Speech Transcript - The Albayzin 2008 Language Recognition Evaluation

0:00:06	i
0:00:07	you start by one percent
0:00:10	this evaluation this was and therefore we made penis paying two years ago
0:00:17	this is the line of the presentation um
0:00:20	i
0:00:21	talk about the compass timeout duration than i would describe the this conditions
0:00:26	which are more much liking the nist evaluation
0:00:30	uh evaluations
0:00:31	we took uh nist evaluations as us
0:00:35	um an example and
0:00:37	i understand that in fact
0:00:39	so we can be uh
0:00:41	maybe uh
0:00:42	and ninety percent off
0:00:44	the condition
0:00:45	then uh i will describe
0:00:47	uh basically
0:00:49	we thought as possible
0:00:51	uh the results and then give some conclusions
0:00:57	well uh
0:00:58	this uh evaluation was uh supported by the spanish thematic network on speech technology to spain
0:01:04	uh it was uh uh but the feed the any of what's up on speech than only one was
0:01:09	that can be bought in november
0:01:12	with us tonight
0:01:14	and in that but in that uh were so um
0:01:17	the what the other two other operations on the speech trouble relation
0:01:22	and screens
0:01:23	speech synthesis
0:01:24	but wrong
0:01:25	the language recognition evaluation
0:01:27	i don't you know
0:01:29	and the what the was another motivation which was that uh or group was interested in developing
0:01:36	language recognition did not you for uh
0:01:39	spokane document retrieval applications
0:01:43	well what we see on all the points we have
0:01:47	we had in mind
0:01:48	when you sign in the evaluation
0:01:50	uh
0:01:51	well to promote collaboration between research group
0:01:54	in this pain also portable
0:01:57	uh secondly uh
0:01:59	to provide
0:02:00	i speech database
0:02:01	a specifically designed
0:02:03	two
0:02:04	uh perform and recognition in the language in spain
0:02:08	therefore languages in spain
0:02:10	not everybody knows that
0:02:12	yeah that
0:02:13	four official languages spoken in spain
0:02:16	and
0:02:18	then a another motivation was uh to ms word accuracy
0:02:22	that the state of the art systems
0:02:24	good a time for for this particular application
0:02:28	because these languages uh
0:02:30	yeah
0:02:31	how people in jointly in spain
0:02:35	so uh maybe this task
0:02:37	could be more challenging than
0:02:39	we could expect
0:02:41	and finally uh mister
0:02:44	this was a diffuse your uh motivation
0:02:47	maybe for some you
0:02:49	to mister the performance of systems developed
0:02:52	on a limited
0:02:52	a month
0:02:53	data
0:02:56	well uh the language
0:02:58	detection task was defined
0:02:59	same way of as for nist
0:03:01	i don't describe this is
0:03:03	the same
0:03:04	or been described
0:03:06	they were
0:03:08	uh
0:03:11	yeah that's
0:03:12	and this can be assumes
0:03:13	simple
0:03:14	uh
0:03:15	uh
0:03:16	what is the
0:03:17	described here this this is like
0:03:19	uh the current system development which is
0:03:22	a special one
0:03:23	and we need differentiate
0:03:26	yeah between uh
0:03:28	systems developed uh in three conditions
0:03:31	using any available materials
0:03:34	and systems uh developed
0:03:36	using only
0:03:38	did the date that we provide
0:03:40	okay
0:03:40	that was very special for this evaluation
0:03:43	we were interested in putting all the teams
0:03:46	at the same point
0:03:47	to develop their systems
0:03:49	and then to evaluate
0:03:50	what they could
0:03:51	do
0:03:52	starting from the
0:03:53	okay
0:03:54	well
