Přepis řeči - Improving Language Recognition with Multilingual Phone Recognition and Speaker Adaptation Transforms

0:00:06	so this is
0:00:07	um
0:00:08	this is work
0:00:09	that we did at S R I
0:00:11	and it's in fact our
0:00:12	foray into language recognition
0:00:15	um
0:00:17	and
0:00:17	it's great because most of
0:00:19	the
0:00:19	the kind of background techniques were already
0:00:22	uh
0:00:22	splendid plenty of detail in the previous
0:00:24	three talks
0:00:25	so
0:00:26	two
0:00:26	go into that again
0:00:27	um
0:00:29	i will
0:00:29	start with some preliminaries
0:00:31	and then
0:00:32	uh tell you about our
0:00:33	or mental set up
0:00:35	and the focus here will be on phonotactic
0:00:37	uh
0:00:38	why should not say phonotactic but phone base
0:00:40	uh recognition
0:00:42	uh i say that because that encompass
0:00:44	or both
0:00:45	phonotactic
0:00:46	uh modelling
0:00:47	uh such as what we've heard about
0:00:48	before as well as
0:00:49	uh mllr
0:00:50	based modelling which of course is also based on phone models hence the commonalities
0:00:55	um
0:00:56	and
0:00:57	i will also look at two different techniques and comparing uh two different ways of doing the phonotactic modelling
0:01:03	um
0:01:04	then conclude with some
0:01:05	pointers to future work and conclusion
0:01:08	so
0:01:09	because we haven't participated in the past
0:01:12	uh language recognition evaluations we actually did have access
0:01:15	two
0:01:16	any data after lre all five
0:01:19	since this work was done L the sea's accuracy release L O seven but we didn't have a chance
0:01:24	process that yeah
0:01:25	so we're dealing with this
0:01:27	i apologise for dealing with a rather
0:01:29	outdated
0:01:29	a task which is the a seven language task
0:01:32	mallory of five
0:01:33	it's a conversational speech sounds a lot of that a voice of america stuff
0:01:38	um the test data consists of
0:01:40	uh about thirty six hundred
0:01:42	test segments
0:01:42	and we only look at
0:01:43	thirty second condition
0:01:45	uh the training data and is
0:01:47	from uh you know the same
0:01:49	seven languages
0:01:50	and the duration
0:01:51	after we perform our automatic segmentation into speech nonspeech
0:01:55	uh boils down to about fifty six hours for these three
0:01:59	so you first languages and that way six hours for the
0:02:02	many falling
0:02:03	that you see here
0:02:05	uh well i'll be reporting the equal error rate averaged over all languages
0:02:09	uh so uh
0:02:11	uh that's just what
0:02:12	how to choose and maybe that
0:02:13	that the best choice but
0:02:14	but that
0:02:15	what you see
0:02:16	um the we performed two fold cross validation so
0:02:21	uh to do calibration and fusion we
0:02:23	uh you know which put the data and how we we
0:02:26	uh estimate on the first have to
0:02:28	second vice versa
0:02:29	combined
0:02:30	results
0:02:31	and this is again because we didn't have
0:02:33	an independent you
0:02:34	that
0:02:35	only working
0:02:36	salary of five
0:02:39	okay
0:02:39	you heard all this
0:02:41	so the the the two main stream
0:02:43	uh
0:02:44	techniques are really the cepstral gmm
0:02:46	um
0:02:47	and lately
0:02:48	you know there's been the incorporation of
0:02:50	uh
0:02:51	uh session
0:02:52	oh variability compensation
0:02:54	with jfa and we implemented that
0:02:56	uh
0:02:57	using you know
0:02:59	kind of a standard framework and just for reference
0:03:01	gives you
0:03:02	something like two point eight seven
0:03:04	uh
0:03:04	the average equal error rate
0:03:06	um
0:03:07	Q data
0:03:08	and uh the alternative
0:03:10	a popular technique of course
0:03:11	the prlm technique
0:03:13	and you've heard all about that already
0:03:15	um
0:03:16	yeah so i won't repeat it here
0:03:18	the uh of course
