Speech Transcript - Speech technologies for data mining, voice analytics and voice biometry

0:00:15	so thank you for a nice introduction
0:00:18	my name is but the shreds and the
0:00:21	at the beginning i was recessionary the banana what is it the of technology
0:00:27	i for one many different fields start thinking of porting recognizer
0:00:33	then
0:00:36	we really try to do it should speaker identification don't know asr
0:00:40	are you
0:00:41	roll the speaker is no particular isn't enough
0:00:45	that many of you have still using
0:00:48	but and down the here
0:00:51	two thousand five
0:00:53	something like a stranger happen to of for us a basically we do our approach
0:00:58	by one company and the
0:01:00	the companies
0:01:01	set to will give you some money but you would like to have different license
0:01:07	for recognizer it was publicly available
0:01:10	also we said okay this is find a do this
0:01:13	bill help us to
0:01:15	the to finance the there is an essential but the third row to be
0:01:19	quite the low expediency to go also
0:01:24	nine months just to the
0:01:27	negotiated the license university
0:01:30	and we realise two things that there is interest from commercial market it can be
0:01:36	recast additional money
0:01:38	and
0:01:41	that we need to do we the heat a better way
0:01:44	so there we started a company called for X
0:01:49	so i would like to talk about to the topics are the two main topics
0:01:54	today
0:01:55	how to the speech tickle you probably such a two
0:02:00	mark and the and then a i would like to the
0:02:05	shows are much related that we see a problem the user point of view
0:02:12	so at the beginning i will talk a few words about the company
0:02:16	then about the text widget use cases
0:02:19	technologies that are behind the our programs
0:02:23	and the so
0:02:25	how big over the technology to the rose and then i mean really indicates someone
0:02:31	grand challenges
0:02:34	people usually don't know what is that you speech
0:02:39	but if you look at this
0:02:44	at the
0:02:47	at a dislike so you can see that they result
0:02:52	that there is before make sure the about speaker it can be
0:02:56	E hundred and there are you gonna be eight sure it can be speaker i
0:03:00	didn't at you can't that example emotion states
0:03:05	meant the speaker speaks and so on there is the goal that the
0:03:10	you can detect the language you can detect dialects a keyword crazy so
0:03:15	you can do the whole speech transcription
0:03:18	maybe the topic is interesting
0:03:22	you can
0:03:23	do something domain incapable then the
0:03:26	but there are other modalities you can have some information about the
0:03:33	and white men the that the speaker is
0:03:35	to whom the speaker speaks you can have other solid so
0:03:39	what is up of user you go
0:03:41	very close animals
0:03:43	or you have a lot of information about equipment that was used
0:03:49	of the to get a relief what we on voice it can be the device
0:03:53	the
0:03:54	for example we also for ticket be transcription of huge it can be according to
0:04:01	you may be the test it in a speech quality
0:04:04	this
0:04:05	is important for user and the users can benefit from this information
0:04:12	about products E R
0:04:13	it also based in two thousand six as us
0:04:16	startup from brno university of technology
0:04:20	it has C
0:04:21	in czech republic in button or just five minute walk though from the university come
0:04:26	rules
0:04:28	if we speak about the user so we have currently users in more than twenty
0:04:32	come companies
0:04:34	so i've got with the agency score said that all the bank
0:04:37	dell corporation that also brought up so service providers and others
0:04:43	the company use the roughly table and the
0:04:47	the little small only for an external funding so those so far
0:04:53	you if we speak the about the process how to transfer technology it probably search
0:05:00	the to the progress and the market
0:05:04	there are several steps
0:05:07	well i'm an important role E V speak about the research
0:05:11	theirself useful for a dollar a universities or inside companies
0:05:16	but the goal is to get the
0:05:20	the best like technology but the more i don't interest is unit could easily the
0:05:28	quality so like a or a set of for the stability of speech will for
0:05:34	the court is not do
0:05:37	the man it main importance
0:05:39	and the what was important for a so it probably also for
0:05:44	you
0:05:45	and this stage we will to be all limit as possible so it's the beans
0:05:50	quotative measurement
0:05:52	to
0:05:52	to the saba
0:05:54	okay so
0:05:56	open-source toolkit and so on
0:05:59	but then you need to getting better technology to
0:06:05	user somehow
0:06:06	this also for this you would be to do
0:06:10	next step but