0:03:55	regarding the set of trials we define it as for nist the closest set
0:03:59	uh this to one nation open set this one ratio
0:04:03	uh we also be fine
0:04:05	fig kind of segments
0:04:06	of uh for a second and second and three seconds segments
0:04:11	um we uh defined we used
0:04:13	the same performance measures
0:04:15	uh also uh defined by nice
0:04:19	you have
0:04:20	uh scene
0:04:22	this must or sin
0:04:23	the previous presentation
0:04:26	average calls
0:04:27	we also use
0:04:28	the seattle area
0:04:30	and finally that the course to
0:04:32	give uh qualitative
0:04:34	uh evaluation
0:04:35	systems
0:04:36	we uh we define the same priors and colours
0:04:40	of the last
0:04:41	to understand what we
0:04:45	well then database way to record it
0:04:49	it was found that call
0:04:50	okay
0:04:51	i recorded the database
0:04:53	from T V
0:04:53	in my home
0:04:56	uh just connecting idea to record the to the
0:04:59	to the decoder
0:05:00	the
0:05:01	couple T V reporter
0:05:03	and is described in the paper that a right
0:05:07	two thousand two
0:05:09	so it it it um important for target languages
0:05:12	spanish out on a second at least yeah
0:05:16	and also all the languages
0:05:17	just to i love open set test
0:05:20	uh the languages
0:05:22	uh
0:05:23	where friends portable used your money and english
0:05:26	five or two case is not so close to spanish that you can
0:05:30	uh
0:05:31	fig
0:05:31	so
0:05:33	for you want to these people uh
0:05:35	find many different
0:05:37	um the spanish too
0:05:38	it should
0:05:39	with the language
0:05:40	well uh audio files uh where uh what files yeah
0:05:44	um
0:05:45	sixteen Q don't hurt
0:05:46	uh
0:05:47	uh the last frequencies
0:05:49	sampling frequency
0:05:50	yeah well they were single channel
0:05:52	fig ten bits per sample compressed
0:05:55	P C M
0:05:56	uh this is another dot
0:05:58	friends with about the nist evaluation
0:06:01	speech signals would all start from T V souls including
0:06:05	a lot of speech or spontaneous
0:06:06	speech
0:06:07	what kind of environment conditions
0:06:09	uh yeah
0:06:11	for instance that could be
0:06:12	three speakers as speaking enough to be
0:06:15	second segment
0:06:17	so
0:06:17	that could be many speakers
0:06:19	speaking in the same
0:06:20	yeah
0:06:21	test set
0:06:23	well we define
0:06:24	he joins subsets of T V shows
0:06:27	to train development and evaluation
0:06:29	this was to make to i a guaranteed at different
0:06:33	more or less
0:06:34	that different speakers
0:06:36	uh where in each
0:06:38	in it
0:06:38	subset
0:06:40	and finally the only data bases pretty
0:06:42	small
0:06:43	four
0:06:44	then use a standard
0:06:45	it's just a fifty fifty hours long
0:06:49	is distributed to
0:06:50	C D V D but
0:06:52	and um
0:06:53	we are just now uh
0:06:56	talking with
0:06:56	the L D C to the distributed
0:06:59	two
0:06:59	L D C
0:07:00	and the train data set in
0:07:03	good
0:07:03	yeah
0:07:04	last
0:07:05	and fifty six hours
0:07:07	nine hours per target language
0:07:09	we don't provide any uh they got to uh
0:07:13	train
0:07:14	or something that
0:07:15	four
0:07:16	oh the seven languages
0:07:17	so just
0:07:18	nine hours
0:07:19	but that the language
0:07:20	that's that's all
0:07:21	and i audits of languages
0:07:23	i'll be uh
0:07:24	in the development dataset and in the evaluation that
0:07:28	which are more or less the we
0:07:30	have more or less the same structure
0:07:32	but
0:07:33	i don't
0:07:34	i would then
