0:00:18thank you
0:00:20the work has been carried out uh and good morning and the work has been carried out uh in the
0:00:25department of electrical and computer engineering at the university of buttons in greece
0:00:29by a at yeah your run D professor them would open as and my sense
0:00:33and uh the work is on the binaural extension of
0:00:37single-channel channel spectral subtraction
0:00:39reverberation i'm
0:00:42reverberation has been a challenging research is you for at least forty kate
0:00:48now the verb techniques are applied either there are standard as standalone process
0:00:53in order to enhance the reverberant signals quality
0:00:56or even uh increase this reverberant speech intelligibility
0:01:01or or as preprocessing steps before other several signal processing algorithms and applications in know the to increase their performance
0:01:10and one one is developing uh binaural dereverberation algorithms
0:01:15uh it should also take into account some constraint
0:01:18that are imposed from the binaural aspect of the or a system
0:01:23so as we all know um when the sound that i've in the left and that a right E channel
0:01:29of the listener
0:01:30it does as with a relative delay and the relative late uh level different
0:01:36and these so-called binaural cues are important for the localization of
0:01:40sound the sound space
0:01:42and this should definitely be preserved from the binaural signal processing in general or
0:01:48and more specifically for the binaural from the binaural dereverberation algorithms
0:01:53on the other hand binaural reverberation
0:01:55has very appealing applications
0:01:58it can be applied in hearing aids
0:02:00in binaural telephony in hands-free devices
0:02:03in most of the code
0:02:04a telecommunications
0:02:07so uh recently we have proposed in our lab uh some single channel dereverberation algorithms
0:02:14we have proposed a framework for improving single-channel channel existing spectral subtraction dereverberation algorithms
0:02:21we have also uh presented a novel method um of high computational complexity that gives
0:02:27uh perceptual sick need
0:02:29perceptually good results
0:02:31which is based on perceptual reverberation modeling
0:02:34and also a fast uh semi-blind reverberation with
0:02:38that's which is based on the hand club recording
0:02:40which targets
0:02:41speech application
0:02:43the state for what step for us was to extent
0:02:47uh sets a technique and the binaural context
0:02:53the most of those uh thing to do was to extend uh the spectral subtraction dereverberation which is
0:02:59uh techniques of low computational complexity when compared to sophisticated
0:03:04and what that the remote pro
0:03:07so the specific gains of this work
0:03:10uh is to propose a single frame frame for the extension of single-channel channel spectral subtraction dereverberation algorithms
0:03:18uh into use and efficient way
0:03:20to prevent of estimation errors
0:03:23and also to evaluate the proposed framework in several state-of-the-art spectral subtraction dereverberation technique
0:03:32expect that subtraction was originally proposed for D knows in application
0:03:37but recently it has been applied for the suppression of late reverberation
0:03:42we all know in room acoustics that after the direct sound
0:03:46the L reflections are i've these are discrete echoes that come from the close surface and produce spectral
0:03:52uh degradation that is perceived as colouration
0:03:55in the diffuse feed the late reverberation arrives
0:03:58which has a the gang noise or a like characteristics
0:04:02and he's perceived as the well-known signal a reverberant tails
0:04:07so in the late reverberation suppression some context spectral subtraction
0:04:11uh gives the any coke estimation by simply um subtracting from the reverberant signal and and then uh and uh
0:04:17estimation of late variation
0:04:20and mostly liberation separation methods that work can this way
0:04:24how to uh estimate exactly these late reverberation spectrum more power spectrum depending on the method
0:04:31and let's look some state-of-the-art methods
0:04:35the methods proposed by where wine gone for we can cut out come from a one will refer to them
0:04:40as W W an S K A
0:04:42i i taking someone assumptions on the reverberant signals that these six
0:04:46while the well known uh reverberation technique from bar to and then be
0:04:50uh uh from oh no we refer to this as a B
0:04:53is um a concern assumption on reverberation characteristic
0:04:57keep in mind that
0:04:59we can easily express the subtraction um
0:05:03a principle as again multiplication
0:05:06uh in the frequency domain by deriving the appropriate gain
0:05:11so the
0:05:11a straightforward approach would be to implement separately in the binaural context uh independently this uh late reverberation suppression technique
0:05:20for the left and the right channel
0:05:23but it has been proved that the lateral signal processing will destroy this binaural cues and