0:06:11	you need to build the
0:06:13	code base that is or almost that is stable
0:06:17	it is fast the has a modified a P has documents a day shown a
0:06:22	proper licensing
0:06:24	us assume so the D V D this is what is what exactly
0:06:30	then that is others that but
0:06:32	a better you know
0:06:35	i need to build product for our customers so you can have nice technology you
0:06:39	could have nice interface is but if you don't have power back to you won't
0:06:42	be able to sell
0:06:44	so here the full cost is
0:06:48	functionally the and that this is donna either by phonics el or by other companies
0:06:53	we never be
0:06:58	not now i the bill mentioned
0:07:03	pretty domain use cases
0:07:06	or pretty main customers there are others but i selected this free
0:07:13	the first the are all sentences in course of course there are so we are
0:07:18	active you know why rasta the fires are is the quality control how to ensure
0:07:26	quality parentheses in
0:07:29	the call the course of terror and there are there is data mining from
0:07:34	voiced i think
0:07:38	the that the quality control what is it about
0:07:42	in both antenna
0:07:44	it you have to really
0:07:46	some kimberly there or supervisor
0:07:49	that's a pairwise well the
0:07:52	then a but i there are so i just this better than do
0:07:57	but i think of course
0:07:59	evaluation of operate that also
0:08:01	analysis of the results
0:08:03	for the team and some reporting
0:08:06	if the there are no speech technology
0:08:09	so you usually only three by a set of recordings is
0:08:15	inspect the then the use of to control wanting to local schools said there but
0:08:20	if you would be point something which is you are able to control
0:08:24	a hundred percent so the topic to get you are able to better use statistics
0:08:28	the
0:08:29	and the this everything is about the
0:08:32	moreover
0:08:33	the cost so of star but
0:08:37	we are able to try to reduce the number of advisers to how well
0:08:41	over operating costs
0:08:44	so it is very you are in but
0:08:47	to shorten the call so
0:08:50	if it you are able to the
0:08:54	you have
0:08:55	find problem so you really errors
0:08:58	some or but i just are not the us up despite the or a remote
0:09:03	well train the
0:09:05	i hear its is possibility to said that what is look training can that
0:09:11	usually a the
0:09:14	the formulation of people you know such posts of the rest is
0:09:18	you know tens of
0:09:21	paris and so we are able to reduce the this
0:09:28	the this amount of data and the again six of some calls
0:09:36	and about this approach is a huge the for quality control the main technology is
0:09:42	the
0:09:43	the at and doing some
0:09:47	this does so i'm not easy so on
0:09:50	topologies
0:09:52	so you are unable to get important statistics
0:09:55	like but better
0:09:58	the dialogue starts the number of speaker turn
0:10:01	us
0:10:03	speech adaption times
0:10:05	of unknown the call centres have
0:10:09	all the equipment so if they have some channels to the conversation the speakers are
0:10:15	not in separate the
0:10:17	well i like channels that we need to do diarization
0:10:22	then it is possible to deploy the key you want to raise the text order
0:10:28	you have some obligatory phase this you don't want the people to
0:10:33	speaker all the words
0:10:36	you would like to have some
0:10:38	of course grip compliance the people should
0:10:41	for all calls three the
0:10:43	and the it is possible at the to the voice speech transcription are mainly
0:10:48	for us
0:10:49	set a should
0:10:52	in this task
0:10:55	every about the data mining the this is other large a topic for also that
0:11:01	is
0:11:03	year again we have like two subtasks
0:11:07	one is but i mentioned of the corset errors
0:11:10	of course and therefore overloading
0:11:13	you may gina that you have
0:11:15	also there are of a few hundred people
0:11:18	and the there is large up
0:11:21	i rolled each of thousands of people start holding the
0:11:26	to all said that the service so you need to whatever it really quickly
0:11:30	we need to explore the export of what is wrong and the
0:11:34	maybe that some information to do initial i we are
0:11:39	a stage
0:11:40	and the japanese so you could be the this is solved by some be
0:11:45	screen
0:11:46	in the call centers showing the topics that are just the discussed that the
0:11:51	percentage
0:11:52	then
0:11:53	i other important the
0:11:56	but you use the like the
0:11:59	i did value speech technologies for business of eight basically now a moan about so
0:12:06	but
0:12:07	indeed i don't know if you have for example how
0:12:12	may be done also
0:12:13	it's looking for places that too but i'm not
0:12:17	in new
0:12:20	new fast foods
0:12:21	you the approach is to but the because go to telephone operator and they ask