0:07:36	uh well when defining when deciding that about the database
0:07:40	uh we only choose uh
0:07:42	tools um
0:07:44	high snr as speech
0:07:46	described in sediments with
0:07:47	right
0:07:48	lemma noise
0:07:49	uh speech overlaps
0:07:51	they or what all of that all of them
0:07:54	well fit the foul
0:07:55	and the guidance documents for training they'll have to know then
0:08:00	restrictions
0:08:02	maybe five minutes
0:08:03	you could train with a five minute segment
0:08:05	with
0:08:07	so um but for seven ms for the betterment of automation
0:08:11	yeah
0:08:12	to cut
0:08:13	uh
0:08:13	lend restrictions
0:08:15	um
0:08:16	we are defined automatic
0:08:18	a way of constructing them
0:08:21	and by ensuring that they would enclosed by silence
0:08:24	more or less
0:08:26	yeah and in fact the subsets
0:08:29	well the subset of
0:08:30	three second segment
0:08:31	is
0:08:32	extracted from the subset
0:08:34	subset
0:08:35	of
0:08:35	then support six seven months
0:08:37	and the same way that and second segment
0:08:40	subset is extracted from the
0:08:42	the the second
0:08:44	segments option
0:08:45	uh quite difficult but
0:08:47	what
0:08:47	we tried is to ensure that differences in
0:08:51	in performance
0:08:53	uh would you only to uh the the land
0:08:56	not too
0:08:57	um
0:08:58	being testing against different material
0:09:02	and
0:09:03	the where sound tolerance in land
0:09:06	we use in fact
0:09:08	a segments between
0:09:09	three and five seconds
0:09:10	ten and twelve seconds and
0:09:12	fig the active duty cycle
0:09:14	where
0:09:15	the door
0:09:16	uh interval
0:09:18	and finally that they don and that's it
0:09:20	and the same for evaluation
0:09:22	but uh
0:09:23	one though send a candidate segment
0:09:26	yeah
0:09:27	sue me not i think for the three ratios so the where
0:09:31	six hundred seven is but duration
0:09:33	and for each iteration there were a one hundred twenty segments but by the language
0:09:38	and
0:09:39	i know that
0:09:40	one hundred twenty seven minutes of all the syllable
0:09:43	it's uh
0:09:45	i have to say that
0:09:46	this uh
0:09:48	it means that
0:09:49	yeah
0:09:50	the where exactly it
0:09:52	oh well too
0:09:53	uh yeah
0:09:56	twenty percent
0:09:57	of um
0:09:59	uh segments where i'll go
0:10:00	from out of seven
0:10:02	as
0:10:02	in the in both the development and evaluation purposes
0:10:06	which might
0:10:07	exactly
0:10:08	was what
0:10:09	what was defined
0:10:10	in the
0:10:11	in the right
0:10:13	thing
0:10:15	um
0:10:17	well
0:10:17	everybody database design
0:10:19	the proportions of known languages
0:10:21	where mate
0:10:22	they rely
0:10:23	but difficult for me to promote it
0:10:25	a different for development evaluation
0:10:28	and to avoid uh
0:10:30	tandem systems to reject specifically
0:10:33	so uh
0:10:34	kerry part of the table of uh
0:10:36	the distribution of segments
0:10:38	for development and evaluation
0:10:40	you can see that
0:10:41	there were
0:10:42	seventy sevens for friends
0:10:43	them for portuguese and
0:10:45	forty four english
0:10:46	and not for from the element in the development set
0:10:50	and evaluation set
0:10:52	the drawings were change
0:10:54	between for example to be sent english and german
0:10:57	so
0:10:58	was
0:10:59	may
0:11:00	this way
0:11:02	uh
0:11:03	evaluation do simply
0:11:05	there were there was on a rotation plan very similar to
0:11:08	that companies
0:11:10	uh the wherefore class conditions open set free open suppressed it consists of three judge that
