0:31:17	both community i mean you guys m and as guys it's very high it will
0:31:21	happen to advance and we can get
0:31:25	perfect in the in condition a b and c all means there's hardly twenty overall
0:31:31	weight that
0:31:32	what that means for really make sure so i know obviously if you were between
0:31:37	us or eer we see the movement if you're working at work or
0:31:42	mated looked at for speech no wild infrared dark speech to all receive are moving
0:31:47	things into more naturalistic data but you believing those corpora they show variability but they
0:31:54	don't be me not necessarily reflect word you would see sometime german in the voice
0:32:00	forensic side of things ensures
0:32:02	somehow would be nice you can see more about him and i'll just full disclosure
0:32:07	unbiased because with the ozarks speaker committee responsible for trying to work to find out
0:32:13	word data could be used for the or sex or
0:32:22	resisting
0:32:24	you know in our company would be always the something agreement with the customers and
0:32:28	we
0:32:29	and not i don't know abilities to redistribute usually this to you builds is data
0:32:36	for the
0:32:38	and community but
0:32:40	i agree with the u and of course i would like to sing these survive
0:32:45	for the speaker went in the wild challenge because it's which was agreed to punch
0:32:50	unity to investigate this key some data just noisy and or different conditions
0:32:59	in case of our company we also participate in the air but
0:33:05	the data we have is the usually unlabeled to and what we have also have
0:33:12	some issues how to use these in our research number topics and we usually
0:33:21	can get only the models will we can get the
0:33:27	we forms all the data so and
0:33:33	in my opinion it is maybe it's a good idea to share i-vectors why not
0:33:41	because no huge just a minute it's possible
0:33:46	well
0:33:49	okay so
0:33:51	we all wear what we'll over a decision is
0:33:55	for surplus have you hear
0:33:57	and thank you your opinion and comments

Commercial applications of speaker and language recognition

Industry & Forensics Track (Short Talks + Panel Session)

Sergey Novoselov (Speech Technology Center Ltd., Russia), Carlos Vaquero (Cirrus Logic, Spain), Antonio Moreno (Agnitio Voice ID, Spain)