0:12:28	please could you
0:12:29	if a statistics of where people that the visit it the our fast foods are
0:12:35	putting day and the
0:12:38	the place where is the highest of that it was good consideration is good place
0:12:42	to start the you fast food
0:12:45	but the you know speech technologies the same
0:12:48	if you have more information for example it's on the phone line a is a
0:12:52	male or female or that they are more people or on the line or the
0:12:59	pairs and was interest in the in some regrets the in boston
0:13:04	it helps you to go
0:13:06	the whole business certain to push to some more
0:13:10	for this we usually use of speech transcription
0:13:14	and then
0:13:15	some of data mining to on top of it so
0:13:19	of course it is possible to at the
0:13:22	so i changing if you want to the session a narcotics
0:13:28	then
0:13:30	the other big groups are bangs
0:13:33	a bank so of colours have all sentence so what i dimension on the past
0:13:38	two slides
0:13:39	is important also here
0:13:41	but the other too large task
0:13:45	was a box are the bands needs to ensure the security part so on the
0:13:54	other side
0:13:55	they need something that these
0:13:58	breeze and the
0:13:59	for the user does that at the
0:14:01	the that doesn't to
0:14:03	being
0:14:05	much complications
0:14:06	so here the voice by but lisa very interesting it can be ways by timidity
0:14:13	using a cape race or it can be ways by comedy that is the dominant
0:14:19	on a big get the using a text independent speaker identification system
0:14:25	and then i other
0:14:28	task is for our protection in major in there are people according to bank so
0:14:35	for example a hour a day
0:14:37	i shows we make i didn't theses and that today are
0:14:42	requesting clones
0:14:45	you it is how to detect that this that if you don't have technology
0:14:49	but if you have technology like the speaker identification is a really simple
0:14:56	now about the intelligence agencies
0:15:00	the intelligence agencies the situation it is usually that the this intelligence agencies have
0:15:07	they really huge amount of data
0:15:11	the amount this i should that the
0:15:13	they are not able to put to see the
0:15:17	manually this can came from
0:15:19	a big use of from telecommunication network and communication the internet and so on
0:15:25	they are looking really for need to
0:15:27	you know i say it's take a
0:15:29	and for these it is possible into to use combination of technologies
0:15:35	so we are using combination of technologies
0:15:39	and you language identification agenda to speaker diarization keyword spotting speech transcription
0:15:46	data mining tools
0:15:48	and also a little fun a correlation with some other metadata for example from this
0:15:54	text is used
0:15:56	and so of course the sequences are very interesting you know operation and forensic speaker
0:16:01	identification
0:16:04	now i will go to be better to the technologies
0:16:09	and the tell you what this important for a practical deployments
0:16:16	here are some of the technologies i want to speak about all of them you
0:16:21	can come and ask if you should it from a question
0:16:26	about the voice activity detection
0:16:29	i would say that this is the most important part therefore practical deployment
0:16:37	you can see our process
0:16:41	by this is the most important part
0:16:44	you can have
0:16:46	very nice results for example on these databases
0:16:50	but
0:16:51	you what do you will explore a target so
0:16:55	the users are working if such channels that the you should quantity of the traffic
0:17:00	it can be
0:17:02	tens of a sense is not speech at all
0:17:05	it is some technical signal for like dialling don so that six
0:17:10	and
0:17:12	everything can if you don't have that
0:17:14	this
0:17:16	built in
0:17:18	it it's a really harder to work with such channels
0:17:22	so we are using energy based the steam of would be eighties energy based
0:17:26	the at the beginning the to remove very large portion of the silences
0:17:32	then technique the signal removal like the tone detect to removal first
0:17:39	back to the spread like that are in mobile station i
0:17:43	signals and so on
0:17:44	and then we have a vad based on F zero tracking
0:17:49	because of the speech have the specific characteristic the that should be a we have
0:17:54	zero
0:17:55	and the
0:17:56	and the respective behaviour of this
0:17:59	F zero and then we have you wanted what based vad
0:18:06	to get a very precise the segmentation
0:18:12	but is this say sets it is very important technology and they are still many
0:18:16	challenges
0:18:19	so it is important
0:18:21	the accuracy of media
0:18:24	directly affects the accuracy of the technologies
