Přepis řeči - AN EVALUATION OF NOISE POWER SPECTRAL DENSITY ESTIMATION ALGORITHMS IN ADVERSE ACOUSTIC ENVIRONMENTS

0:00:15	oh uh i
0:00:17	i mentioned uh this work is uh
0:00:19	i one for and the initial a trained that for a more quickly
0:00:24	on the the a project of do you and at least that's not C G
0:00:30	or or i or one
0:00:32	i know i had the opportunity to work for uh
0:00:35	some uh is search for those
0:00:36	sparse it
0:00:37	i is is that uh
0:00:39	results of out
0:00:40	right
0:00:42	what's that what would like to
0:00:43	you was i a the top P and evaluation of noise power spectral lines
0:00:47	summation them
0:00:49	in at the best an and one
0:00:52	but the outline of mine each face
0:00:55	give the motivation of the whole
0:00:57	and now i in to use you an overview of that is that's that we can see that the uh
0:01:02	a frame or
0:01:04	then uh the evaluation measures that use
0:01:07	and uh the
0:01:08	and of my uh
0:01:10	that a they would be uh experimental results and uh
0:01:14	conclusion
0:01:16	but you know that and noise power spectral density estimation or or uh and noise power
0:01:21	estimation summation is a a a crucial part of a speech enhancement the frequency domain
0:01:27	and uh this N T and in new algorithms have mean in to use in this uh in your
0:01:33	uh but unfortunately there is no uh compressive and you
0:01:38	uh
0:01:39	frame for evaluating
0:01:41	uh
0:01:42	the performance of noise probably made
0:01:45	therefore uh
0:01:47	the aims of the framework that you got uh
0:01:50	look for was
0:01:51	to uh present the
0:01:53	performance of some uh the on and B sent and was power estimators
0:01:58	and uh
0:02:00	we
0:02:01	uh
0:02:02	for the more at the uh uh a new measure two
0:02:05	do you are a more comprehensive uh evaluation of the perform
0:02:11	we can see there uh a a lose them in our framework uh the minimum statistics noise power estimate
0:02:18	uh proposed boy uh rain a marketing two thousand one which is one of the is state of the or
0:02:23	algorithm
0:02:25	a second uh i is that was uh a minimal
0:02:28	controlled recursive averaging or
0:02:31	and crawl
0:02:32	and which is also the the state of the art i is and in this area
0:02:36	quite and two thousand two
0:02:38	V i with them belonging to M prague category
0:02:42	uh the improved version of this algorithm
0:02:45	in cross and uh em crawl which is a and hand
0:02:49	uh minimal controlled recursive averaging
0:02:52	and am cry maybe
0:02:54	uh we consider it's bass notes tracking approach or it's and T what was what hendrix of doesn't eight
0:03:02	to algorithms exams uh based on a minimum unit square error or estimation
0:03:07	which are available in two
0:03:09	uh a different approaches uh mini
0:03:11	mmse U and mmse had two thousand ten
0:03:17	a a a a a uh for evaluation uh to he issues that there uh taken into account and uh
0:03:23	the face
0:03:24	issues that
0:03:26	we one oh the uh evaluation being uh independent of uh a speech enhancement system
0:03:32	uh there is then is that we want to uh separate the effects of any a speech enhancement system
0:03:38	on the performance of noise power estimator
0:03:41	we just focus on the
0:03:43	uh estimation error or of noise
0:03:45	a track
0:03:47	a second issue is that uh we need
0:03:50	it to have a
0:03:51	a suitable able friends noise for our evaluation
0:03:55	and uh uh you have as the
0:03:57	three uh
0:03:59	reasons for our uh consideration
0:04:01	first during link speech activities
0:04:04	uh the an instantaneous and noise but is not available
0:04:08	uh and noise
0:04:10	and most
0:04:10	noise power estimation approach
0:04:14	also sorry in noise reduction approaches require a smooth
0:04:17	uh version of noise
0:04:18	estimate
0:04:20	and uh the last one not that this one if we want to
0:04:23	uh that to use the impact of uh random fluctuations in the
0:04:28	origin original noise uh a pretty okay well
0:04:30	therefore four three
0:04:32	uh
0:04:32	consider the
0:04:35	let for noise not the original
0:04:37	the reference noise which is this means not
0:04:40	the innocent in
0:04:42	so
0:04:44	uh
0:04:45	the first evaluation measure that we can see is
0:04:48	uh mean estimation error or used the most common
0:04:51	uh use
0:04:53	the estimation
0:04:54	uh
0:04:55	evaluation measure
0:04:56	and is defined as the average if there's based being the
