0:00:13a with the money
0:00:14yeah yeah
0:00:15ladies and and you're
0:00:17uh uh and and sends a long
0:00:19communication code
0:00:21to that
0:00:22is due of a code good chinese got than all science
0:00:25uh it's a great honour for me to speak up out to channel post got need based on what they
0:00:30have to use the moody and noise properties
0:00:32i have divided them my
0:00:34presentation i been into four
0:00:36house
0:00:37you know first pass i would like to
0:00:39say something about background and the motivation
0:00:42in an S hat
0:00:43i would like to insert
0:00:44i would like to study a to channel all get getting at reasons in theory
0:00:49in has three
0:00:51two schemes are proposed
0:00:53to improve noise reduction without introducing audible
0:00:56so peace distortion
0:00:58and in a last pass
0:01:00experiments or out will be giving and we we are makes and conclusions
0:01:06as as we know
0:01:07uh
0:01:08uh the noise the word it's becoming rather and with noise almost everywhere
0:01:13and a background reduce use both speech quality and the speech
0:01:16intelligibility
0:01:18what's more small also increase the dissonance for T
0:01:21so in practice
0:01:23speech enhancement
0:01:24uh including a single symbol channel speech enhancement and the channel which an enhancement is often in have twelve
0:01:33oh we can't have a speech enhancement sort a spatial feeling and the whole to as in a in a
0:01:37have a a in at the table
0:01:39we found that
0:01:41post-filter filter is more efficient for a diffuse noise field
0:01:46and uh we compare the single channel speech enhancement with the model channel
0:01:51filtering filter read and you know no uh table
0:01:54uh from a male table we can find that we can find a to imitation of the single channel speech
0:01:59enhancement
0:02:00first
0:02:01now station noise
0:02:03of of them can all be well suppressed
0:02:05second
0:02:06speech intelligibility
0:02:08can no be well
0:02:10can be significantly improved
0:02:12in generally
0:02:13so
0:02:14multi-channel a a is of them pretty for
0:02:19we present a three gain functions that of a used
0:02:22for post filtering
0:02:25uh previous study
0:02:26the we ask that
0:02:28and noise only segments
0:02:29we
0:02:30the noise reduction should be in finny in theory
0:02:34but
0:02:35and we have a that use be M means that the noise reduction
0:02:39is all and not a not
0:02:41very
0:02:42uh we want to study
0:02:43the reason in the theory
0:02:45this also
0:02:46brings out a new
0:02:48but your and how can we do to improve noise reduction without introducing audible speech
0:02:54distortion
0:02:57that's all i would like to speak about uh
0:03:01background and the modulation
0:03:03now that's ten to the first question why don't know it's action
0:03:07is no in infinity
0:03:09even even a noise only segment
0:03:12we have to some soon two assumptions
0:03:15first
0:03:15the thing or a lot
0:03:17i don't two microphones
0:03:18a gaussian distribution
0:03:20second
0:03:21the noise it's it's and the at to model and that these two assumptions
0:03:27the P we all the ground four but chi square distribution
0:03:31with two degrees of three then
0:03:33and in theory the real and the image only task of look the P both plus
0:03:40it's a but it
0:03:42oh what you me to be use in all level as distribution do you it makes it difficult to obtain
0:03:47that a
0:03:48that this should fusion of the cross
0:03:50spectra
0:03:51so we use of the gaussian distribution
0:03:55to approximate a lower level as
0:03:58distribution we present it in a
0:04:01you know low location
0:04:06a a as we know the all spectra and cross
0:04:09spectral can be a a ten by averaging L
0:04:12independent of friends
0:04:15all the P
0:04:16or a uh
0:04:18P read all right
0:04:19so
0:04:20level
0:04:22that this your job of the or those vector that it's speech and of problem
0:04:26cross a
0:04:27can be of a the has to be
0:04:29we the pdf of low all spectra and a cross spectra
0:04:33the pdf of the gain function can be a at and in theory
0:04:38we brought of the pdfs of logging function in is a was the difference a a
0:04:44smoothing factor of a
0:04:47uh as can be seen from this figure
0:04:50the
0:04:51theoretical or out
0:04:53face power while with the empirical with out
0:04:57oh to obtaining the pdf of the gain functions
0:04:59we can and a nice
