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