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