0:05:01the theoretical on means of noise reduction
0:05:04and amount of musical noise as link
0:05:06oh think that some pearl
0:05:08we present
0:05:09a a like to you meets of the noise action
0:05:13the the only
0:05:14the like to call me miss off noise reduction can be a but by
0:05:18can but expected value of the gain function
0:05:22and we proud to the
0:05:24noise reduction of a different values of uh
0:05:29was a different well to of the
0:05:31smoothing factor in these figure as can be seen from this figure go of the smoothing fact that increases this
0:05:37the noise reduction can be improved
0:05:40but this that E we can i as is question why don't noise reduction is no in you even at
0:05:46noise only segments
0:05:48and if and only if
0:05:50that's the in fact uh approach is one noise reduction
0:05:54can be in
0:05:58but this study sounding max can be summarised as follows
0:06:02if not move in fact uh it's no last in that
0:06:06you car noise if you have well you to a like heart of the gain function having large very
0:06:12it's better to use a large value of a less smoothing factor to increase both noise reduction and we use
0:06:17musical noise
0:06:19we have to make a chair off between and
0:06:22noise reduction and estimation pass
0:06:25but properly
0:06:26so acting this smoothing factor
0:06:29now a less to a second question how can we do to improve noise reduction
0:06:34without introducing audible speech distortion
0:06:38we propose to study in
0:06:40in this paper
0:06:41the first thing is that adaptive time-frequency frequency keys P
0:06:45as we know for a two channel post filtering out regions
0:06:49the sudden and change of the system only a "'cause" at a a that speech on site and off size
0:06:55so it's better to use a small value of the smoothing factor
0:06:59and that it's that
0:07:00speech on side and all side
0:07:02which use a tools that screen
0:07:05for that it a speech on site
0:07:08the smoothing factor
0:07:11it it is it i mean the by the signal-to-noise ratio
0:07:15for that is that a speech all side
0:07:17the smoothing in fact uh it's gradually increased from zero to one
0:07:23that's take again
0:07:24if sample
0:07:26we
0:07:27the it's now
0:07:28we
0:07:29that let us the don't it's out of the proposal
0:07:32D
0:07:33we can find a from this
0:07:35the
0:07:36a little figure
0:07:38and at it that speech on side
0:07:43and desired
0:07:44so based on side a all size
0:07:46that's smoothing in fact uh has a small value
0:07:49but is very close to zero
0:07:52and a noise only segments
0:07:54the smoothing factor fact is close to one
0:07:57the
0:07:58the without out i expected it
0:08:04the second the skiing is the adaptive which or noise floor action
0:08:09in order to mask me car noise can much speech enhancement
0:08:13all thing use a constant with joe noise for all
0:08:16with a even better
0:08:18the after the equation
0:08:20i have a bit on this
0:08:22psychoacoustic fat
0:08:23is difficult
0:08:25for a home to mask a of noise
0:08:28that's
0:08:29is is the of
0:08:30it's better that of further suppress the
0:08:32tonal no and so we propose a modified gain function
0:08:37two
0:08:38uh we use the number a
0:08:40K L to it at the tonal all components
0:08:44but using in a modified gain function
0:08:46the noise reduction
0:08:48could be improved without introducing audible as
0:08:51speech
0:08:52enhancement
0:08:53because we only increase the amount of noise
0:08:56with action and of peaks of U
0:08:59estimated noise power spectral density
0:09:03that's the an example
0:09:06the noise
0:09:07no no it's it's be thing else like your
0:09:11why noise
0:09:12and then a noise it's is added to clean speech
0:09:17and a segmental signal-to-noise ratio about an pain is a base
0:09:22uh that's this end
0:09:24to is
0:09:27is all ills
0:09:29uh the first one it's the noise this speech
0:09:34right
0:09:34a speakers no
0:09:36hmmm
0:09:40so
0:09:43uh
0:09:45the are we can also see that with the the the experiment bits out from this period
0:09:50from this
0:09:51period right
0:09:53uh
0:10:10uh
0:10:10we can find that
0:10:12the enhanced speech in enhanced speech
0:10:14with a a that you race still noise for action yeah
0:10:18the thing else like a around as the operator of
0:10:21and with the in hand with the proposed
0:10:25D
0:10:25and that the thing is like component