0:03:20	what
0:03:20	popular about it is that you can then combine multiple language specific like
0:03:24	fusion
0:03:25	at the score level and get much
0:03:27	results
0:03:27	to solve for
0:03:29	and for calibration
0:03:30	fusion
0:03:30	we also didn't attempt anything
0:03:32	uh out of the ordinary
0:03:33	uh in fact we haven't even
0:03:35	i tried to um
0:03:36	oh i haven't
0:03:37	you know
0:03:38	incorporated this uh the gaussian
0:03:40	uh
0:03:40	i can't modelling yet so we just use the multiclass vocal
0:03:44	uh which is based on a lot cleaner
0:03:46	aggression
0:03:47	um
0:03:47	did you both user
0:03:48	and
0:03:49	um
0:03:50	calibration
0:03:51	so
0:03:52	first section
0:03:53	is about
0:03:53	phonotactic language modelling
0:03:56	um this again is a
0:03:58	standard technique by now saw before so
0:04:00	stead of doing one best phone decoding we do we do for that
0:04:03	the coding
0:04:04	um
0:04:06	uh this was actually
0:04:08	you know adopted twice
0:04:10	uh lindsay
0:04:11	uh
0:04:12	proposed
0:04:12	for language I D and
0:04:14	uh as right
0:04:15	C proposes
0:04:15	for
0:04:16	speaker I D
0:04:17	um
0:04:18	and in both cases that shows pretty dramatic improvements and actually i wanted to
0:04:21	respond to uh
0:04:22	to uh something that uh uh haptic
0:04:25	said
0:04:26	um
0:04:27	uh
0:04:27	because
0:04:28	quest
0:04:29	uh or because of the previous talk
0:04:31	i actually do not think that the lattice
0:04:33	the coding
0:04:34	in that increases your variability and in fact i think it reduces your credibility
0:04:38	because you're not making hard decision
0:04:40	so whereas in one hypothesis
0:04:42	you know the recogniser
0:04:43	so the whatever charlie might decide between
0:04:45	having a frequency of one
0:04:47	or you know point nine nine nine
0:04:49	in the
0:04:50	in the lattice
0:04:51	approach you actually
0:04:52	uh represent both the one best and
0:04:54	and later hypotheses so you have
0:04:57	you have all the hypotheses represented
0:04:59	and they just differ
0:05:00	by small numerical values i think it actually gives you a more robust
0:05:04	uh
0:05:05	feature
0:05:06	and that was actually demonstrated
0:05:07	by andy had in the original paper
0:05:10	post
0:05:11	yeah
0:05:12	and the other reason why it works but of course gives you more
0:05:15	granularity
0:05:16	feature
0:05:18	uh
0:05:20	so much about the
0:05:21	how we do the feature extraction
0:05:23	um we
0:05:24	we didn't have time to really develop new phone recognisers for this so we just took
0:05:29	three phone recognition systems that we had lying around
0:05:32	uh one is for american english one is for spanish and thirty four levels you know
0:05:37	and you can see here the phone sets differ in their sizes
0:05:39	and uh
0:05:40	furthermore the training data
0:05:42	is vastly different
0:05:43	uh english we have you know basically as much data
0:05:46	one
0:05:47	um and for that reason we gender dependent modelling for the other two languages with less training data we we
0:05:52	do a gender independent modelling
0:05:54	but other than that they all use
0:05:55	same kind of standard asr
0:05:57	right
0:05:58	uh
0:05:58	plp front and vocal tract length normalisation
0:06:01	hlda
0:06:02	uh for dimensionality reduction
0:06:04	and the crossword triphones intently
0:06:07	uh acoustic model training
0:06:10	uh the decoding of course is done without one attack
0:06:13	constraints
0:06:14	um
0:06:14	but we do use
0:06:16	following the results from lindsay we use the context dependent
0:06:19	triphones
0:06:19	the code
0:06:21	uh
0:06:23	and also go again following uh you know very nice
0:06:26	figure from them as a couple years ago
0:06:28	uh we use the cmllr adaptation and decoding