0:11:16	restrict it
0:11:17	and three durations of the web
0:11:19	to attract
0:11:20	five
0:11:21	for it
0:11:22	this condition on uh it's fifteen percent just one single primary system
0:11:27	and any number of compressed before alternative systems
0:11:31	they wanted to pursue
0:11:32	ah the solution should be
0:11:34	so my submitted by by teams
0:11:37	in this uh evaluations format
0:11:40	at this file with one hundred trials section
0:11:43	fig spline
0:11:45	um but this depends what am i committed to specifically specified
0:11:49	whether or not the scores may be interpreted
0:11:52	us
0:11:52	look like
0:11:53	oh look
0:11:53	like that the errors
0:11:55	or not
0:11:56	and also to send this presents and to participate um
0:12:01	in the ann arbour scene with us tonight
0:12:03	and then with evolution evolution works
0:12:06	okay
0:12:07	systems where one
0:12:08	uh in fact according to their average goals
0:12:12	and are defined that way
0:12:14	in this
0:12:15	fancy
0:12:16	and
0:12:17	though
0:12:18	was run it
0:12:19	and right now i'm not work for the best system
0:12:23	the system the only you'll be in
0:12:25	the least
0:12:25	the average goes in there
0:12:27	see a
0:12:28	thirty condition
0:12:30	close to restrict it
0:12:31	on
0:12:32	other subset of
0:12:34	uh
0:12:35	still be second seven
0:12:38	well
0:12:38	this was there
0:12:40	the scale of their valuation
0:12:42	uh
0:12:42	in
0:12:44	few words
0:12:45	the work
0:12:46	three months for developing your system
0:12:48	and there were three weeks to uh
0:12:50	uh process they want vision of
0:12:53	and
0:12:54	i have to say that uh the database produced the database was produced
0:12:58	in
0:12:59	three models
0:12:59	from april to june
0:13:01	depends on a
0:13:02	and we also recorded some more um
0:13:06	data in september
0:13:08	to find something to
0:13:08	two
0:13:09	uh
0:13:11	and uh complete
0:13:12	the evaluation on the test
0:13:14	okay
0:13:16	that that you can find it in the paper
0:13:19	well
0:13:20	uh
0:13:21	i now i begin to describe herself
0:13:24	yeah
0:13:25	the work for participants
0:13:27	displayed in teams
0:13:29	percent including systems
0:13:31	things were from spain and what about
0:13:34	and uh there were two teams percent in a state of the art systems more or less
0:13:39	and the two first the first ones
0:13:41	T one T two
0:13:43	and the other two percent it assistance not
0:13:46	specifically designed for uh
0:13:48	a language recognition applications so
0:13:51	the the source world
0:13:52	just the table of uh the average cost
0:13:55	four
0:13:56	uh thirty second segment
0:13:58	you can
0:13:59	see that
0:14:00	there
0:14:00	performance as well
0:14:02	very bad
0:14:04	so uh in the following that it will only
0:14:08	because
0:14:08	talk about
0:14:09	results of these two two
0:14:11	to to
0:14:13	okay
0:14:14	well
0:14:15	yeah no i somersaults uh first the
0:14:18	condition i
0:14:20	talk about
0:14:21	is
0:14:21	the the mandatory one
0:14:23	for which they almost all the teams have to
0:14:26	the centre system
0:14:28	you can see here
0:14:29	cool
0:14:30	that's good
0:14:31	um
0:14:33	this uh
0:14:34	like what this
0:14:35	one is uh
0:14:37	for a contrastive system funding from T one with
0:14:40	in fact
0:14:41	uh got the best result
0:14:43	the best the average goes
0:14:45	but the best primary system was also
0:14:47	from T one
0:14:49	uh
0:14:50	they have the
0:14:51	then um
0:14:54	okay
0:14:55	yeah
0:14:56	when
