i
but
thank you
and
the work has been carried out uh and good morning and the work has been carried out uh in the
department of electrical and computer engineering at the university of buttons in greece
by a at yeah your run D professor them would open as and my sense
and uh the work is on the binaural extension of
single-channel channel spectral subtraction
reverberation i'm
reverberation has been a challenging research is you for at least forty kate
and
now the verb techniques are applied either there are standard as standalone process
in order to enhance the reverberant signals quality
or even uh increase this reverberant speech intelligibility
or or as preprocessing steps before other several signal processing algorithms and applications in know the to increase their performance
and one one is developing uh binaural dereverberation algorithms
uh it should also take into account some constraint
that are imposed from the binaural aspect of the or a system
so as we all know um when the sound that i've in the left and that a right E channel
here
of the listener
it does as with a relative delay and the relative late uh level different
and these so-called binaural cues are important for the localization of
sound the sound space
and this should definitely be preserved from the binaural signal processing in general or
and more specifically for the binaural from the binaural dereverberation algorithms
on the other hand binaural reverberation
has very appealing applications
it can be applied in hearing aids
in binaural telephony in hands-free devices
in most of the code
a telecommunications
so uh recently we have proposed in our lab uh some single channel dereverberation algorithms
we have proposed a framework for improving single-channel channel existing spectral subtraction dereverberation algorithms
we have also uh presented a novel method um of high computational complexity that gives
uh perceptual sick need
perceptually good results
which is based on perceptual reverberation modeling
and also a fast uh semi-blind reverberation with
that's which is based on the hand club recording
which targets
speech application
so
the state for what step for us was to extent
uh sets a technique and the binaural context
and
and
the most of those uh thing to do was to extend uh the spectral subtraction dereverberation which is
uh techniques of low computational complexity when compared to sophisticated
and what that the remote pro
so the specific gains of this work
uh is to propose a single frame frame for the extension of single-channel channel spectral subtraction dereverberation algorithms
two
uh into use and efficient way
to prevent of estimation errors
and also to evaluate the proposed framework in several state-of-the-art spectral subtraction dereverberation technique
um
expect that subtraction was originally proposed for D knows in application
but recently it has been applied for the suppression of late reverberation
we all know in room acoustics that after the direct sound
the L reflections are i've these are discrete echoes that come from the close surface and produce spectral
uh degradation that is perceived as colouration
in the diffuse feed the late reverberation arrives
which has a the gang noise or a like characteristics
and he's perceived as the well-known signal a reverberant tails
so in the late reverberation suppression some context spectral subtraction
uh gives the any coke estimation by simply um subtracting from the reverberant signal and and then uh and uh
estimation of late variation
and mostly liberation separation methods that work can this way
how to uh estimate exactly these late reverberation spectrum more power spectrum depending on the method
and let's look some state-of-the-art methods
yeah
the methods proposed by where wine gone for we can cut out come from a one will refer to them
as W W an S K A
i i taking someone assumptions on the reverberant signals that these six
while the well known uh reverberation technique from bar to and then be
uh uh from oh no we refer to this as a B
is um a concern assumption on reverberation characteristic
keep in mind that
we can easily express the subtraction um
a principle as again multiplication
uh in the frequency domain by deriving the appropriate gain
so the
a straightforward approach would be to implement separately in the binaural context uh independently this uh late reverberation suppression technique
for the left and the right channel
but it has been proved that the lateral signal processing will destroy this binaural cues and
uh it will make the localization in the produced signal uh be distorted so
and in the bibliography
be i hitting a can team has proposed
uh spectral subtraction extension which is based on the delay and sum beamformer
uh
by beamforming by actually at thing at the left and the right the channels and synchronizing then
um
it produces a reference signal it then makes the late reverberation estimation and the signal and then it apply spectral
subtraction independently
uh in the left and the right yeah
and so the binaural cues are present
in these work
uh i will make an extra samson that the relative delay between uh that to um E S i actually
depends on the weight of the human head and
it can be assumed that it would be uh smaller than the typical analysis windows
so we for this work uh we meet the delay and sum beamformer state
and we propose a binaural extension which is based on a single channel uh spectral subtraction dereverberation on
based and lateral again of station
a see the signal flow of the proposed approach
uh
separately from the web left and the right a rubber and frames with the two different estimations
and uh know the to derive the bi lateral games
then these gains are combined
with a chosen a again of the patient seen
in order to to give us the binaural game
then
again my to the regularization seem that prevents from of very estimation roles that we introduce here is applied
in order to give us a constraint binaural again
which is separately independently
applied on the left and the right frame
the gain adaptation for the gain adaptation in this work was chosen the or to use uh started is
by taking the marks again in it's frequency being uh we had seemed more it's operation and fewer processing artifacts
by taking the average gain would be the compromise between the reverberation reduction and the processing folk
while the minimum gain give significance of print so oppression but
it can be easily introduce artifacts
so the selection of the gain of the patients one was made according to the application scenario
you know there
these blind method as are uh use and introducing uh signal artifacts and to not to to prevent from such