0:06:31	um and by the way we tried a regular mllr as well and it didn't perform as well
0:06:36	you know and i guess that's an agreement with
0:06:38	another of the previous
0:06:39	talk
0:06:41	okay
0:06:43	so uh the first
0:06:44	uh
0:06:45	i think we would like to propose
0:06:47	is to get rid of or largely get rid of all these different kernel phone decoder
0:06:51	and instead we
0:06:53	we can define
0:06:54	a universal
0:06:55	phone set that covers several languages in our case we made up such a set
0:07:00	uh
0:07:01	fifty two phones
0:07:02	and what you do is you map
0:07:04	a new map your your individual language specific dictionaries to a
0:07:08	uh common shared phone set
0:07:10	and then you retrain
0:07:11	you're acoustic models
0:07:13	uh using the map ref
0:07:15	uh and of course the language models
0:07:17	uh if you perform
0:07:18	uh for decoding with the language model with the phonotactic models should also be retrained
0:07:23	the phone recognition accuracies uh as measured
0:07:27	on um
0:07:28	on individual languages are very close
0:07:30	two
0:07:31	what you get with universal phone set you you're not really you uh
0:07:34	sacrificing much in terms of that
0:07:36	see
0:07:37	and this is the these are the following these are the
0:07:39	language specific steps that we combine
0:07:42	map in this fashion so
0:07:43	we took a
0:07:44	um american english data
0:07:46	of two
0:07:47	the right is mainly native and nonnative speakers
0:07:50	and we
0:07:50	because we know that in much of what
0:07:53	we do both native and
0:07:54	the nonnative speakers of
0:07:55	her
0:07:56	uh
0:07:56	not in the natural frequencies but with
0:07:58	with
0:07:59	more than native nonnative uh
0:08:01	focus
0:08:02	we actually waited them so that they have equal amount of data roughly
0:08:05	uh and then um and then and spanish and egyptian arabic now note we use the egyptian arabic here
0:08:11	because that happen
0:08:12	to be a dataset where we have
0:08:14	about lies
0:08:15	uh
0:08:16	transitioned so we can actually perform this from happening in there you know
0:08:20	pretty straightforward way
0:08:21	and also these
0:08:22	two data sets that have very little data spanish and uh egyptian
0:08:26	they are weighted more heavily to
0:08:28	it's even more about
0:08:30	in terms of the overall model
0:08:33	then this is might be a detail and
0:08:35	known to everybody but
0:08:36	uh
0:08:37	window to this i thought i'd point it out
0:08:39	uh so
0:08:40	when we do the the log likelihood ratio scoring
0:08:43	we actually a do not
0:08:45	use all the languages
0:08:46	in the denominator but only the languages that are not
0:08:49	the target language
0:08:50	and that gives you slightly better result
0:08:57	okay
0:08:58	so here the results using
0:09:00	the uh prlm approach
0:09:02	so
0:09:03	have the three individual
0:09:04	um the individual
0:09:06	P
0:09:07	are aligned
0:09:08	still
0:09:08	american based on the american english recogniser level in arabic and spanish recognisers
0:09:13	and uh american english asthmatic
0:09:16	back
0:09:16	i because it has the most training data gives you the best
0:09:18	individual results
0:09:20	uh and might also be good because american english actually had where english has it
0:09:24	relatively high number of
0:09:26	uh oh
0:09:26	testing phone so that gives you a lot of resolution
0:09:29	in your in your
0:09:29	coding
0:09:30	um and then when you do the standard
0:09:33	um
0:09:34	prlm with first
0:09:36	two recognisers and then three recognisers you get progress
0:09:39	truman
0:09:40	overall from the
0:09:41	single best
0:09:42	uh which is the american english to the three way
0:09:45	you know be prlm
0:09:47	uh with about thirty four
0:09:48	centrality
0:09:49	yeah
0:09:50	and