0:14:57	when channel
0:14:57	this was i
0:14:58	okay
0:14:59	to say that this was this was in restrictive conditions
0:15:02	these systems to come see a
0:15:04	big difference with
0:15:06	T seem to on T V
0:15:08	team one
0:15:09	uh because uh
0:15:11	they were already uh develop their systems
0:15:14	using
0:15:15	this to the data provided in this one which
0:15:18	not using
0:15:19	any other sisters
0:15:20	they rely on all the data and all the like
0:15:23	okay
0:15:24	so when changing to the three conditions
0:15:28	uh
0:15:29	with
0:15:29	see the systems uh got
0:15:32	uh
0:15:32	much better
0:15:33	performance
0:15:34	around
0:15:35	five percent equal error rate
0:15:37	but the
0:15:39	in fact we were surprised
0:15:41	by this result because we expect it
0:15:43	much better results
0:15:45	around one percent or less
0:15:47	yeah
0:15:49	and uh we uh
0:15:51	made some experiments afterwards
0:15:53	the the the one mission which
0:15:55	on system
0:15:57	a system that got on there
0:16:00	fig
0:16:01	he percent or whatever right
0:16:02	in the general language recognition task defined in use
0:16:05	two thousand seven
0:16:07	evaluation
0:16:08	and we've got
0:16:09	five
0:16:10	yeah
0:16:11	forty five percent whatever right
0:16:13	so
0:16:14	five
0:16:15	it seems that uh this task
0:16:17	mm
0:16:18	the task defined for
0:16:19	for about seeing this evaluation
0:16:22	is
0:16:23	uh more difficult than
0:16:24	i'm spec
0:16:25	okay
0:16:27	there are some possible issues not the same that's another thing that data results comparable comparable
0:16:33	between the knees
0:16:34	evaluation on this evaluation
0:16:36	maybe not
0:16:38	the statistical significance
0:16:40	there are not many uh trials
0:16:42	only
0:16:43	yeah six hundred there
0:16:44	uh
0:16:46	but the nation
0:16:47	okay
0:16:48	and there are also some possible explanations maybe the acoustic variability a speaker's channel
0:16:53	background noise
0:16:55	there were different conditions
0:16:56	and also their phonetic and lexical we
0:16:59	but for these
0:17:01	uh
0:17:01	the phonetic on lexical similarity among body language
0:17:05	or more than one
0:17:07	the same country
0:17:09	oh no
0:17:10	many years many centuries
0:17:11	what we
0:17:12	don't leave so
0:17:14	maybe this is the
0:17:15	then race
0:17:16	in any case
0:17:18	size have said that that seems uh challenging enough for that
0:17:22	a lot of other research
0:17:24	in
0:17:25	language recognition technology
0:17:27	well
0:17:27	yeah this is
0:17:29	we have been talking about their clothes
0:17:32	set
0:17:32	condition now i'm talking about
0:17:35	that opens the condition
0:17:36	the best performance in this case was
0:17:38	worse
0:17:39	like for
0:17:41	yeah because there are
0:17:43	uh well known languages
0:17:44	in there
0:17:44	the trials
0:17:46	and with the systems that the system works around nine percent accurate
0:17:51	this case
0:17:52	which is almost
0:17:53	two times they were raiding the
0:17:55	close to completion
0:17:57	so that
0:17:57	three conditions
0:17:59	yeah well
0:18:00	or conclusion is that some unknown languages are being confused with body language is maybe or to be some friends
0:18:06	we don't know
0:18:09	well
0:18:09	yeah you have uh
0:18:11	there was these results
0:18:12	uh the second rate for languages for target languages
0:18:15	uh uh for the best system
0:18:18	so you can hear you can see for the close
0:18:21	set condition i'm for that opens the condition
0:18:24	and