0:05:29uh it will make the localization in the produced signal uh be distorted so
0:05:34and in the bibliography
0:05:37be i hitting a can team has proposed
0:05:40uh spectral subtraction extension which is based on the delay and sum beamformer
0:05:46by beamforming by actually at thing at the left and the right the channels and synchronizing then
0:05:52it produces a reference signal it then makes the late reverberation estimation and the signal and then it apply spectral
0:05:59subtraction independently
0:06:01uh in the left and the right yeah
0:06:03and so the binaural cues are present
0:06:08in these work
0:06:09uh i will make an extra samson that the relative delay between uh that to um E S i actually
0:06:16depends on the weight of the human head and
0:06:18it can be assumed that it would be uh smaller than the typical analysis windows
0:06:23so we for this work uh we meet the delay and sum beamformer state
0:06:28and we propose a binaural extension which is based on a single channel uh spectral subtraction dereverberation on
0:06:35based and lateral again of station
0:06:38a see the signal flow of the proposed approach
0:06:42separately from the web left and the right a rubber and frames with the two different estimations
0:06:49and uh know the to derive the bi lateral games
0:06:52then these gains are combined
0:06:54with a chosen a again of the patient seen
0:06:57in order to to give us the binaural game
0:07:00again my to the regularization seem that prevents from of very estimation roles that we introduce here is applied
0:07:06in order to give us a constraint binaural again
0:07:09which is separately independently
0:07:12applied on the left and the right frame
0:07:16the gain adaptation for the gain adaptation in this work was chosen the or to use uh started is
0:07:21by taking the marks again in it's frequency being uh we had seemed more it's operation and fewer processing artifacts
0:07:28by taking the average gain would be the compromise between the reverberation reduction and the processing folk
0:07:34while the minimum gain give significance of print so oppression but
0:07:38it can be easily introduce artifacts
0:07:41so the selection of the gain of the patients one was made according to the application scenario
0:07:47you know there
0:07:48these blind method as are uh use and introducing uh signal artifacts and to not to to prevent from such
0:07:55of estimation not different
0:07:58we have
0:07:59uh probe proposed here we introduce here again a market to the regularization step
0:08:04which is implemented
0:08:06uh in the low signal to reverberation or should detector
0:08:09the assumption here is that um
0:08:13musical noise or yeah other of estimation that the facts
0:08:16will a um
0:08:18we are more probably to uh be present in low signal to reverberation racial frames
0:08:24and this these and didn't regularization sing
0:08:28uh depends on a regularization application of to see that
0:08:33on a regularization ratio are
0:08:35these are user defined parameters that can be a just
0:08:39in order to um control the suppression rate
0:08:42so this that um
0:08:44while properly uh just adjusting these parameters can compensate for estimation error
0:08:49and prevent musical noise
0:08:52further explain uh the use of these parameters
0:08:56these are typical spectral gain functions
0:09:00and now by keeping seat that to zero point two and are equal to
0:09:05uh are equal for an a equal or are equal eight we can see how the gain functions
0:09:12but keeping think to uh uh are constant we can change the
0:09:17two zero point four and zero point sick
0:09:20so we
0:09:21from here we can see that a that can be used for the but note um
0:09:26control of the separation range
0:09:28why of the parameter R can be used for fine tuning the method
0:09:34uh let's present some results
0:09:36uh these results
0:09:39are uh um made with um measure at um
0:09:44i impulse responses
0:09:45these uh a specific uh a is since a given from the i can that the base yeah that the
0:09:50in the stairway away for uh with a reverberation time of
0:09:54zero point seven approximately
0:09:56note the to evaluate the results
0:09:58uh we used to metrics the signal to reverberation
0:10:01or a should difference when compared to the reverberation
0:10:04to the reverberant signal
0:10:06so pos difference is be note that the um
0:10:09more significant reduction
0:10:11and also um medic the pesq Q uh difference when comparing to the reverberant signal
0:10:17which relates more to the perceptual
0:10:19uh quality of the final result
0:10:22uh we implement
0:10:24uh this
0:10:25three by a binaural gain adaptation the patient started is
0:10:28as well as a delay and sum beamformer or in three state of the art a spectral subtraction dereverberation algorithms
0:10:34V L B W W gone of gay
0:10:37and as we can see
0:10:38uh all of the then any can me significantly reduce reverberation