0:05:01	yeah reference noise
0:05:03	oh and
0:05:04	estimated well
0:05:05	if you can see that the look you of uh the noise power that are you have shown here
0:05:12	uh the reference noise power boy
0:05:14	uh
0:05:16	a that are you power to and uh as that's the hot
0:05:20	you see uh that the operation the
0:05:23	but shoot the look or a issue over all frequency bins and frame is
0:05:28	where capital K is the number of frequency bins and a capital R i
0:05:33	is the number of frame the
0:05:35	is is that you did not model uh
0:05:38	evaluation measure
0:05:40	and the pope was one is the estimation error value
0:05:43	which is
0:05:45	uh propose in this uh a frame
0:05:47	and in fact if if you imagine the
0:05:51	ratio between
0:05:52	uh a reference noise ball
0:05:55	no if reference noise power and estimated noise power
0:05:58	overall all of frequency bins and frame in this is as the re
0:06:03	we divide these meant to some uh of units and uh
0:06:07	we estimate the variance of is
0:06:10	so you need then the computed value values are
0:06:13	uh i over the number of some units
0:06:17	but
0:06:18	and trap it all N
0:06:20	is there a number of us all units
0:06:22	in a role of the matrix and N is the number of sub units in colour
0:06:28	this body has operator
0:06:30	is the operator which computes the variance of his soap unit in the and so on "'em" score
0:06:36	that in the next nine
0:06:39	uh
0:06:39	here are we shall uh for example the
0:06:42	do you i think the sub you need
0:06:44	uh but the number of frequency bins of this up in it is
0:06:47	uh
0:06:48	case so on
0:06:49	the number of frames in this
0:06:51	i so
0:06:53	so uh i for this uh up units V
0:06:56	compute the value as uh which is uh uh is
0:07:00	uh equation to estimate the value
0:07:03	and here
0:07:04	uh for example um mu
0:07:06	i and is there
0:07:08	mean of the
0:07:10	uh you know of the
0:07:11	expected values
0:07:13	for us stop unit in the and strong
0:07:16	uh in our experiments we consider at the number of uh the number of
0:07:20	yeah frame in this is for just sub unit be to fifteen
0:07:24	and the number of frequency bins ten
0:07:28	yeah
0:07:28	yeah
0:07:30	or each uh present you the
0:07:32	the experiments are settings of the algorithm
0:07:35	uh as i mentioned before i
0:07:37	eight a yeah i it so it's right
0:07:39	where implement implemented
0:07:42	a sampling frequency of all signals is eight khz
0:07:46	uh they've window length as well as the uh you have length
0:07:50	a uh is it one is fifty six samples or sit two the cans
0:07:55	uh
0:07:56	consider it uh
0:07:58	to source of clean is speech signal and are taken from timit database
0:08:03	uh one made a speech on a a one theme in speech
0:08:06	each of with the duration of two you
0:08:10	right yeah concatenating the
0:08:11	short segments
0:08:13	of for example a us
0:08:15	six speakers different speakers
0:08:18	and uh for uh simulating the
0:08:21	uh
0:08:22	adverse environments
0:08:23	acoustic can stick one ms we consider a a seven different type of noise
0:08:28	uh taken from uh sound at as database
0:08:31	this simple less and the
0:08:33	uh
0:08:35	uh
0:08:37	easy S one is the what question noise to estimate uh the on a stationary and white gaussian noise that
0:08:43	we can see that
0:08:44	uh
0:08:45	bank i of noise
0:08:47	uh and sinusoidally modulated white question noise
0:08:50	uh the a noise car noise and traffic want traffic to noise
0:08:54	but the traffic to noise uh mainly can uh contains
0:08:58	uh a home
0:08:59	uh sounds
0:09:01	and the difference between traffic one traffic one is that the traffic to noise
0:09:05	is a more structured and how
0:09:10	at the range of input snrs used from minus five db to twenty db with this sub size of uh
0:09:15	five db
0:09:17	uh
0:09:18	as i talk uh
0:09:20	before the reference noise that's uh was
0:09:23	important for our evaluation
0:09:25	uh finally we decided to
0:09:28	uh can can see that that we can receive temporal a smooth single of the noise pretty a ground
0:09:33	with the
0:09:35	uh
0:09:35	a most you factor of a point nine
0:09:40	yeah i i shall a a uh pretty at a crumb or the noise power
0:09:44	uh which is a uh or
0:09:47	uh frequency bins
0:09:49	uh and is
0:09:50	plot to or frame in this is
0:09:53	uh here you see the