0:18:25	the green who is for bus
0:18:28	which got the best
0:18:29	uh performance
0:18:31	and then uh
0:18:32	right put his fork at a time
0:18:34	we've got a
0:18:35	worst performance
0:18:36	in opens the condition
0:18:38	and you can see the uh that bus
0:18:42	the change in the performance
0:18:43	for bass
0:18:44	really
0:18:45	it's more
0:18:47	also forced by means which is the
0:18:50	i think
0:18:51	but this is the kid a blue
0:18:53	and the power point it's not easy and which also
0:18:56	uh wasn't
0:18:57	yeah
0:18:57	it's performance but not as
0:18:59	much as a forecast for qatar
0:19:02	so are we have uh analyse this in more time
0:19:06	with
0:19:06	this table
0:19:07	i have to say that
0:19:09	there is are a right uh never in the paper
0:19:12	uh these numbers are
0:19:13	five
0:19:14	the uh error rates
0:19:16	uh
0:19:17	you need some before somehow
0:19:19	we missing there
0:19:21	in the
0:19:22	dialogue now and
0:19:23	be false alarm aside diagonal
0:19:25	um
0:19:27	uh we mistake them as coast
0:19:29	so yes this
0:19:30	they they did but they soon as the same but
0:19:32	the
0:19:34	the numbers are not
0:19:35	what the paper says
0:19:37	they are
0:19:37	okay
0:19:38	ah as you can
0:19:40	there is a reliable recall here
0:19:43	the white meaning
0:19:44	zero there are and black meaning
0:19:47	one
0:19:48	the maximum possible error
0:19:51	uh yeah
0:19:51	so this is for the close and so condition
0:19:54	and when changing to the open set
0:19:57	conceive here
0:19:59	really
0:20:00	usually that for at a time
0:20:01	no languages
0:20:03	tyler
0:20:04	you find that
0:20:05	um
0:20:06	many uh
0:20:07	trials
0:20:08	corresponding to known languages where
0:20:10	confused with qatar
0:20:12	that's
0:20:12	the origin of
0:20:14	that uh changing the core
0:20:17	for the open set condition
0:20:19	okay
0:20:20	uh
0:20:21	this assault and not going to comment
0:20:23	this because it's the same for us for me is
0:20:26	the the performance
0:20:27	uh watson's us
0:20:29	the
0:20:30	double of the land
0:20:31	uh is
0:20:32	less
0:20:34	of the around
0:20:35	the segment
0:20:37	and uh
0:20:38	this is for
0:20:39	more interesting for me
0:20:41	is
0:20:41	because uh
0:20:43	you can see
0:20:44	what happens when you restrict
0:20:46	they get the the bottom we conditions
0:20:49	uh you have here um
0:20:51	yeah two different teams
0:20:53	the blue ones
0:20:54	being one
0:20:55	the right one is team too
0:20:57	and
0:20:58	for the uh three
0:21:00	condition
0:21:01	well then uh got more or less the same
0:21:04	performance
0:21:06	this
0:21:06	to go
0:21:08	but for the restricted condition when
0:21:10	restricting the materials
0:21:12	they could use
0:21:13	to double their systems
0:21:16	what uh the T one
0:21:18	okay
0:21:19	the the performance
0:21:21	quite close to the
0:21:22	to the other condition where is the the other one
0:21:25	uh
0:21:26	the performance was much
0:21:28	much worse
0:21:29	the difference
0:21:30	it's
0:21:32	uh
0:21:33	forty percent word
0:21:34	a or its goals
0:21:35	to uh
0:21:36	four hundred percent
0:21:38	what are its goals
0:21:40	these
0:21:41	okay
0:21:41	so
0:21:42	i think this is important because
0:21:44	you can
0:21:45	i for me
0:21:46	this is spent is much more robust
0:21:48	now the other one
0:21:50	because it
0:21:51	not
0:21:52	does
0:21:52	does not depend
0:21:54	on
0:21:55	so much on the materials provided to to trying to to to train it
0:21:59	okay
0:22:00	well
0:22:01	conclusions
0:22:04	um
0:22:05	well