0:18:27	us some sort are actually trying just you can have music
0:18:31	of
0:18:32	that they're other speakers sounds of like people tend to
0:18:37	well a four or something like this
0:18:41	you have a an alignment silence
0:18:45	use a different technical signals
0:18:47	what is a challenge is the vad one variable well snrs
0:18:54	we at on distort each section was
0:18:57	well what is also through you that important i think it is unknown parameter to
0:19:02	a
0:19:03	non automatic way
0:19:04	of green vad because we know that we can do it's one to the deep
0:19:08	or specific channel
0:19:10	by training just as before
0:19:12	some good classifiers
0:19:14	but how to get a rise this it is still difficult and of colours distant
0:19:18	mikes
0:19:22	and well the language identification
0:19:26	currently if we are able to recognize about fifty languages
0:19:31	and the
0:19:32	what is even more important
0:19:34	that's that the user can add a new language
0:19:40	themselves
0:19:42	this is important especially for the intelligence come community because today will never
0:19:48	tell you able to instead of the languages are great interest the that
0:19:53	what the correct you on sent to you won't be able to collect the data
0:19:56	to have much easier X axis
0:19:59	to such data
0:20:02	we are using i-vector based the technology
0:20:06	and it is commanded training okay and
0:20:09	we have first of all men which means that
0:20:12	the language print is a less than
0:20:15	well
0:20:16	"'kay" a record
0:20:18	bear
0:20:21	a better file
0:20:26	do it in this is the technology behind the there S this several stages
0:20:34	year we have feature extraction
0:20:37	collection of statistics using the ubm
0:20:40	usually the
0:20:42	we use gmm and the that is
0:20:46	are aesthetically the by some subspace so the subspace it's estimated on large quantity of
0:20:52	data to model the
0:20:55	for
0:20:57	variability in the speech
0:21:01	so in that we get the
0:21:04	of estimate so for but when we are in the subspace
0:21:09	so this part was prepared by for next year
0:21:13	but then and there is other part to
0:21:16	but that it is the classifier of languages we use a multi class logistic regression
0:21:22	here
0:21:23	and that this is done by using
0:21:31	about speaker recognition
0:21:34	the there are many task like speaker verification a speaker to
0:21:38	set of speaker spotting link analyses
0:21:42	for
0:21:42	after normalization some house
0:21:44	sometimes social network analysis
0:21:47	we can work in text independent or text dependent more
0:21:51	i-vector based the approach
0:21:55	we use diarization
0:21:57	i think it what is important here we have
0:22:01	use a based the system training for calibration
0:22:04	that again helps
0:22:05	people a lot
0:22:07	it is here
0:22:11	a so that the use of the same as a in case of
0:22:15	and which identification
0:22:18	what about a year we remove other what it but it is on speaker variability
0:22:25	i would be have some normalisation of ways pain so simply by
0:22:29	mean subtraction that can be done it user side
0:22:33	and
0:22:34	then
0:22:36	if we have well
0:22:39	scoring
0:22:43	we compare ways putting it's a
0:22:46	this
0:22:48	pretend the by one excel and it is do you get it but the
0:22:53	we allow our user to
0:22:56	the rain or i don't the this classifier
0:22:59	this is very important because
0:23:03	it's harder to get any recording for from clients
0:23:07	but
0:23:08	if you deliver a such system to clients the
0:23:11	and they are able to adapt the system you the amount of data can be
0:23:15	a really small it can be
0:23:17	for example fifty speaker does just few recordings of each
0:23:21	well i'm
0:23:22	we saw that the
0:23:23	a normal telephone channels that we are able to get about the forty percent improvement
0:23:30	for
0:23:31	the new deployments
0:23:34	and if it is about some
0:23:36	us special of or
0:23:39	for example many directions
0:23:40	we saw a hundred percent improvements just
0:23:44	with this
0:23:45	simple book
0:23:47	and of course so what is closely you
0:23:51	important is calibration
0:23:54	of
0:23:56	you know case that we are drinker like that and then calibration because that this
0:24:00	is also not seen too much in
0:24:03	and nice the because they're the recordings are
0:24:07	about two and half minutes long the but if you have
0:24:10	the huge
0:24:11	but variability in do lying to you need to do anything three this L C
0:24:15	the shore recording studio
0:24:19	solve it
0:24:20	by do some up for a times
0:24:25	what are the challenge in a language identification and the speaker identification
0:24:33	i think the that the main challenges are very short recording so it can be
0:24:39	one less than a three seconds
0:24:41	but the
0:24:43	very important for us is
