Datasets:

elplaguister

/

DTS_line_datasets

like 0

because you get a sense also of posture. Posture , and we could like , you know , block out the person 's face or whatever

QMSum_87

What people were wearing.

QMSum_87

Yeah.

QMSum_87

The fashion statement.

QMSum_87

but but , you know , these are important cues ,

QMSum_87

Oh , Andreas was

QMSum_87

How big their heads are.

QMSum_87

I mean the the how a person is sitting is

QMSum_87

But if you just f But from one picture , I don't know that you really get that.

QMSum_87

Yeah. Andreas was wearing that same old sweater again.

QMSum_87

Right ? You 'd want a video for that , I think.

QMSum_87

It 'd be better than nothing , is is i Just from a single picture I think you can tell some aspects.

QMSum_87

A video , yeah.

QMSum_87

Think so ?

QMSum_87

I mean I I could tell you I mean , if I if I 'm in certain meetings I notice that there are certain people who really do eh The body language is very uh is very interesting in terms of the dominance aspect.

QMSum_87

And And

QMSum_87

Hmm.

QMSum_87

Yeah. And and Morgan had that funny hair again.

QMSum_87

Yeah. Well , I mean you black out the that part.

QMSum_87

Hmm.

QMSum_87

But it 's just , you know , the the body

QMSum_87

He agreed.

QMSum_87

you know ?

QMSum_87

Of course , the where we sit at the table , I find is very interesting , that we do tend to cong to gravitate to the same place each time.

QMSum_87

Yeah.

QMSum_87

and it 's somewhat coincidental. I 'm sitting here so that I can run into the room if the hardware starts , you know , catching fire or something.

QMSum_87

Oh , no , you you just like to be in charge , that 's why you 're sitting

QMSum_87

I just want to be at the head of the table.

QMSum_87

Yeah.

QMSum_87

Take control.

QMSum_87

Speaking of taking control , you said you had some research to talk about.

QMSum_87

Yeah. Yeah.

QMSum_87

Yeah , I 've been playing with , um uh , using the close - talking mike to do to try to figure out who 's speaking. So my first attempt was just using thresholding and filtering , that we talked about about two weeks ago , and so I played with that a little bit , and it works O K , except that it 's very sensitive to your choice of your filter width and your threshold. So if you fiddle around with it a little bit and you get good numbers you can actually do a pretty good job of segmenting when someone 's talking and when they 're not. But if you try to use the same paramenters on another speaker , it doesn't work anymore , even if you normalize it based on the absolute loudness.

QMSum_87

But does it work for that one speaker throughout the whole meeting ?

QMSum_87

It does work for the one speaker throughout the whole meeting. Um Pretty well.

QMSum_87

How did you do it Adam ?

QMSum_87

Pretty well. How did I do it ?

QMSum_87

Yeah.

QMSum_87

What do you mean ?

QMSum_87

I mean , wh what was the

QMSum_87

The algorithm was , uh take o every frame that 's over the threshold , and then median - filter it , and then look for runs.

QMSum_87

Yeah. Mm - hmm.

QMSum_87

So there was a minimum run length ,

QMSum_87

Every frame that 's over what threshold ?

QMSum_87

so that A threshold that you pick.

QMSum_87

In terms of energy ? Ah !

QMSum_87

Yeah.

QMSum_87

OK.

QMSum_87

Say that again ? Frame over fres threshold.

QMSum_87

So you take a each frame , and you compute the energy and if it 's over the threshold you set it to one , and if it 's under the threshold you set it to zero , so now you have a bit stream of zeros and ones.

QMSum_87

Hmm. OK.

QMSum_87

And then I median - filtered that using , um a fairly long filter length. Uh well , actually I guess depends on what you mean by long , you know , tenth of a second sorts of numbers. Um and that 's to average out you know , pitch , you know , the pitch contours , and things like that. And then , uh looked for long runs.

QMSum_87

OK

QMSum_87

And that works O K , if you fil if you tune the filter parameters , if you tune how long your median filter is and how high you 're looking for your thresholds.

QMSum_87

Did you ever try running the filter before you pick a threshold ?

QMSum_87

No. I certainly could though. But this was just I had the program mostly written already so it was easy to do. OK and then the other thing I did , was I took Javier 's speaker - change detector acoustic - change detector , and I implemented that with the close - talking mikes , and unfortunately that 's not working real well , and it looks like it 's the problem is he does it in two passes , the first pass is to find candidate places to do a break. And he does that using a neural net doing broad phone classification and he has the the , uh one of the phone classes is silence. And so the possible breaks are where silence starts and ends. And then he has a second pass which is a modeling a Gaussian mixture model. Um looking for uh whether it improves or or degrades to split at one of those particular places. And what looks like it 's happening is that the even on the close - talking mike the broad phone class classifier 's doing a really bad job.