0:09:50	then these single decoder that uses only the multilingual
0:09:54	a recording
0:09:55	a gives you three point O one which is very close to the
0:09:58	combined
0:09:59	oh and of course vastly simpler and faster
0:10:01	strongly
0:10:02	because
0:10:03	and if you combine these all you have a four way P P R line now
0:10:07	you get another nice to uh improvement so you go from the previous result
0:10:11	the
0:10:12	uh you know what
0:10:13	the three language specific stuff
0:10:14	terms
0:10:15	to a full weight prlm
0:10:17	with a with a pretty significant
0:10:19	uh twenty four percent additional reduction so
0:10:22	usually when you add more and more of these
0:10:24	language specific systems the improvement kind of peter out as you might expect
0:10:28	but if you apply
0:10:30	the multilingual
0:10:31	uh system then you get another big
0:10:35	uh just some details so again this might all be common knowledge but
0:10:39	uh we did find that that was actually no gain from four grams three grams was
0:10:44	the best
0:10:45	in terms of the language model um
0:10:47	you know the overall lack
0:10:48	see
0:10:49	so somehow the
0:10:51	you know the the the
0:10:52	the programs are too sparse all the models are not adequate to capture the information
0:10:56	programs
0:10:57	and it actually
0:10:58	a good to do a fairly
0:11:00	uh suboptimal smoothing in terms of language model performance
0:11:04	the assembly at once moving works best
0:11:06	works better than doing fancy things and i
0:11:10	um
0:11:12	okay
0:11:15	so now we do have some we had something which
0:11:18	is very easily done
0:11:19	in like a bird
0:11:20	system
0:11:21	uh my understanding so we
0:11:22	we use uh
0:11:24	we used we augment the standard
0:11:26	uh cepstral front end
0:11:27	with a mlp features with multilayer perceptron and neural network features
0:11:31	uh which works right very well when we do word recognition
0:11:35	uh and and other tasks
0:11:36	and we also show that a
0:11:38	front end that is trained on saying english too
0:11:41	the form english phone
0:11:43	uh
0:11:43	uh discrimination
0:11:45	uh actually generalises to other layers so you could train
0:11:47	and you're not
0:11:48	to discriminating users english phones at the frame level
0:11:51	and then
0:11:52	uh
0:11:52	use that
0:11:53	train front end
0:11:54	to train say a mandarin recogniser and you would see a nice
0:11:58	so
0:11:58	this is this confidence that is
0:12:00	this
0:12:00	a front end although it is trained on only one language will actually
0:12:04	a generalised to other languages which is exactly what we want for the
0:12:07	language
0:12:09	uh did we find that the across the board for all languages
0:12:13	um
0:12:13	we get uh a small but
0:12:15	consistent improvement in the
0:12:17	the recognition accuracy at the phone level
0:12:20	phone phone recognition or
0:12:21	see
0:12:22	um and now we're gonna throw this at the multilingual prlm
0:12:25	so we can augment the
0:12:27	multilingual prlm
0:12:28	the
0:12:29	uh with
0:12:30	uh with this
0:12:31	uh mlp feature front
0:12:32	and you see an improvement here
0:12:34	that uh is about
0:12:36	um
0:12:37	uh is it that is about
0:12:40	um
0:12:41	you want to see it yet though some three point O one
0:12:43	two two point eight one
0:12:45	and if you do this
0:12:47	combination
0:12:48	with the with the other
0:12:49	language specific
0:12:50	P L M systems
0:12:52	you got improvement from two point oh nine to one
0:12:54	so nice
0:12:55	nice improvement from adding those
0:12:57	yeah i uh
0:12:58	an opium
0:12:59	as others have seen but we wanted to verify it
0:13:02	uh for this
0:13:03	uh
0:13:03	for this framework with the multilingual
0:13:07	okay so now we're gonna try something diff
0:13:08	so another thing that we use
0:13:10	with some success