0:24:46	keeping to the training corn a user side
0:24:49	because why less than three seconds to each if you have department of speaker identification
0:24:56	and the you would like to deploy eighteen bank the people don't want to speaker
0:25:02	they would like to have
0:25:04	the decision even before they start speaking
0:25:07	so i would say that the ten second these
0:25:12	the maximum of
0:25:14	a line that it that is a set the
0:25:17	and you really free second to for a verification
0:25:21	you can do we the we but text dependent systems
0:25:25	i is harder to do it with text independent systems but in case of text
0:25:29	independent systems
0:25:32	these two steps are report to study on background so bias it to do use
0:25:36	user is talking
0:25:38	the operator
0:25:40	of
0:25:44	i that is question how to ensure
0:25:47	accuracy over large number of acoustic channels and languages
0:25:52	the technologies are more and more general
0:25:55	independent but there still is someone
0:26:00	independence
0:26:02	what is was link important there are a graphical tools so that how
0:26:06	the user is to visualize the information to do the calibration
0:26:10	because
0:26:12	if you do want to do this for user the user will never we the
0:26:17	in self
0:26:19	what we see also very challenging is
0:26:21	language identification and we could ideas deviation a no voice over ip networks because
0:26:27	there you have pockets you have gets lost
0:26:31	and the you if you have this costs
0:26:34	you usually cortex a are doing something that they are either put their zero also
0:26:39	okay are sensing is i speech
0:26:41	but this is not so the speaker to the that it this is something that
0:26:45	was
0:26:45	generated by decoding
0:26:47	so that's also it's very important topic
0:26:50	and of course the distance might
0:26:54	now i would
0:26:56	say few words a ball so diarisation a because that this is very important technology
0:27:01	you useful for example the call centres but model also
0:27:06	for anger
0:27:10	other users
0:27:12	we are using approaches one approach is really possible the not so
0:27:17	much weaker the
0:27:19	this is approach are based on
0:27:22	clustering of i-vector so we basically split the audio too small chunks and to do
0:27:27	clustering go for i-vectors
0:27:29	but the
0:27:30	then the i don't take the
0:27:33	fully bayesian approach to the initial you know might take the by
0:27:37	fabio one a
0:27:39	patrick kenny worked with this to
0:27:42	on the reset assures
0:27:44	quite with the text and it it'll be the
0:27:47	D P this is approach to bear you don't do a heart decision
0:27:55	during of the process of you have everything good
0:28:03	probably sticks and the you are going to do the decision and at the beginning
0:28:06	it at the end
0:28:10	this approach is i would say
0:28:14	but if you're at the but you want see
0:28:17	well my next slide it's not
0:28:19	fully to
0:28:20	but
0:28:21	memory cons i'm mean and the quite small
0:28:26	so what are challenging
0:28:28	so in diarisation
0:28:31	in my point of view the diarization
0:28:34	still technology that needs quite a lot of research
0:28:39	really so that it is very sensitive to initialization
0:28:44	it is very sensitive to
0:28:47	non speech sounds
0:28:49	do you usually it is about the wall so you got more gaussian before
0:28:55	for example if there are you sure that the
0:29:00	new sounds that you haven't seen in your training data needs to be sorted for
0:29:04	example we ask
0:29:06	the system
0:29:06	to keep two speakers
0:29:08	but the output was so the that the
0:29:11	we got two speakers in a
0:29:14	one or like
0:29:16	like under one labeler and the second speak that are you know
0:29:21	there were segments
0:29:23	it was other us
0:29:26	speaker sounds
0:29:27	i think of a lot in this case the
0:29:31	so what is important the it's a very
0:29:35	would the duration of your vad if you have
0:29:39	i just sounds the that the speech
0:29:43	it can hardly due to the adaptation
0:29:47	a so it's a it is but very sensitive to two so that speech and
0:29:51	also and then which is
0:29:53	what we see that the you human of with this is things systems you can
0:30:00	easily very should diarisation error rate the
0:30:02	close to what one percent the one is data
0:30:07	but
0:30:08	well what we also saw
0:30:10	and the
0:30:13	you the
0:30:15	first it's us
0:30:16	we could be that the rest of the one percent the
0:30:20	is that the there won't be pro by means segmentation about it's fails so
0:30:26	forty four for this recording good did this is the
0:30:30	usually of speaker to sweep but very similar voice but this happens
0:30:38	okay so i think there is a shana
0:30:41	that was a lot done during past two years but the data
0:30:45	the challenge is quite the