QMSum_87

Who was it trained on ?

QMSum_87

Uh , I have no idea.

QMSum_87

Hmm.

QMSum_87

I don't remember. Does an do you remember , Morgan , was it Broadcast News ?

QMSum_87

I think so , yeah.

QMSum_87

Um So , at any rate , my next attempt , which I 'm in the midst of and haven't quite finished yet was actually using the uh , thresholding as the way of generating the candidates. Because one of the things that definitely happens is if you put the threshold low you get lots of breaks. All of which are definitely acoustic events. They 're definitely someone talking. But , like , it could be someone who isn't the person here , but the person over there or it can be the person breathing. And then feeding that into the acoustic change detector. And so I think that might work. But , I haven't gotten very far on that. But all of this is close - talking mike , so it 's , uh just just trying to get some ground truth.

QMSum_87

Only with eh uh , but eh I I I think , eh when when , y I I saw the the the the speech from PDA and , eh close talker. I I think the there is a a great difference in the in the signal.

QMSum_87

Oh , absolutely.

QMSum_87

Um but eh I but eh I I I mean that eh eh in the in the mixed file you can find , uh zone with , eh great different , eh level of energy.

QMSum_87

So s my intention for this is is as an aide for ground truth. not

QMSum_87

Um I I think for , eh algorithm based on energy , eh , that um h mmm , more or less , eh , like eh eh , mmm , first sound energy detector.

QMSum_87

Say it again ?

QMSum_87

eh nnn. When y you the detect the the the first at at the end of of the detector of , ehm princ um. What is the the name in English ? the the , mmm , the de detector of , ehm of a word in the in the s in an isolated word in in the background That , uh

QMSum_87

I 'm I 'm not sure what you 're saying , can you try

QMSum_87

I mean that when when you use , eh eh any

QMSum_87

I think he 's saying the onset detector.

QMSum_87

Yeah.

QMSum_87

Onset detector , OK.

QMSum_87

I I think it 's probably to work well eh , because , eh you have eh , in the mixed files a great level of energy. eh and great difference between the sp speaker. And probably is not so easy when you use the the PDA , eh that Because the signal is , eh the in the e energy level.

QMSum_87

Right.

QMSum_87

in in that , eh eh speech file is , eh more similar. between the different eh , speaker , um I I think is eh , it will i is my opinion.

QMSum_87

Right. But different speakers.

QMSum_87

It will be , eh more difficult to to detect bass - tone energy. the the change. I think that , um

QMSum_87

Ah , in the clo in the P D A , you mean ?

QMSum_87

In the PDA.

QMSum_87

Absolutely.

QMSum_87

Yeah. Yeah.

QMSum_87

Yeah , no question. It 'll be much harder.

QMSum_87

Yeah.

QMSum_87

Much harder.

QMSum_87

And the the another question , that when I review the the the work of Javier. I think the , nnn , the , nnn , that the idea of using a neural network to to get a broad class of phonetic , eh from , eh uh a candidate from the the the speech signal. If you have , eh uh , I 'm considering , only because Javier , eh only consider , eh like candidate , the , nnn , eh the silence , because it is the the only model , eh eh , he used that , eh eh nnn , to detect the the possibility of a a change between the between the speaker ,

QMSum_87

Right.

QMSum_87

Um another another research thing , different groups , eh working , eh on Broadcast News prefer to , eh to consider hypothesis eh between each phoneme.

QMSum_87

Mm - hmm. Yeah , when a phone changes.

QMSum_87

Because , I I I think it 's more realistic that , uh only consider the the the the silence between the speaker. Eh there there exists eh silence between between , eh a speaker. is is , eh eh acoustic , eh event , important to to consider.

QMSum_87

Mm - hmm. Mm - hmm.

QMSum_87

I I found that the , eh silence in in many occasions in the in the speech file , but , eh when you have , eh eh , two speakers together without enough silence between between them , eh I think eh is better to use the acoustic change detector basically and I I I IX or , mmm , BIC criterion for consider all the frames in my opinion.

QMSum_87

Mm - hmm. Yeah , the you know , the reason that he , uh just used silence was not because he thought it was better , it was it was it was the place he was starting.

QMSum_87

Yeah.

QMSum_87

Yep.

QMSum_87

So , he was trying to get something going ,

QMSum_87

Yeah.

QMSum_87

and , uh e e you know , as as as is in your case , if you 're here for only a modest number of months you try to pick a realistic goal ,

QMSum_87

Yeah , yeah , yeah , yeah. Yeah. Yeah , yeah , yeah , yeah.

QMSum_87