Přepis řeči - SPEAKER AND NOISE FACTORISATION ON THE AURORA4 TASK

0:00:13	i
0:00:14	yes
0:00:15	factorization
0:00:18	i
0:00:18	for a task
0:00:20	i
0:00:21	paper have proposed a
0:00:23	channel well
0:00:24	and she
0:00:27	i
0:00:29	good morning our run uh were come to talk a for come to to um at all
0:00:34	however have of be very uh please to present our reason to work arounds
0:00:38	a title ways speaker and noise factorization are or for task
0:00:43	and ball on you tell all and this is a joint work with by supervisor all
0:00:48	mark gales
0:00:49	and
0:00:50	so he's here still will we'll
0:00:52	a a a uh uh uh
0:00:54	first slayer out a top or something about the
0:00:57	a a model based approach for robust speech running
0:01:00	speech recognition
0:01:02	and is the is uh uh a uh uh a a a lot of the skiing that that have been
0:01:07	developed over the years that does
0:01:09	uh a specific acoustic
0:01:11	uh us to distortion including speaker adaptation and noise robustness
0:01:16	and the we'll talk about uh we we we discuss discussed options we have to handle multiple acoustic factors
0:01:22	in this call so that the concept of acoustic factorisation is introduced
0:01:26	and as an example we do
0:01:29	uh we derive when you adaptations
0:01:31	which we call joints that handle speaker and noise that uh distortions
0:01:36	and then just uh
0:01:37	i rooms and conclusion
0:01:39	so for we start from the uh environment with as we all know the speech signal can be influenced by
0:01:45	a factor
0:01:46	think i i as in this diagram we have speaker differences
0:01:50	i a channel mismatch
0:01:52	and also some some sort of a back noise and room reverberation noise
0:01:57	also also to do also of this factors can uh fact that speech and decode added it want to variations
0:02:03	and degree or all to uh decoder speech in speech signals
0:02:07	so this makes the
0:02:09	a a robust speech recognition by challenge task
0:02:12	and
0:02:13	so in this work we consider using the model based of porsche
0:02:17	to handle
0:02:18	multiple close factors
0:02:21	so
0:02:22	a in in the in this framework we have a
0:02:25	can all can not cope you look
0:02:26	come of cool model be able to model the disorder versions
0:02:30	and we use a a a a a a set of a transformed to the used at uh i used
0:02:35	to adapt a can not come model to different of course the conditions
0:02:39	and
0:02:41	or the U S a different transforms
0:02:43	has been to about the to hand do a pacific
0:02:46	single acoustic factors including speaker adaptation and noise compensation schemes
0:02:52	and
0:02:53	so yeah
0:02:54	uh a hard but hard to combine
0:02:57	this transforms to handle multiple close to factors
0:03:00	and you know if active and efficient way it is this central topic out this talk
0:03:06	oh
0:03:07	that's look at a first look at speaker adaptation be all know as being a transfer is
0:03:12	uh well
0:03:13	adapt acoustic models
0:03:15	this uh this is on mean transforms
0:03:18	and is this in a transform is very simple but very effective in practice
0:03:23	but this uh a limitation of this this thing a transform that is
0:03:28	uh uh we have a uh right relative a large number of parameters to estimate so we can't do
0:03:33	robust estimation a single options
0:03:36	this
0:03:36	so this this thing in transform cannot be
0:03:39	is not suitable for very rapid at adaptation
0:03:43	so and and then a point interesting point to make it uh is
0:03:47	uh this thing a transform or or uh a was or you don't design of for speaker adaptation but you've
0:03:53	we this strands was a generating a transfer
0:03:56	this can also extend to you my meant to you want men to adaptation
0:04:01	and next so we look at as noise come compensation schemes
0:04:05	normally a mismatch function of a be defined for the impact well environment
0:04:10	uh this is
0:04:11	is the first equation is a use the nonlinear
0:04:15	uh
0:04:15	some is about mean occlusions that relate
0:04:18	i describes how lead to channel distortion and added noise can a fact that clean speech
0:04:24	and it's on this mismatch functions
0:04:26	uh model based approach um
0:04:29	modified a models to