0:30:47	and of course you can speak about
0:30:50	text and distant mikes for
0:30:53	the of
0:30:55	processing cove of
0:30:57	the or like for example of deviance
0:31:02	it in both keyword spotting
0:31:05	so we are
0:31:10	the we are using approach is what one approach is something probably you know all
0:31:15	few
0:31:16	no see the
0:31:18	probably from project
0:31:19	is the lvcsr based the keyword spotting
0:31:22	is this is very good
0:31:25	but the
0:31:26	small
0:31:28	and it's expensive for development
0:31:30	the other keyword spotting the that the T V are using this acoustic bass the
0:31:37	year indifference the that the
0:31:39	it here we usually use a larger acoustic model
0:31:42	here it's a simple on your network based acoustic model
0:31:50	the there is no language model or data simple language model but here it's much
0:31:56	cheaper of for our development
0:31:58	so in case of
0:32:00	lvcsr we are stopping creep hundreds of hours of training data
0:32:04	in case of plastic you want splitting a
0:32:07	we are stuck used by the office of acoustic data or human less
0:32:14	the speech transcription a what we are using a
0:32:18	this is probably not important of all of you are working can this
0:32:22	feel that
0:32:23	we are using the system based on a
0:32:26	bottleneck features that the can combination we've other features hlda vtln
0:32:33	gmm based system or and your network based system is not explaining okay
0:32:38	speaker adaptation
0:32:39	and gram language model and generate the
0:32:42	usually confusion networks
0:32:47	what are the challenging
0:32:49	from the deportment point of view here
0:32:54	well of course the accuracies
0:32:57	still important
0:32:59	but i would say that the it's not so the most important challenge
0:33:04	the challenge is us be the
0:33:07	lower memory consumption
0:33:10	how to train new systems for the automatically course we would like to do it
0:33:14	for
0:33:16	so how to donna
0:33:17	hundreds of recognizers in a parallel
0:33:21	before all compute efficient computation one
0:33:25	resources
0:33:26	and also
0:33:27	how to
0:33:28	to the lecture normalization is
0:33:31	speaker
0:33:32	adaptation of for any length of
0:33:35	speech utterance
0:33:37	course for example if we transcribe
0:33:40	along all source lectures some whatever
0:33:43	we try to put the
0:33:45	this much a adaptation is possible but if you are working with very short the
0:33:50	segments so like
0:33:51	three seconds or less
0:33:54	the adaptation
0:33:56	below heart how do you and the usually you will see worse results
0:34:01	but
0:34:02	the system
0:34:03	that was so one solution is to remove those this adaptation
0:34:08	but
0:34:09	the system to be less robust to train
0:34:18	not now of a how to sell of speech transcription
0:34:25	what we found that if you must speech transcription
0:34:29	and the you want to sell this technology is quite heart
0:34:34	you need to have but
0:34:36	something that this on top of this technology at that the real presently information to
0:34:42	users
0:34:43	this is the
0:34:45	because
0:34:47	there is too much text
0:34:49	and
0:34:52	this that there are still some errors
0:34:56	what is our experience that the
0:34:59	the user
0:35:01	bill never be happy about the accuracy of those
0:35:04	the speech recognition system if there are errors in more so the uses to mention
0:35:10	this are also
0:35:11	you if the words are correct the data start combine of a preposition suffixes in
0:35:16	this is correct the
0:35:17	a star complain about the some punctuation marks or grammar
0:35:21	this is but
0:35:24	if we use so
0:35:26	of for a summer
0:35:27	and the representation how to look at the data
0:35:31	bill help you
0:35:32	to sort of technology and that we are doing the in such weighted maybe do
0:35:37	configuration
0:35:38	we've the existing test bay it takes a base data mining tools
0:35:42	integration is donna
0:35:44	you usually on that the level
0:35:47	all of
0:35:49	of confusion networks also we have also the other a captive audience
0:35:57	this is one a to the to use of
0:36:00	this was the double
0:36:01	by our part company so like
0:36:04	so you have set session in gina
0:36:06	here you can have very complex squarey here are
0:36:11	documents it's a better found
0:36:13	the document
0:36:15	but you need to
0:36:19	bright some somehow the query the query can be very complex
0:36:23	so is here is
0:36:25	gladiator
0:36:27	but the was so it's one possible ability you but if you want to
0:36:33	we have described topic so you can use this but it there
0:36:37	or you can go from update time you can result i still they are you
0:36:41	can look at what is the correlation among works
0:36:44	and the
0:36:45	you can you can
0:36:46	take this could happen automatically two classes