0:09:57	the noise power of white question noise uh bank is to while question noise
0:10:02	uh
0:10:02	on to the
0:10:03	traffic
0:10:04	a noise
0:10:05	and you see the
0:10:06	a a stationary T of the noise
0:10:09	and uh are also a stationary
0:10:10	someone
0:10:17	yeah
0:10:17	here is their results of our yeah evaluation
0:10:21	in terms of uh a not
0:10:23	estimation or or yeah it's make mean estimation error or
0:10:26	and uh estimation error variance for what question noise
0:10:31	that is uh the pick that
0:10:33	for a a eight i'll
0:10:35	uh all this space was tracking mmse mse hmmm X
0:10:38	i mean was it it's six uh on the
0:10:41	mmse you
0:10:43	and is a
0:10:45	uh is uh depicted for uh
0:10:48	uh six level of signal soon as a issue
0:10:52	uh be different colours
0:10:54	and you see that uh or
0:10:57	no signals and was you uh most most of the algorithms perform or less the same
0:11:02	or by increasing the signal to lose a issue
0:11:05	uh
0:11:06	uh some of i with sam's
0:11:08	or it seems to be more susceptible
0:11:11	and uh here
0:11:14	or show the results for
0:11:15	and not a stationary what gaussian noise uh sinusoidally modulated one
0:11:19	uh and you see some of the algorithms on not robust in tracking the noise power
0:11:25	uh but the others us
0:11:27	like us of base notes tracking
0:11:30	mmse hand leaks and mmse you
0:11:33	or or a in the low level of signals signal so those a seems to be one
0:11:38	and fast in tracking of the noise power or
0:11:40	what uh
0:11:42	the um
0:11:44	in terms of the uh
0:11:46	estimation error variance also you see
0:11:48	that the uh same result used to live in the ranking go five reasons
0:11:53	more
0:11:55	is is the
0:11:56	results or babble noise
0:11:59	that's
0:12:00	uh
0:12:01	you presents here
0:12:05	and uh uh here is that not for traffic to noise
0:12:08	uh and we selected actually these noise to for uh the sub bass notes tracking i with them
0:12:14	uh there are uh that in this algorithm uh one of the national uh assumptions used that
0:12:20	uh
0:12:21	the noise uh
0:12:22	shouldn't be
0:12:24	uh
0:12:25	a structure or how because
0:12:27	uh
0:12:28	the
0:12:29	uh
0:12:30	uh this
0:12:31	how how many signals
0:12:33	uh can be calm it
0:12:34	the extra we uh low rank model
0:12:37	and can be confused beat the speech signal in the signal subspace
0:12:41	what of course are some modifications in these uh
0:12:44	algorithms that in to do in the paper
0:12:47	uh but uh for the
0:12:49	algorithms we talked to uh a modulation you see that
0:12:53	uh and want of the mean estimation or
0:12:56	which is uh
0:12:57	what's than the
0:12:59	mm as
0:13:01	so
0:13:05	yeah a is the
0:13:07	actually a short here as some of the results of our evaluation
0:13:11	uh for uh
0:13:13	a limited time
0:13:15	uh
0:13:16	and uh one of the important points that we can't to from the evaluation is that
0:13:21	uh estimation in or writing as a trying was i addition iteration are uh in size uh for the evaluation
0:13:28	of the perform
0:13:30	because a using estimation error variance
0:13:33	we can uh measure the amount of fluctuations in the noise power
0:13:37	uh
0:13:38	uh in the noise estimated noise power
0:13:41	for example of uh to mess so it's performance uh a very close to show the in terms of uh
0:13:47	is mean estimation all
0:13:50	why having the estimation error variance we can
0:13:53	a a get the that to uh and more comprehensive you've performance and i uh the most rate might claim
0:14:00	point
0:14:00	showing an example
0:14:02	uh for example can see if you made speech signal to by uh sinusoidally modulated noise what question noise
0:14:10	at twenty db signal to a issue
0:14:13	in this figure are uh the
0:14:15	you look is the speak in a speech for power
0:14:19	uh at the green a curve is the estimated noise at to live quite in with them
0:14:25	and uh the red curve is the estimated noise by
0:14:29	and crime maybe are result and
0:14:32	there are noise is
0:14:33	uh
0:14:35	the black four
0:14:36	uh
0:14:37	you see that be a different behavior of algorithms in tracking of noise power
0:14:42	for example uh
0:14:45	and cry maybe i'll with them
0:14:46	uh denoted by a red curve
0:14:48	a has the
0:14:50	on the estimation of noise power
0:14:52	uh
0:14:53	well not
0:14:54	a fast
0:14:55	the action to it's tracking of noise
0:14:58	uh a in the
0:15:01	in a i'll is a gives uh