0:10:27and you can pretty T
0:10:29suppressed
0:10:39uh uh
0:10:40but any i would like to compare to propose a re than with the
0:10:44additional signal to channel post filtering
0:10:48a low is at of an no set setup
0:10:49a some as as for
0:10:51we used
0:10:52to map foreigns produced by boat at that distance be seen a two microphone is half a meter
0:10:58the reverberation distance of a no is about one meter the noise speaker it's look at about four meters away
0:11:04from the center of the two microphones
0:11:07that it's that a speech
0:11:08it's located in front of a center my from at a distance of a
0:11:13have a meter
0:11:15we with of the measured coherent
0:11:18of the noise because it in
0:11:21you know task to no
0:11:22using in thought he to lie
0:11:25as a comparison
0:11:26we
0:11:27we problem
0:11:29the
0:11:30the coherence of the diffuse noise field
0:11:33using in a sorry the lie
0:11:35we can see from this figure that
0:11:38the
0:11:39the coherence of the noise has a small value
0:11:42so
0:11:43we can assume that that
0:11:45the speech
0:11:46no the noise it's great it
0:11:50we present the comparison and out of the set of mental as and not improvement
0:11:55and the pesq improvement in these two tables
0:11:59as can be seen from these tables
0:12:01with the proposed as a read and
0:12:04consistent improvements of pulls the
0:12:07segmental a single to noise ratio
0:12:10and the pesq Q can be a but a
0:12:15now that's ten to the conclusion
0:12:19in practice
0:12:21uh uh that's kind a
0:12:22the conversion
0:12:24now we can't i'm sort of quite channel why a noise reduction is no if a need you and
0:12:29at noise it's only segment
0:12:33uh
0:12:33we also propose
0:12:35to help re
0:12:36two
0:12:37improve noise reduction without introducing audible speech distortion in practical consideration
0:12:44one it's the adaptive time-frequency smoothing scheme
0:12:48and the D out a it's the noise property G mean adaptive joe noise floor selection D
0:12:55but any i would like to emphasise that
0:12:57but two schemes could be applied to any speech enhancement and then need to be
0:13:02then need to estimate also of cross vector
0:13:06and the gain function
0:13:08thank you for a attention
0:13:12i
0:13:15do real very
0:13:17i we
0:13:17but from many questions
0:13:19so any questions
0:13:24any any questions from all
0:13:28know
0:13:31i is it possible to try to demonstration restriction
0:13:34be sorry
0:13:35yeah to to use the microphone yeah
0:13:43do do you use you in
0:13:45excuse
0:13:46so there's the question for you
0:13:48well
0:13:48right
0:13:49do you use a computer nor is used a little D
0:13:53where a diffuse noise
0:13:55action really
0:13:56uh
0:13:57you know
0:13:58i actual or just them the
0:14:00you you used to do to a with or me right so for the assumption
0:14:04or the with some correlation white
0:14:08so
0:14:09or just one the ring are
0:14:11do do you
0:14:13a how we you square X i experiment incorporating
0:14:16click you the we so chose
0:14:19uh
0:14:20um
0:14:22the direct and noise
0:14:24or in any kind of the voice such just but we're always of or something
0:14:31you someone
0:14:36uh i think this is a very good question
0:14:38because
0:14:39uh
0:14:40uh i think post filtering the post that that we don't i expect it a coherence space
0:14:46but feel tuning at queen and we have
0:14:49we have
0:14:49to was size of that the noise is
0:14:52and at it
0:14:53and the and noise
0:14:54the noise field is
0:14:56uh
0:14:58if a noise is created a i think is
0:15:01is small
0:15:03it's more efficient to use the
0:15:06uh
0:15:07a
0:15:11to use
0:15:13sorry
0:15:18to use a special of of to me a reader
0:15:25you O any questions
0:15:28you used to solve system okay
0:15:31uh what the what we the demo
0:15:34yeah phone
0:15:37okay okay it's
0:15:49means
0:15:58i
0:15:59oh
0:16:03oh
0:16:08uh this is that you hand to speech without using Z
0:16:11the base you or noise for
0:16:15selection action and this is a but i
0:16:18oh
0:16:19oh
0:16:20i
0:16:22yeah
0:16:23oh
0:16:24i
0:16:25i
0:16:26i
0:16:28and a or a signal so components
0:16:31ask
0:16:32as press
0:16:35and then or or
0:16:36or and you know
0:16:38the noise is i
0:16:39i
0:16:40oh
0:16:45ah
0:16:52okay any question
0:16:55Q and a any questions
0:16:57remote more questions
0:16:59okay if not the strings so speaker