0:13:11	in speaker
0:13:12	identification of course is the M R trends
0:13:14	um
0:13:15	so
0:13:16	why should we be able to do this
0:13:17	or
0:13:18	language recognition
0:13:19	so the idea
0:13:20	seen it
0:13:20	uh talk about probably he at the workshop is you have a language independent
0:13:25	set of phone models
0:13:27	and uh you use mllr adaptation
0:13:29	so you estimated transform
0:13:31	to move certain phone class
0:13:33	from there
0:13:33	language independent locations or
0:13:36	speaker independent or whatever the in
0:13:37	the
0:13:38	pennants is that you that you care about
0:13:40	to a defendant
0:13:42	to to a location that is specific to a subset of your data such as
0:13:45	language or
0:13:47	and then use the transform coefficients
0:13:49	as features
0:13:50	and you model them with as we have
0:13:53	and an hour
0:13:54	case
0:13:54	we have eight phone classes
0:13:56	each uh
0:13:57	the feature vector
0:13:58	has thirty nine components
0:14:00	and uh the you know the the affine transform of the thirty nine by forty matrix that we get
0:14:05	about thirteen thousand
0:14:07	the twelve thousand
0:14:08	raw
0:14:09	we perform right normalisation as we do in our speaker I D's
0:14:13	and that's our feature vector
0:14:15	and then we do support vector machine training with linear kernels
0:14:18	and
0:14:19	you know the hyperplane is really the the model
0:14:21	the language model
0:14:22	model for the language
0:14:24	case
0:14:24	and the L A D's scores
0:14:26	this
0:14:26	from
0:14:27	your
0:14:27	test sample
0:14:31	yeah the results and this is a very crude system but
0:14:34	bear with me
0:14:35	so um we try this
0:14:37	first with english
0:14:38	uh mlr reference model
0:14:40	so we use
0:14:41	female english
0:14:42	speakers only an hour
0:14:44	in our uh reference model
0:14:45	and we get a
0:14:46	you know we get some results
0:14:47	some people are right
0:14:49	um
0:14:49	we can play this game when we actually combining male and a female transform and we get a better result
0:14:55	insistent with
0:14:56	with what we
0:14:56	see and
0:14:56	speaker i work
0:14:58	uh but when we use a single gender independent
0:15:01	multilingual animal a reference model
0:15:04	we do much better
0:15:05	so this just goes to show first that it works in principle
0:15:08	secondly
0:15:09	that again the multilingual phone models work
0:15:12	we better than the line
0:15:13	fig
0:15:14	oh
0:15:17	now we want to
0:15:18	get this result down to be more competitive with our standard
0:15:21	uh it's a cepstral uh girls
0:15:24	so first of all we can
0:15:25	uh we can we can use a little trick
0:15:28	the training conversations actually pretty long compared to the test
0:15:31	the conversation
0:15:32	that's the test set
0:15:33	so we can actually split our training
0:15:35	conversations
0:15:36	into thirty seconds
0:15:38	uh
0:15:38	segments and get many more data points for the svm training
0:15:42	we can also optimise the number of gaussians in our F
0:15:45	models
0:15:46	to be smaller that forces the mlr to do
0:15:48	more adaptation work
0:15:50	in the transform uh as opposed to just using different regions
0:15:53	if you're gaussian mixture
0:15:55	and finally we can
0:15:56	we can do that now
0:15:58	to try to project out
0:16:00	uh within language
0:16:01	where the light
0:16:03	so um
0:16:04	uh so that
0:16:05	uh that's all done you kind of incrementally and you see that the
0:16:09	that the average uh equal error rate goes down from you know the seven
0:16:14	to just blow for
0:16:16	so
0:16:16	we're not quite
0:16:17	they are yet
0:16:18	far as the
0:16:19	baseline of the cepstral gmm goes but it's much more
0:16:26	okay
0:16:27	um
0:16:28	now again another incremental improvement we augment the plp front end of the mlr system with a twenty five mlp