0:04:31	and it to better represent a noisy speech distributions
0:04:35	yeah you the D is used to here
0:04:38	i
0:04:39	the second you creations
0:04:40	shoes how we can adapt acoustic models using vts should based the cool um
0:04:46	which has but is the
0:04:47	but the comp compensation schemes
0:04:49	you can see if only creations that a
0:04:52	do to we use a use the
0:04:54	a mismatch function this
0:04:56	transfer transfer use highly construing and nonlinear
0:04:59	so we can
0:05:00	uh uh we can see that's uh relativist film
0:05:03	for a member of prime is to estimate
0:05:06	so we can do very red we very rapid at that adaptation sings to
0:05:10	a noise transform can be estimated a for a single options
0:05:14	so
0:05:15	and you know about be i talk about in speaker i the speaker adaptation a noise transforms
0:05:21	a noise
0:05:21	compensation schemes
0:05:22	so hard to combine the
0:05:24	in in practice we have a very simple various
0:05:27	straight forward uh a combination schemes of we call this joint a we called this
0:05:32	that's more combination
0:05:33	and the E here you cushion here uh describe some how we can do
0:05:39	a first uh adapt to a a uh a week
0:05:41	the first adapt the acoustic models using vts transforms
0:05:45	and failing dart we have learning a transformed to reduce is mismatch
0:05:49	and this
0:05:51	uh and and the diagram shoes with uh we
0:05:54	we do is sing
0:05:55	how we do is
0:05:56	a given a acoustic addition be as to be noise friends one speaker transform
0:06:01	uh a a proper update or
0:06:02	and and if i that speaker or all noise transform to be to estimate re-estimated post
0:06:08	are are and so
0:06:10	uh at so we can see a a a a a a uh a limitation is obvious that uh
0:06:15	did you know transform should be estimate
0:06:17	on a block of data so this
0:06:19	kind of a combination can out you very rapid a rapid at that adaptations this T
0:06:24	it requires a block up a a a a block update data
0:06:28	and i to me to a we and do you uh uh uh in another way we call this of
0:06:32	acoustic factorisation
0:06:34	in in uh uh we have
0:06:36	we decompose the transform
0:06:38	and a come constrain the each transform to more low as best the good
0:06:42	as the best to suppress the close the factor
0:06:45	in this case we has speak transform and noise transform
0:06:48	which also have a each others
0:06:50	this gives us the some free even two
0:06:53	to to use this transform for example you've we know that same speaker as
0:06:58	a speaking
0:06:59	you know the changing noise conditions
0:07:01	and we want to
0:07:03	uh we we want to update the noise condition for and went to and two we just
0:07:08	a to speech transfer is as we now was speaker has not changed
0:07:11	and the can to noise update uh i i adaptation would just to do a a a nice adaptation
0:07:17	and
0:07:17	a similar way
0:07:19	oh
0:07:20	but environment that use is and change
0:07:22	but a speaker has
0:07:23	it has
0:07:24	has has changed to another speaker speakers we can do
0:07:27	uh make
0:07:28	a speaker transform i out
0:07:29	updating
0:07:30	that we do this noise transforms
0:07:33	so this
0:07:34	that
0:07:34	a a kind of acoustic factorization E
0:07:37	factorization a a lots of this peak transform can be used in a range of noise condition
0:07:41	and similar for noise transform
0:07:43	and that when you sure with this is this approach is that
0:07:47	the transfer what uh should be used the uh we use the transforming of factor i the fashion
0:07:52	that to to estimate a transform need we need to join to estimate both speaker and noise trends are since
0:07:58	that eight or uh of a fact
0:08:01	a a a a a a of of
0:08:03	uh
0:08:04	a a a that of fact the by two
0:08:06	uh to acoustic factors simultaneously
0:08:10	a base on this comes at we derive a new adaptation schemes we call joint that that the king
0:08:16	and this C D to on the right hand side shoes how we manipulate as transforms
0:08:21	first what uh do in contrast to the previous should we do we it T plus them are
0:08:27	this approach
0:08:28	at that you use a a a reversed R a transform with applied to being a transform first and uh
0:08:34	and the modified