0:10:43as we expected the mean gain adaptation seem we'd uses more reverberation while the marks gain less
0:10:49and when seeing the
0:10:51where P Q difference which makes more sense in a from a perceptual point of view we can see that
0:10:56the W W method with the mean game technique
0:10:59uh gives slightly but the results
0:11:03these results are taken in the at the uh from the all the would that the base
0:11:09these cafeteria has uh
0:11:11high reverberation time of one point three seconds
0:11:15and um
0:11:17as we can see that is the reverberation reduction here is um
0:11:24and it seems that
0:11:26such techniques in the sets reverberant conditions
0:11:30uh and enhance the final signals
0:11:33but on the other hand uh the enhancement is less than the previous case
0:11:39again uh the W W to can uh technique i had achieved uh but the results
0:11:45in terms of
0:11:46um S R are and press
0:11:49and uh the best results were uh were observed for the average gain adaptation seen
0:11:57so we not there to presents some further evaluation we conducted
0:12:02subjective evaluation test
0:12:04this test was based on the I T U B
0:12:08eight thirty five and recommendation
0:12:13seventeen test subjects participated in the test
0:12:16uh we made by a look test not the to get to test the um two
0:12:20choose the best of the station
0:12:22and seem for the set it's techniques so for the L B and W W technique
0:12:28the average gain adaptation was chosen while for the S T A an meaning i the M meaning gain technique
0:12:33was chosen
0:12:35and the test subjects were asked
0:12:37two or rate the speech not real nice
0:12:40they reverberation intrusive an S and the overall quality of
0:12:44this speech signals
0:12:47for a in a most K from zero to five
0:12:51so from these results
0:12:54we can see that uh the test subjects
0:12:57rate the dereverberated signal
0:13:00i net less natural in all cases
0:13:04and we notice a significant reverberation reduction
0:13:09and also
0:13:10at least the L B and W W techniques preserve the signal quality the overall signal while
0:13:18and a for gently we need
0:13:21headphones phones know the to diffuse some them one
0:13:23but if anyone is interested
0:13:26uh that then was out of a are available in the web of our group
0:13:30um B website is also in the paper uh is written in the paper
0:13:36so to sum up
0:13:39we have introduced a framework for five binaural spectral subtraction dereverberation
0:13:43which is based on bi lateral gain adaptation
0:13:47the gain map and the regularization seeing that we introduced can read use the over estimation errors
0:13:52and produce some uh uh and
0:13:56from some uh
0:13:58uh the gradations uh processing the gradations
0:14:02the selection of the adaptation seem and the D M parameters
0:14:05uh can be made according to the application scenario
0:14:09and there is also significant reverberation reduction
0:14:12uh while the overall speech quality and the binaural cues are
0:14:17can be present
0:14:19how there
0:14:20we noticed some loss of speech naturalness
0:14:24so for the for us this indicates the need for native binaural mode it's
0:14:28models that take into account the binaural properties of the to the system
0:14:33this is on what where working right now
0:14:37thank you very much
0:14:42okay um we have time for a few questions
0:14:46you that questions can you just use the microphone over there
0:14:52any questions from the audience
0:14:59and questions
0:15:01okay maybe i just start
0:15:03okay a how do you
0:15:05oh man on the uh uh the accuracy of this been all real
0:15:09oh and uh cues preservation
0:15:12uh this is a big problem because actually we
0:15:15the a perceptual test
0:15:17that can and exactly and um
0:15:22read the of on the on these
0:15:24these need to really control the um
0:15:27and um
0:15:30and so it was really difficult to do so it's that's actually
0:15:34uh i think that i i'm not aware of uh and it test for reverberation a graph that
0:15:40and and um
0:15:42uh exactly uh predict the these
0:15:44uh binaural cues preservation
0:15:47um this is the for the for further investigation
0:15:51so you have not done any subject you test on this
0:15:53on these snow
0:15:59the questions
0:16:03you know we best you really
0:16:08another question is how do you did the mean this power meters
0:16:12you know G R G M R
0:16:15these parameters
0:16:16actually depend on the frame length
0:16:19and on the reverberation time on how to store to this you not signal
0:16:24and we give some uh range for the parameters in the paper
0:16:31actually the they are
0:16:33different frequencies range
0:16:35for it's sampling frequency needs frame length
0:16:38that's the user can that know the to take the optimal results
0:16:41so for your experiment or for simulations are a bit sorry
0:16:45use no we made by look test
0:16:47to tune the parameters
0:16:49for these
0:16:50a rules what different environments yes
0:16:58any questions
0:16:59so you've not last thanks the speakers again