0:16:35	features
0:16:36	so
0:16:37	a number of features goes from um
0:16:40	uh thirty
0:16:40	uh thirty and nine times forty
0:16:43	to that one
0:16:44	the other block diagonal
0:16:46	opponent that
0:16:47	accounts for adapting the mlp features
0:16:49	which is
0:16:49	twenty five
0:16:50	yeah
0:16:52	six so
0:16:52	overall the feature dimension increases from
0:16:55	the twelve thousand
0:16:56	to uh to just
0:16:57	but you are under eighteen
0:16:59	okay
0:17:00	and
0:17:01	the performance goals
0:17:02	oh
0:17:03	okay
0:17:03	see
0:17:04	improves
0:17:05	uh and well to thirteen
0:17:07	central
0:17:08	i i reduction
0:17:12	okay so now i want to go back
0:17:13	two
0:17:14	phone
0:17:15	phonotactic modelling
0:17:17	and as we've seen
0:17:19	um
0:17:20	i you know hardly anybody uses language models anymore i'm gonna use
0:17:23	yeah
0:17:24	for uh for phonotactic modelling
0:17:26	uh so we wanted to
0:17:27	do the same
0:17:28	uh and see if
0:17:29	if what we saw before still work
0:17:31	so um
0:17:33	as we found also
0:17:35	you know many years ago that in uh in speaker ideas that svm models
0:17:39	uh
0:17:39	plight phone tandem features
0:17:41	uh what better than language
0:17:43	and that
0:17:43	to me
0:17:44	because of
0:17:44	right
0:17:45	um
0:17:47	so
0:17:47	here we want to apply this to the multilingual phone right
0:17:50	hmmm
0:17:50	wars and we use the uh T if there are no good
0:17:54	yeah campbell
0:17:55	uh and we do not perform any rank normalisation
0:17:58	sound like
0:17:59	yeah
0:18:00	sure
0:18:01	the uh again we play this game that we use
0:18:03	split our training uh
0:18:05	conversation sides into
0:18:06	segments that match the length
0:18:08	the test data
0:18:09	and that gives us more
0:18:10	uh that is as more training samples
0:18:13	a four or a smear
0:18:15	uh
0:18:16	so this was our baseline using a language model over the phone and ram
0:18:20	uh and that was the old is all
0:18:22	then when we do an S yeah
0:18:24	that with
0:18:25	with the same feature space trigrams
0:18:27	we do slightly worse
0:18:29	but uh
0:18:30	whereas previously we did not get again with foreground
0:18:34	we now i
0:18:34	you again with foreground
0:18:36	so with
0:18:36	with the additional features that
0:18:39	and uh the result actually get better than up
0:18:42	uh finally we can uh we confuse the two uh phonotactic systems the lm based
0:18:48	the svm based system
0:18:50	and we got another
0:18:52	uh back
0:18:53	so
0:18:54	uh
0:18:55	apparently the svm is a better tool when it comes
0:18:57	to modelling very sparse
0:18:59	uh features
0:19:00	uh
0:19:01	and that's why we see again from going from
0:19:04	like that
0:19:05	we also tried uh using now
0:19:07	but that no gain from that
0:19:09	yeah replicating
0:19:10	something we tried and speaker I D
0:19:12	it didn't work
0:19:13	um however we haven't tried to the uh you know dimensionality reduction
0:19:16	techniques like proposed in the uh in the previous talk so that's certainly something
0:19:20	okay
0:19:22	okay and just kind of the grand finale
0:19:25	what we put everything together
0:19:26	this is our single best system the
0:19:28	oh phone recognition or phonotactic S yeah
0:19:32	with that result
0:19:33	over there
0:19:35	one word yes
0:19:36	and this is our other baseline the cepstral gmm
0:19:40	and then we can
0:19:41	incrementally add
0:19:42	at uh
0:19:43	a phonotactic or or phone based systems
0:19:46	uh we see again from combining the caps O all the combination
0:19:50	start with the cepstral gmm
0:19:52	uh so
0:19:53	we first of all we see that doing mllr type modelling on cepstral features
0:19:58	uh does combine with the with the cepstral gmm
0:20:02	um
0:20:03	the