0:08:36	uh
0:08:37	clean speech to nice the speech choose to crucial by doing so
0:08:41	i work
0:08:42	you can transform is a acting on the clean speech
0:08:46	and
0:08:47	the the in speaker independent clean speech and the S transform is a up you acting on this speech
0:08:54	speaker dependence
0:08:56	which true shouldn't
0:08:57	this
0:08:58	uh
0:08:59	all do
0:09:00	that are these are problems we
0:09:02	we expect in speaker adaptation all
0:09:05	noise compensation so we expect is
0:09:08	is to transform she can be
0:09:10	uh uh can be a so
0:09:12	can be a some sort of of factor tries all also noise to each other so we can apply
0:09:17	didn't me
0:09:18	we can
0:09:19	the them
0:09:20	so that did i when use D we how we evaluate a hard you by it is uh uh a
0:09:26	joint joint to transform seeing a in the
0:09:29	X runs
0:09:30	so we have for we have this song
0:09:33	a we a you condition data are that's is from noise one
0:09:37	peak K as me the noise phones for an speech transform joint state
0:09:41	and the for and for the same speaker and then and uh i i can a noise condition
0:09:46	we do a bit just a dude noise transform and uh these speak trance of we have all ten the
0:09:51	in i don't in
0:09:52	in the previous uh estimation
0:09:55	and
0:09:55	jen at
0:09:56	at that acoustic models
0:09:58	so that of it has the free and that's
0:10:00	things
0:10:01	uh uh since not only points friends far
0:10:03	uh required do
0:10:05	a a a update
0:10:06	so this can be done or a single options so we can do this
0:10:10	joint to speaker and noise i
0:10:12	a a adaptation
0:10:14	a single options
0:10:15	which is very flexible
0:10:18	so as scroll to the X ones
0:10:20	uh for as when we we you bout the i-th runs on or four task
0:10:24	this is a a is derived from most wrong as a joke one and task
0:10:29	and we have for test set find
0:10:32	there errors uh the in set a
0:10:35	set a a is uh a test or one which is clean set
0:10:39	and test
0:10:39	in set B we have sick
0:10:41	different to a six different types of noise at it
0:10:44	and set C and said D is
0:10:46	comes from the far-field microphones
0:10:49	for the close to model training we do some of are pretty standard stuff
0:10:53	and
0:10:54	this is the X runs from a bashful batch X in a i'm
0:10:58	batch more X in i mean
0:11:00	the speaker and noise transform for i estimate for
0:11:03	a for you shop that's for test set
0:11:06	so
0:11:07	it this no sharing bic uh off speaker transforms
0:11:10	we can see that's uh
0:11:12	by during speaker and noise adaptation
0:11:14	a combine the speaker and noise adaptation
0:11:17	we she we yeah she and signal and things over
0:11:22	noise adaptational
0:11:23	a noise adaptation on only
0:11:25	and we L i can see that sings joint it's just the reverse all they're of each T from of
0:11:30	i am i'm not transform so the order
0:11:32	in share is not a very sensitive to it
0:11:36	it it it is really uh a
0:11:39	it it it it impacts performs it it does not impact for one too much
0:11:44	so
0:11:44	uh but we want to emphasise that this is a batch more X runs
0:11:49	we we we uh this
0:11:51	we recall which requires a a update is to estimate france transforms
0:11:56	so it a is is better uh it he's not very flexible to be used
0:12:00	so what is more interesting is the factorization X ones
0:12:04	we can which uh in this X runs
0:12:06	we have we can uh these estimates speaker transform for a from the clean set we should test or one
0:12:13	and and we applied to speech transforming out the noise conditions
0:12:17	we can see from the uh
0:12:19	so the row of the table
0:12:21	that's a we
0:12:22	we uh this
0:12:23	the speech transform from big from clean speech that's hard for a for the set B C is uh the
0:12:29	out noise is set
0:12:31	and then function at native S plus a that's not
0:12:35	a that's not general at the that actually decrease performance
0:12:38	because the each here the M O transfer in this case
0:12:42	i uh is
0:12:43	i i is acting on the vts adapted the being so uh so uh is
0:12:48	is it is
0:12:49	uh a you can it his uh a social a suspect and noise condition and
0:12:55	can not be used uh that you know i don't noise conditions