0:22:06	we i thought sent it uh that what was an evaluation involving
0:22:10	the official language in spain
0:22:12	ask around a listener spanish
0:22:14	you seen uh material was a recording from till you drop
0:22:19	davis
0:22:20	since then uh playing
0:22:22	state of technology got around five percent equal error rate
0:22:25	in the close set
0:22:26	three development condition
0:22:28	just
0:22:28	what's that
0:22:29	the rest for for them
0:22:32	and
0:22:33	we think
0:22:34	that
0:22:34	fine task
0:22:35	tasks
0:22:36	in this uh evaluation
0:22:38	my support for the developments in language information technology
0:22:42	and uh will we form
0:22:44	yeah darkness is its sensitivity to the bottom restrictions
0:22:48	depending depending
0:22:50	on the system
0:22:52	uh
0:22:53	from two different systems
0:22:54	that
0:22:56	uh
0:22:57	fig creasing calls
0:22:58	what's different
0:23:00	for them
0:23:01	my thing to be
0:23:03	this
0:23:03	uh
0:23:05	condition i don't know if
0:23:06	you are interested in restricting
0:23:08	the materials
0:23:10	but i think
0:23:11	it could be interesting for me assimilation
0:23:14	maybe i don't know
0:23:15	you are interested but
0:23:17	um
0:23:19	on finally we found not the same performance um opening languages
0:23:23	the best performance was formed for bass
0:23:25	and the
0:23:26	was performance was from for a time
0:23:29	speculating about these
0:23:30	we can uh say that bus
0:23:33	is uh a special language
0:23:35	not romance languages
0:23:37	uh data
0:23:38	its origins are different
0:23:40	oh
0:23:40	all the languages in spain
0:23:43	and contamination roma's language
0:23:45	which you may be
0:23:46	usually confused
0:23:48	people by the systems
0:23:50	with portuguese or
0:23:51	maybe friends
0:23:53	or maybe
0:23:54	spain uses pennies or at least
0:23:57	well
0:23:58	and
0:23:59	finally i have to say or couldn't work
0:24:01	is organising
0:24:03	in this in this evaluation
0:24:04	we are not just now
0:24:06	or anything
0:24:07	this evaluation that was seen two percent and language recognition evaluation
0:24:12	yeah yeah
0:24:12	we have a record it i knew we have extended
0:24:15	the
0:24:16	how like a database
0:24:17	which was
0:24:18	one used
0:24:19	before
0:24:20	to uh define come back to
0:24:23	we have i did portuguese and english study languages
0:24:27	maybe you are interested
0:24:28	these languages
0:24:29	a happy new they're set of unknown languages
0:24:32	um have included
0:24:34	i knew this condition for noisy speech
0:24:39	is this getting we
0:24:40	you can't of easter
0:24:42	yeah until july
0:24:43	fifteen
0:24:45	i'm september you have
0:24:47	more or less three months
0:24:49	if you use them now
0:24:52	until september twenty seven
0:24:54	to uh the video systems
0:24:57	and two weeks
0:24:58	to uh process
0:25:00	uh evaluation that
0:25:02	then and the key file and results were released
0:25:05	one double
0:25:07	fifteen
0:25:08	and the warsaw
0:25:10	yeah
0:25:11	language recognition what is and what's not
0:25:13	we we had
0:25:14	in november i beagle spain
0:25:18	in a contest of how to do something that is uh
0:25:22	what's up uh
0:25:23	spain
0:25:25	okay
0:25:26	uh you can repeat step in this
0:25:28	uh
0:25:30	well
0:25:31	and if you look is that please
0:25:33	dissipate
0:25:35	that's all
0:25:43	they should
0:25:48	you mentioned at the beginning
0:25:50	when you collect the database
0:25:52	you might sure
0:25:53	uh
0:25:54	now treat each speaker
0:25:56	no no i'm not sure