0:15:04	some over a
0:15:05	estimation of noise power we some
0:15:07	uh
0:15:08	a fluctuations
0:15:10	and these fluctuations that are actually
0:15:13	a uh
0:15:15	uh following or a tracking the
0:15:17	speech component
0:15:20	so is important for us to a predict that
0:15:23	if the error or is related to on the estimation or or estimation or the fluctuations
0:15:29	here uh by using estimation
0:15:32	a mean estimation and error you see
0:15:34	uh that the
0:15:39	uh you see that the
0:15:41	estimated value they mean estimation error is uh
0:15:44	the same very close
0:15:49	and uh but a and and also sort the performance of in cried with them use some hope better than
0:15:54	uh
0:15:56	so
0:15:56	uh but there that in prime three
0:15:59	but in terms of the estimation error run ins you see that the
0:16:04	uh in prime a maybe i'd with them gives a less
0:16:07	uh
0:16:08	and hence
0:16:09	and
0:16:10	this shows the
0:16:11	uh three for ability of the
0:16:14	employ a maybe because you
0:16:15	meant
0:16:16	uh gives a less for uh fluctuations or
0:16:21	is more a smooth
0:16:22	the tracking of noise power
0:16:27	here
0:16:27	by uh
0:16:29	and of my presentation be giving some conclusions of the frame work
0:16:33	uh first
0:16:35	conclusion is that this some noise power estimators are sorts of the the
0:16:39	and sensitive to the increase meant
0:16:42	of the
0:16:42	signals to ratio
0:16:44	and uh for some of them you see uh the robustness in this test signals so as issue but for
0:16:51	others no
0:16:52	and this is can uh uh of uh can N is that uh a you having
0:16:57	uh estimation error variance we and a
0:17:00	gets better
0:17:01	uh
0:17:02	inside
0:17:03	two words
0:17:04	comp where comparing the most power estimator
0:17:07	and uh in fact uh
0:17:10	these fluctuations maybe
0:17:12	uh put a
0:17:13	a voice some a musical noise at the end of this beast and enhancement for the enhance the speech
0:17:19	so is important to
0:17:20	predict then want of
0:17:22	but uh fluctuations
0:17:23	and uh the says conclusion is that uh for non is stationary and noise types
0:17:29	uh
0:17:30	few algorithms uh can give us
0:17:33	the
0:17:33	fast tracking of noise power
0:17:35	and uh a according to our experiments we found that the
0:17:39	uh mmse hand leaks uh i i with M is the
0:17:43	a most of was one
0:17:44	uh and
0:17:45	it can
0:17:46	i gonna to guarantee that
0:17:48	the and has a speech at the end of a speech enhancement you you was
0:17:52	more improvement in signals the most issue
0:17:55	for what we don't a can name that these
0:17:58	and was probably estimator we give us better intelligibility
0:18:02	it should be tested thing
0:18:03	and another the uh
0:18:05	well
0:18:07	i Q and
0:18:08	yeah i
0:18:13	so at the end we actually know
0:18:15	which are rules
0:18:17	we should use bleach
0:18:19	solutions
0:18:20	but uh
0:18:20	that might be a
0:18:22	one question
0:18:23	what's about complexity could you comment on that
0:18:27	you know
0:18:28	if we uh
0:18:30	consider the the algorithms speech uh track
0:18:33	uh better in terms of uh mean estimation error or
0:18:36	uh uh uh
0:18:38	mm as to exists no complex
0:18:40	i mean uh and in comparison to some was space knows striking to
0:18:45	a performs that so a fast
0:18:47	okay i thank you
0:18:49	the same question
0:18:50	for a question
0:18:54	john
0:18:57	to to this just to your just time tracking or inside one of asked this question your
0:19:02	i in the babble noise at look to snow three uh power was some changing as much one of your
0:19:09	plots there
0:19:10	i i worked a little bit more constant than i thought it might be
0:19:13	um are are you actually you using map or using a large crowd noise
0:19:18	uh large scale oh parts are you know you sure back but we normally you think and you can actually
0:19:23	hear individual speakers or individual word so is uh is what isn't it distinguishable the
0:19:29	it it it's more broad band or or an edge more two
0:19:32	pinkish maybe a right
0:19:34	right in this figure for example yeah like the
0:19:38	not
0:19:38	that that's what i'm looking at "'cause" you have a couple of spots were kind of shot you know
0:19:44	okay no further a questions thank you once more i

AN EVALUATION OF NOISE POWER SPECTRAL DENSITY ESTIMATION ALGORITHMS IN ADVERSE ACOUSTIC ENVIRONMENTS

Speech Enhancement