0:12:58	and and what is more interesting is that if we estimate transforms forms the speak transform font you got nice
0:13:04	a set
0:13:05	test or for all which is uh i since a restaurant noise
0:13:09	and we have
0:13:10	the
0:13:11	a a a a a would joint to screw adapting skating actually
0:13:15	uh you
0:13:16	that that a get a a a uh guess a sum
0:13:18	i that some better the result
0:13:22	this is an interesting so
0:13:23	and
0:13:25	uh had this night might be a in a need and indicated that
0:13:29	i would join transform uh i'm i'm a transformed in joint
0:13:33	maybe more something that should be more that by
0:13:36	i a by vts transform which is say which means that our factorization maybe not perfect
0:13:43	that them but the a number of is
0:13:45	uh are are uh are a up that nation use a as using the transform a speak transform as they
0:13:51	from i is a shave
0:13:53	for point for or which should use just which is very close to more expert
0:13:59	a a more lax foreigns
0:14:00	this demonstrate
0:14:01	it
0:14:02	we can we can fact tries act a we can fact a speak transform and using sprites speak transforming up
0:14:08	you very good noise stations
0:14:11	so
0:14:13	so i i here i rival at my conclusions
0:14:16	in this talk at i i we argue that um
0:14:19	a handling doing or close to factors Z is important in
0:14:23	being very complex realistic in closing moment
0:14:26	and we present a are powerful and flexible polish test based on the
0:14:30	acoustic factorisation with your the derive when you adaptation skiing because a joint
0:14:36	and and this allows very rapid the speaker and noise adaptation
0:14:41	this
0:14:41	speak transform can can be used a cross them are local acoustic issues
0:14:45	and just a little bit a about this a new X in
0:14:49	a we have to to compare our approach
0:14:52	is the uh uh a feature enhancement you enhancement to style is
0:14:57	a style a but a noise robustness schemes is
0:15:01	and am adaptation a speaker transforms
0:15:04	a speaker adaptation
0:15:05	we we observe that of joining the all all performed this a you have he feature these M are employed
0:15:12	um um
0:15:13	such and factorization mold
0:15:15	and this is demonstrated
0:15:17	a the the power of
0:15:18	but the bayes framework
0:15:20	and the we uh and the inside is from where we have a very powerful and flexible to does sees
0:15:27	that he's acoustic factorisation
0:15:29	second
0:15:36	do have a a time for a couple of questions but have a process to to you might be are
0:15:42	both behind the projectors
0:16:04	so i have the
0:16:06	so questions
0:16:07	to use a speaker
0:16:11	a to the factorization so
0:16:14	quote a close at all
0:16:16	three
0:16:17	to that extent you so the assumption is about
0:16:22	oh
0:16:24	uh
0:16:25	uh yeah i i i think this you are right to that this actually is
0:16:29	um
0:16:30	do not very uh it is not perfect fact uh of factor arise since me
0:16:35	as you can see that uh
0:16:37	uh
0:16:38	but we have a that by on the on the transfer or and we also have a channel distortion actually
0:16:44	is a also a bias on that
0:16:46	that we do
0:16:48	we can't
0:16:49	uh
0:16:49	and but is since
0:16:51	but for the main main part the
0:16:54	we transform is ending at as well and the nice friends for its in a transform
0:16:59	so this to leave different types of transform one combined
0:17:03	is actually the D can be
0:17:04	uh uh
0:17:06	uh uh uh a uh uh a factor rights
0:17:09	and uh as the that there's the X were and demonstrated that
0:17:13	oh
0:17:14	we can't use
0:17:15	you can because a is the fact right
0:17:18	property is quite good
0:17:20	um two
0:17:21	the count to say in met matt in mathematically medically
0:17:24	uh the the if it also a model to each other but we can see
0:17:28	from this we can use the speech transforming wireds conditions
0:17:32	so that's
0:17:33	that's T
0:17:34	a is kind of factor
0:17:35	uh uh also an art
0:17:40	a questions
0:17:45	i i sure can for the speaker

SPEAKER AND NOISE FACTORISATION ON THE AURORA4 TASK

Robust ASR

Přednášející: Yongqiang Wang, Autoři: Yongqiang Wang, Mark Gales, University of Cambridge, United Kingdom