0:20:04	the multilingual
0:20:06	uh
0:20:06	prlm system
0:20:08	is the best
0:20:08	i think the combined with the baseline
0:20:10	see a whopping
0:20:11	send
0:20:12	uh reduction
0:20:13	there yeah
0:20:14	um and then adding on top of these
0:20:16	these two
0:20:17	uh and you can have you know all the others and you get you go down for another twenty percent
0:20:22	relative so we
0:20:23	essentially you had the error rate from you know two point eight seven that one
0:20:27	or
0:20:28	uh which
0:20:30	looks like a pretty nice reduction
0:20:32	um
0:20:33	the well i really
0:20:35	told you the highlights only
0:20:36	the fact that
0:20:37	two different kinds of phonotactic modelling actually combine
0:20:40	um the fact that you type
0:20:41	cepstral modelling combine
0:20:43	um and the interesting thing is here
0:20:46	that
0:20:46	adding multiple uh this
0:20:48	what behind the P prlm
0:20:50	adding multiple
0:20:51	language
0:20:52	pacific
0:20:53	uh phonotactic model
0:20:55	does not help
0:20:56	okay
0:20:56	the one thing that all these other things it's no longer useful to actually have the line
0:21:00	pacific phone recognition
0:21:05	okay just a quick rundown of some
0:21:07	one of our future direction
0:21:09	so obviously we want to
0:21:11	uh
0:21:11	verify these results with
0:21:13	more recent
0:21:14	uh lre dataset
0:21:15	we want in particular trials on the language
0:21:17	uh the dialogue type the path
0:21:20	and um
0:21:21	you know the svm approach this already
0:21:24	it seemed as
0:21:24	for some the previous talks
0:21:26	can be pursued in the parallel with multiple language specific phone set
0:21:30	um
0:21:31	but more interestingly i think we should we train the mlp features
0:21:34	to actually be a well matched
0:21:36	to the multilingual phone set that we're using now
0:21:39	at the end but
0:21:40	so
0:21:41	uh that's it
0:21:41	yeah button additional improvement
0:21:43	the um
0:21:46	uh
0:21:46	we could all do not very interesting we
0:21:48	could you mlp features for all the language
0:21:50	pacific phone recognisers
0:21:52	the
0:21:52	handling them
0:21:53	we might not really
0:21:54	sue because we
0:21:55	trying to get rid of the lines
0:21:57	fig one recogniser
0:21:58	and we can of course then go to more high level feature uh features that we've tried and worked well
0:22:04	in speaker I D such as prosodic features
0:22:06	and constrained uh caps
0:22:09	okay so here the the
0:22:11	the people messages so we try
0:22:13	there is fine
0:22:14	uh phone based systems
0:22:16	uh for language I D
0:22:17	using techniques that we
0:22:18	uh that we had previously seen uh to work well in asr and also in speaker I D
0:22:25	uh
0:22:25	we for the first time to our knowledge we tried using mllr svm modelling for the language reckon
0:22:30	have a network
0:22:31	um
0:22:32	the uh multilingual i guess
0:22:35	the biggest
0:22:35	take a math
0:22:36	that
0:22:36	the multilingual phone model approach is
0:22:39	is
0:22:40	works
0:22:40	better
0:22:41	and as
0:22:42	is simpler
0:22:43	then using a combination of a language
0:22:47	did you
0:22:47	parable
0:22:49	and it still gives you some games if you combine
0:22:51	language mister
0:22:53	uh phone recognition
0:22:54	um the mlp front end uh can
0:22:57	proof but not that so what others found that mlp fine then you gain line recognition carries over to these
0:23:03	two techniques
0:23:04	that we explored here
0:23:05	and
0:23:06	uh the mllr in the cepstral gmm uh approach
0:23:09	for cepstral modelling also combine quite well
0:23:12	um
0:23:14	well the rest of set already so
0:23:16	that's it
0:23:24	any questions
0:23:25	right
0:23:30	thank you very much for nice to know
0:23:32	and
0:23:33	at the beginning you said that the multilingual phone um did you mention