of estimation not different
um
we have
uh probe proposed here we introduce here again a market to the regularization step
which is implemented
uh in the low signal to reverberation or should detector
the assumption here is that um
musical noise or yeah other of estimation that the facts
will a um
we are more probably to uh be present in low signal to reverberation racial frames
and this these and didn't regularization sing
uh depends on a regularization application of to see that
and
on a regularization ratio are
these are user defined parameters that can be a just
in order to um control the suppression rate
so this that um
while properly uh just adjusting these parameters can compensate for estimation error
and prevent musical noise
further explain uh the use of these parameters
these are typical spectral gain functions
and now by keeping seat that to zero point two and are equal to
uh are equal for an a equal or are equal eight we can see how the gain functions
saying
and
but keeping think to uh uh are constant we can change the
two zero point four and zero point sick
so we
from here we can see that a that can be used for the but note um
control of the separation range
why of the parameter R can be used for fine tuning the method
uh let's present some results
uh these results
um
are uh um made with um measure at um
i impulse responses
these uh a specific uh a is since a given from the i can that the base yeah that the
base
in the stairway away for uh with a reverberation time of
zero point seven approximately
note the to evaluate the results
uh we used to metrics the signal to reverberation
or a should difference when compared to the reverberation
to the reverberant signal
so pos difference is be note that the um
more significant reduction
and also um medic the pesq Q uh difference when comparing to the reverberant signal
which relates more to the perceptual
uh quality of the final result
uh we implement
uh this
three by a binaural gain adaptation the patient started is
as well as a delay and sum beamformer or in three state of the art a spectral subtraction dereverberation algorithms
V L B W W gone of gay
and as we can see
uh all of the then any can me significantly reduce reverberation
as we expected the mean gain adaptation seem we'd uses more reverberation while the marks gain less
and when seeing the
where P Q difference which makes more sense in a from a perceptual point of view we can see that
the W W method with the mean game technique
uh gives slightly but the results
these results are taken in the at the uh from the all the would that the base
and
these cafeteria has uh
high reverberation time of one point three seconds
and um
ooh
as we can see that is the reverberation reduction here is um
smaller
and it seems that
such techniques in the sets reverberant conditions
uh and enhance the final signals
but on the other hand uh the enhancement is less than the previous case
again uh the W W to can uh technique i had achieved uh but the results
in terms of
um S R are and press
and uh the best results were uh were observed for the average gain adaptation seen
so we not there to presents some further evaluation we conducted
um
subjective evaluation test
this test was based on the I T U B
eight thirty five and recommendation
and
seventeen test subjects participated in the test
uh we made by a look test not the to get to test the um two
choose the best of the station
and seem for the set it's techniques so for the L B and W W technique
the average gain adaptation was chosen while for the S T A an meaning i the M meaning gain technique
was chosen
and the test subjects were asked
two or rate the speech not real nice
they reverberation intrusive an S and the overall quality of
this speech signals
um
for a in a most K from zero to five
so from these results
we can see that uh the test subjects
rate the dereverberated signal
i net less natural in all cases
however
and we notice a significant reverberation reduction
and also
at least the L B and W W techniques preserve the signal quality the overall signal while
and a for gently we need
um
headphones phones know the to diffuse some them one
but if anyone is interested
uh that then was out of a are available in the web of our group
um B website is also in the paper uh is written in the paper
so to sum up
and
we have introduced a framework for five binaural spectral subtraction dereverberation
which is based on bi lateral gain adaptation
the gain map and the regularization seeing that we introduced can read use the over estimation errors
and produce some uh uh and
um
preserve
from some uh
uh the gradations uh processing the gradations
the selection of the adaptation seem and the D M parameters
uh can be made according to the application scenario
and there is also significant reverberation reduction
uh while the overall speech quality and the binaural cues are
can be present
how there
we noticed some loss of speech naturalness
so for the for us this indicates the need for native binaural mode it's
models that take into account the binaural properties of the to the system
and
this is on what where working right now
thank you very much
okay um we have time for a few questions
you that questions can you just use the microphone over there
any questions from the audience
and questions
okay maybe i just start
okay a how do you
oh man on the uh uh the accuracy of this been all real
oh and uh cues preservation
uh this is a big problem because actually we
the a perceptual test
that can and exactly and um
read the of on the on these
these need to really control the um
and um
environment
and so it was really difficult to do so it's that's actually
uh i think that i i'm not aware of uh and it test for reverberation a graph that
and and um
uh exactly uh predict the these
uh binaural cues preservation
um this is the for the for further investigation
so you have not done any subject you test on this
on these snow
the questions
you know we best you really
uh
another question is how do you did the mean this power meters
you know G R G M R
or
uh
these parameters
actually depend on the frame length
and on the reverberation time on how to store to this you not signal
and we give some uh range for the parameters in the paper
so
uh
actually the they are
different frequencies range
for it's sampling frequency needs frame length
that's the user can that know the to take the optimal results
so for your experiment or for simulations are a bit sorry
use no we made by look test
to tune the parameters
for these
a rules what different environments yes
any questions
so you've not last thanks the speakers again