0:36:50	well i mean you have these so you can
0:36:52	here are need the correct someday time
0:36:55	then train
0:36:56	statistical based approaches
0:36:58	or if you can deploy stuff for example to see you what is the
0:37:05	how well
0:37:08	the topics
0:37:08	and morph in time
0:37:14	of the input
0:37:16	not now i have
0:37:19	two slides so how we transfer the call
0:37:22	what is the time please
0:37:27	all cases so it each
0:37:30	okay so it's we just quickly
0:37:33	this is how we
0:37:37	to transfer the call so it to use this at the i think in a
0:37:41	two thousand seven wanna be decided to write our speech for the of score
0:37:46	the well the reasonable so that you wanted to have something very
0:37:51	stable very fast that
0:37:53	and the before the proper interfaces
0:37:58	the speech for it has morgan two thousand five hundred topic objects coding or the
0:38:04	hour of speech processing go
0:38:06	it is a more then
0:38:10	minimal first lines of source code and it ceases steely
0:38:14	still use it might enable
0:38:18	how
0:38:21	be approach to refer to the recession
0:38:24	we it the research is usually done using standard tools like that
0:38:28	S T K in a car be by transcripts
0:38:32	i think it it's all through the nose these two kids that this is
0:38:38	for of
0:38:41	hmms reconnect the but this is for neural network training good colour be it's made
0:38:47	in the by then pour we
0:38:49	and so and so on
0:38:52	but the that diana we can the to use our code base
0:38:58	and we can implementing new system and a two hour speechcorder quickly in a
0:39:05	just two days the
0:39:06	well final nor seen a single line of C plus court is written
0:39:12	everything is don
0:39:14	flew configuration file this could do this configuration file
0:39:18	can
0:39:19	look like this
0:39:20	you have some objectivity this object so
0:39:23	well this description is the map
0:39:26	two
0:39:29	C plus interface the user to set functions
0:39:33	and then i we have some framework out to connect to be subjects to better
0:39:39	so some
0:39:40	you of fun we have four or the artemis entity
0:39:43	but if you need a algorithm we just goal and to buy one simple chip
0:39:49	for simple objects
0:39:52	a about interface is
0:39:54	what
0:39:55	the customers are used to
0:39:58	a locally specific interfaces
0:40:02	us
0:40:03	so i don't want to change data bits is so we
0:40:07	men then the double the
0:40:09	large
0:40:10	set of interface this C plus channel aussie sharper and marcy be protocol uses for
0:40:18	ivr so that is nice open source project the
0:40:21	press instead of face to build the our on how based so set B C's
0:40:26	and so on
0:40:30	the this is common a framework
0:40:32	for
0:40:33	but based
0:40:35	over a solution
0:40:37	we were speech set of our application server the base ever and some clients
0:40:45	this is just example of
0:40:48	our testing client
0:40:51	okay so some not now i will just summarised
0:40:55	three slides about the
0:40:57	some ongoing challenge is that the I C now
0:41:02	partner very important challenge these
0:41:05	data
0:41:07	a training data is the smog a small company it is a difficult for us
0:41:12	to get the data it is expensive
0:41:17	and the this out that the a common approach of allows us to at just
0:41:23	two and we just by here
0:41:25	so we are working the
0:41:28	for cheaper mesa how to do these
0:41:32	i think it'd great inspiration is google
0:41:36	so but not a we did something similar in you know language identification
0:41:42	in which i didn't if we are able to bit the data the that the
0:41:46	we can use of for training go for organisational
0:41:52	a speech recognition systems that can be deployed on balls like don't the you know
0:41:57	quantization of speech for telephone speech and the
0:42:00	for broadcast
0:42:03	so of what one possibility that we export the was to use broadcast
0:42:10	for this
0:42:11	but not so the whole content but the
0:42:15	automatically to take the
0:42:17	phone calls in the broadcast this ensures
0:42:20	a high variability of speaker does the dialects and the was speaking styles
0:42:28	language can be a very fight the using the when automatic language identification will so
0:42:33	we need to some
0:42:34	a small amount of data to bootstrap the this approach but then it is possible
0:42:41	this is speaker identification of the speakers of the variability conventional by current
0:42:47	speaker id technology
0:42:50	the you would like to
0:42:52	transcribe the it is some of the speech we think crowd sourcing
0:42:57	and that use have really unsupervised the training for adaptation
0:43:03	of
0:43:05	currently we discuss the D so we've
0:43:10	several company sent to would like to form a conception for this
0:43:14	you have some expending admins experience to when we did this