0:23:37	she works
0:23:38	proximate
0:23:38	the same as the language dependent one
0:23:41	do you have
0:23:42	i mean numbers or
0:23:43	oh no it's not here
0:23:44	no i i have them
0:23:46	you know what home but
0:23:47	but i didn't think it was really relevant
0:23:49	'cause when we imagine phone recognition accuracy
0:23:51	we actually usually apply phonotactic model
0:23:54	but in
0:23:54	for language I D purposes we throw away the phonotactic model because we want
0:23:59	to be very sensitive to the
0:24:00	uh
0:24:01	you know to the
0:24:02	particulars of the line
0:24:03	because
0:24:12	maybe that was the very beginning of the top but we was more details about this
0:24:16	discriminative
0:24:17	mlp features that you're feeding into the
0:24:19	hmmm
0:24:19	phone recogniser or do you like
0:24:21	then the posterior
0:24:23	by some postprocessing or button next yeah so playing with
0:24:26	yes they are actually quite
0:24:27	plaques
0:24:28	uh they were by the way we didn't train anything particular for this language
0:24:32	yeah we just
0:24:33	that's something that we have used in in word rec
0:24:35	yeah
0:24:36	uh in fact it was all the way you know it's basically
0:24:38	these
0:24:39	you just were optimised for
0:24:40	word recognition
0:24:41	and conversational english
0:24:43	telephone
0:24:44	um
0:24:48	uh
0:24:49	so
0:24:49	we take um
0:24:51	we take actually
0:24:52	which a plp features
0:24:53	over a nine frame window
0:24:55	and then perform the usual kind of
0:24:58	M L T
0:24:59	mlp uh
0:25:00	uh training with those
0:25:02	those input features
0:25:03	um we also
0:25:04	form we also use the hats features
0:25:07	which are kind of a derivative of the trap features
0:25:09	uh
0:25:10	going back to
0:25:10	you know like uh a man
0:25:12	work
0:25:12	uh so those capture more long term
0:25:15	uh
0:25:15	critical band energies
0:25:17	um
0:25:17	and then we combine the posteriors from these two mlps
0:25:21	into a single set of posterior vectors
0:25:23	and then um
0:25:24	then we would use it to twenty five dimensions using using uh
0:25:28	yeah
0:25:32	any questions
0:25:38	what is your noise to you
0:25:40	um
0:25:41	i
0:25:42	does
0:25:42	falling on
0:25:43	from scrooge um
0:25:45	uh
0:25:46	the mlp setup
0:25:47	what are you trying
0:25:49	do you
0:25:49	you
0:25:50	beautiful
0:25:51	your mind
0:25:52	ooh
0:25:53	so it's an english phone set so it has another forty five
0:25:56	uh categories
0:25:57	that's performing
0:25:58	frame level class
0:25:59	cation
0:26:00	so you trying to predict
0:26:02	the phone at each frame
0:26:03	uh
0:26:04	english phone of each frame
0:26:06	regardless of the length
0:26:08	so as i said we did not train language specific or even the multilingual mlp
0:26:12	we we just be using the english uh specific mlp that we had
0:26:18	the really
0:26:19	and perhaps try
0:26:20	do you
0:26:21	oh
0:26:22	you might
0:26:22	hmmm
0:26:23	you know that's what it put in the future work as one of the obvious
0:26:26	the proof
0:26:27	that you could actually
0:26:28	generalised concept
0:26:29	cover all languages
0:26:30	and then we try the animal
0:26:32	hmmm
0:26:40	i
0:26:41	one
0:26:42	um
0:26:43	um
0:26:46	huh
0:26:47	oh
0:26:49	uh
0:26:52	it's a fake
0:26:53	it's a it's a uh it's a mapping designed by a phonetician
0:26:57	yeah
0:26:58	this
0:26:58	oh
0:26:59	uh
0:27:01	we plan to do that but
0:27:03	we have
0:27:10	i see
0:27:13	right

Improving Language Recognition with Multilingual Phone Recognition and Speaker Adaptation Transforms

SESSION 10: Language recognition – phonotactics

Přidáno: 14. 7. 2010 11:08, Autor: Andreas Stolcke, Murat Akbacak, Luciana Ferrer, Sachin Kajarekar, Colleen Richey, Nicolas Scheffer, Elizabeth Shriberg (SRI International), Délka: 0:27:19