0:25:57	i
0:25:58	try to uh
0:25:59	distribute programs
0:26:01	in different sets for instance one brother one T V so
0:26:04	was only for evaluation
0:26:06	another T V show was only for development and
0:26:10	you
0:26:11	for instance yeah
0:26:12	this I T V so called colour
0:26:15	in bass
0:26:16	this T V show was
0:26:17	i don't think it all for training
0:26:20	but not for development not
0:26:22	i'm not sure
0:26:23	that if they are not
0:26:24	the same speakers
0:26:26	i like
0:26:27	i i
0:26:27	we tried
0:26:28	two
0:26:29	to um
0:26:31	to manage
0:26:31	the
0:26:33	just understood
0:26:33	oh
0:26:34	i'm just speculating that
0:26:36	uh
0:26:37	the speakers are also like
0:26:39	what
0:26:39	well developed
0:26:40	see
0:26:40	oh
0:26:41	if
0:26:42	there's a lot
0:26:44	uh_huh
0:26:45	repeated speaker in
0:26:47	elements
0:26:48	no
0:26:49	uh
0:26:50	you could lead to recognise
0:26:51	speaker
0:26:51	right
0:26:52	lang
0:26:53	no because uh we try to put on so and
0:26:57	and
0:26:58	many speakers in in
0:27:00	that is not problems like um
0:27:03	um
0:27:04	broadcast news
0:27:05	where there is only one or two speakers speaking all the time
0:27:08	or more
0:27:09	much yeah
0:27:10	much
0:27:11	time
0:27:12	we try to select various
0:27:13	T V so
0:27:14	different T V shows
0:27:16	uh sincerely debates
0:27:18	and talk shows where many people speak
0:27:21	and they're also interviews so
0:27:24	so
0:27:25	maybe is to what you're telling
0:27:27	you are suggesting
0:27:29	but
0:27:30	i don't think so we
0:27:32	i don't know
0:27:39	question uh
0:27:40	with the growing to do the data format so you recorded to wideband speech
0:27:45	and uh
0:27:46	you also find that so uh
0:27:48	it was a little harder task
0:27:50	yeah
0:27:51	expect
0:27:52	speech
0:27:52	in this
0:27:54	nation
0:27:55	no
0:27:56	so you would have more information like
0:27:59	speech
0:28:00	but
0:28:01	um
0:28:02	it might also that might be the fact
0:28:04	right
0:28:05	we stick unique
0:28:06	what effect it might be the fact that
0:28:08	most people have been developed for telephone
0:28:11	oh
0:28:12	speech
0:28:13	what would you be in there
0:28:16	well
0:28:16	i i not happy with that obviously
0:28:18	because we are developing technology for
0:28:21	for T V
0:28:22	for for T V signals uh record it
0:28:25	in one by one
0:28:26	but by conditions
0:28:28	so uh i i understand the reasons to organise
0:28:32	then use the one races because
0:28:34	was the sponsor or what
0:28:35	the sponsor ones
0:28:36	two
0:28:38	by finance in the the
0:28:40	they're all nations
0:28:41	but
0:28:41	from the point of view of the of the research uh community i think
0:28:45	we should uh
0:28:48	try to
0:28:49	to organise how to kind of whatever whatever it seems more devoted to technology got it
0:28:54	the bottom ends
0:28:55	unless to the application i
0:28:57	it's my opinion but
0:28:59	uh
0:29:00	we have to decide
0:29:01	we had to maybe to find
0:29:03	sponsors and
0:29:04	i don't know
0:29:06	if
0:29:06	uh
0:29:06	that's
0:29:08	possible or not
0:29:16	understood
0:29:17	and
0:29:17	i just
0:29:18	session because the men should discussion going
0:29:22	so
0:29:22	thank you
0:29:23	oh okay

The Albayzin 2008 Language Recognition Evaluation

SESSION 7: Speaker and Language recognition - Evaluations and performance testing

Added: 14. 7. 2010 11:08, Author: Luis Javier Rodriguez-Fuentes, Mikel Penagarikano, German Bordel, Amparo Varona (University of the Basque Country), Length: 0:29:29