0:43:19	you know project the for language identification ldc anthony's the
0:43:24	of it turns out the E to be very successful and the melody so like
0:43:29	mainstream language identification
0:43:31	we have one line
0:43:32	up or go for adaptation
0:43:36	be backed by our
0:43:38	after companies the spinoff from but and that we believe that we could put to
0:43:44	reduce the cost for the opened of new recognizer
0:43:47	to variable and models
0:43:49	so if you are interested in the you would like to one and more just
0:43:53	sent me email and to
0:43:55	we can discuss this
0:43:58	the then other trying to we see is the that the
0:44:03	we have quite the roots the technology about the still the deep one man the
0:44:09	is hardly bring some of somebody six of each customer list
0:44:13	but if the specified
0:44:16	the each if we have departments many cantonese
0:44:19	we never know what to be the final
0:44:23	accuracy of up
0:44:25	you're technology and of unevenly to do adaptation again some
0:44:30	project so that i mention on the previous slide so we have
0:44:35	to word this
0:44:37	usually if you speak of the technologies that
0:44:40	we claim the
0:44:43	the customers that the technology
0:44:46	is
0:44:48	language-independent the
0:44:50	channel independent the but always there is some for two for station
0:44:55	the only possible way i see to reduce this risk
0:44:58	is to built on to evaluate that these technologies
0:45:03	on a many languages and to know that the results in advance before the technology
0:45:08	so
0:45:10	so for this again the data collection project and can have
0:45:15	and to you are thinking about you want
0:45:17	to extends some approaches to something like to work through much of spoken languages
0:45:24	because for language identification we have collection of about fifty languages and
0:45:29	good all the rapidly
0:45:32	and the
0:45:33	finally remark
0:45:36	what we see that is that the percentage is
0:45:41	full cost mainly on accuracy or more most of the research articles
0:45:46	are describing some improvement in accuracy but if we speak about commercial market the
0:45:53	i think that any improvement and the you know speech the
0:45:57	or will
0:45:59	something that cannot
0:46:00	and i do use
0:46:02	the cost of hardware
0:46:04	can help and can
0:46:07	help you to
0:46:10	have successful technology
0:46:13	so we saw you in some approaches the hardware cost is a really large can
0:46:19	be
0:46:20	fifty percent of the cost of the project and so on
0:46:24	so this is everything from me and thank you for attention if you have any
0:46:30	questions please ask
0:46:39	any questions
0:46:46	so we how did you do that didn't have to go to cepstral but better
0:46:51	or something like that
0:46:53	we are considering this approach should to but
0:46:57	you know it at the beginning it's harder to get the
0:47:01	money from adventurers
0:47:02	the so this started the in the trade that event to customers and the ask
0:47:08	or negotiate at some contacts
0:47:12	and the we just started to be for contract so basically the
0:47:17	custom development and the
0:47:19	B and some money on this custom development and then we compute developing technology
0:47:24	and we start something good technology and then
0:47:28	even to product and stuff
0:47:32	i have a question your
0:47:35	your solutions are on site or is it based on cloud services
0:47:43	most of the solution so
0:47:46	i don't say it actually bosses possible be because
0:47:50	that we can use of the technology one site but we have the base the
0:47:55	or interfaces for example that is the best interface that can be used for
0:48:00	called department
0:48:03	but so you have a lot of cloud deployments or not please models and not
0:48:07	know most of our current improvements are
0:48:10	of
0:48:11	a local click the like of one side the departments
0:48:15	but we have
0:48:19	the spinoff at the battalion is it of technology this is to play well
0:48:22	that is for example the recording go lecture here
0:48:25	this is already got based this is gonna but
0:48:30	i don't at the of lectures
0:48:35	questions
0:48:44	so you started off connected with
0:48:47	with university do you still do you have now it's to say that you projects
0:48:51	that at the next cnns are an issue with that in terms of their
0:48:56	we only with the company a with the government
0:49:00	we are doing this see in different races we didn't have students
0:49:05	it's for next cer we some or twelve some people at the but
0:49:11	some contacts we have joint project
0:49:14	the sort out differently so
0:49:21	alright
0:49:22	that's one thank you thank you

Speech technologies for data mining, voice analytics and voice biometry

Applications Day

Petr Schwarz (Phonexia & Brno University of Technology)