topic_shift
bool
2 classes
utterance
stringlengths
1
7.9k
session_id
stringlengths
7
14
false
Mm - hmm.
QMSum_120
false
I I I think that you could make the opposite argument that the well - matched case is a fantasy.
QMSum_120
false
Mm - hmm.
QMSum_120
false
You know , so ,
QMSum_120
false
Uh - huh.
QMSum_120
false
I think the thing is is that if you look at the well - matched case versus the po you know , the the medium and the and the fo and then the mismatched case , um , we 're seeing really , really big differences in performance. Right ? And and y you wouldn't like that to be the case. You wouldn't like that as soon as you step outside You know , a lot of the the cases it 's is
QMSum_120
false
Well , that 'll teach them to roll their window up.
QMSum_120
false
I mean , in these cases , if you go from the the , uh I mean , I don't remember the numbers right off , but if you if you go from the well - matched case to the medium , it 's not an enormous difference in the in the the training - testing situation , and and and it 's a really big performance drop.
QMSum_120
false
Mm - hmm.
QMSum_120
false
You know , so , um Yeah , I mean the reference one , for instance this is back old on , uh on Italian uh , was like six percent error for the well - matched and eighteen for the medium - matched and sixty for the for highly - mismatched. Uh , and , you know , with these other systems we we helped it out quite a bit , but still there 's there 's something like a factor of two or something between well - matched and medium - matched. And so I think that if what you 're if the goal of this is to come up with robust features , it does mean So you could argue , in fact , that the well - matched is something you shouldn't be looking at at all , that that the goal is to come up with features that will still give you reasonable performance , you know , with again gentle degregra degradation , um , even though the the testing condition is not the same as the training.
QMSum_120
false
Hmm.
QMSum_120
false
So , you know , I I could argue strongly that something like the medium mismatch , which is you know not compl pathological but I mean , what was the the medium - mismatch condition again ?
QMSum_120
false
Um , it 's Yeah. Medium mismatch is everything with the far microphone , but trained on , like , low noisy condition , like low speed and or stopped car and tested on high - speed conditions , I think , like on a highway and
QMSum_120
false
Right.
QMSum_120
false
So
QMSum_120
false
So it 's still the same same microphone in both cases ,
QMSum_120
false
Same microphone but Yeah.
QMSum_120
false
but , uh , it 's there 's a mismatch between the car conditions. And that 's uh , you could argue that 's a pretty realistic situation
QMSum_120
false
Yeah.
QMSum_120
false
Mm - hmm.
QMSum_120
false
and , uh , I 'd almost argue for weighting that highest. But the way they have it now , it 's I guess it 's it 's They they compute the relative improvement first and then average that with a weighting ?
QMSum_120
false
Yeah.
QMSum_120
false
And so then the that that makes the highly - matched the really big thing.
QMSum_120
false
Mm - hmm.
QMSum_120
false
Um , so , u i since they have these three categories , it seems like the reasonable thing to do is to go across the languages and to come up with an improvement for each of those.
QMSum_120
false
Mm - hmm.
QMSum_120
false
Just say " OK , in the in the highly - matched case this is what happens , in the m the , uh this other m medium if this happens , in the highly - mismatched that happens ".
QMSum_120
false
Mm - hmm.
QMSum_120
false
And , uh , you should see , uh , a gentle degradation through that.
QMSum_120
false
Mmm.
QMSum_120
false
Um. But I don't know.
QMSum_120
false
Yeah.
QMSum_120
false
I think that that I I I gather that in these meetings it 's it 's really tricky to make anything ac make any policy change because everybody has has , uh , their own opinion
QMSum_120
false
Mm - hmm.
QMSum_120
false
and I don't know.
QMSum_120
false
Yeah.
QMSum_120
false
Yeah.
QMSum_120
false
Uh , so Yeah. Yeah , but there is probably a a big change that will be made is that the the baseline th they want to have a new baseline , perhaps , which is , um , MFCC but with a voice activity detector. And apparently , uh , some people are pushing to still keep this fifty percent number. So they want to have at least fifty percent improvement on the baseline , but w which would be a much better baseline.
QMSum_120
false
Mm - hmm. Mm - hmm.
QMSum_120
false
And if we look at the result that Sunil sent , just putting the VAD in the baseline improved , like , more than twenty percent ,
QMSum_120
false
Mm - hmm.
QMSum_120
false
which would mean then then mean that fifty percent on this new baseline is like , well , more than sixty percent improvement on on o e e uh
QMSum_120
false
So nobody would be there , probably. Right ?
QMSum_120
false
Right now , nobody would be there , but Yeah.
QMSum_120
false
Good. Work to do.
QMSum_120
false
Uh - huh.
QMSum_120
true
So whose VAD is Is is this a ?
QMSum_120
false
Uh , they didn't decide yet. I guess i this was one point of the conference call also , but mmm , so I don't know. Um , but Yeah.
QMSum_120
false
Oh.
QMSum_120
false
Oh , I I think th that would be good. I mean , it 's not that the design of the VAD isn't important , but it 's just that it it it does seem to be i uh , a lot of work to do a good job on on that and as well as being a lot of work to do a good job on the feature design ,
QMSum_120
false
Yeah.
QMSum_120
false
so
QMSum_120
false
Yeah.
QMSum_120
false
if we can cut down on that maybe we can make some progress.
QMSum_120
false
M Yeah.
QMSum_120
false
Hmm.
QMSum_120
false
But I guess perhaps I don't know w Yeah. Uh , yeah. Per - e s s someone told that perhaps it 's not fair to do that because the , um to make a good VAD you don't have enough to with the the features that are the baseline features. So mmm , you need more features. So you really need to put more more in the in in the front - end.
QMSum_120
false
Yeah.
QMSum_120
false
So i
QMSum_120
false
Um ,
QMSum_120
false
sure. But i bu
QMSum_120
false
Wait a minute. I I 'm confused.
QMSum_120
false
Yeah.
QMSum_120
false
Wha - what do you mean ?
QMSum_120
false
Yeah , if i
QMSum_120
false
So y so you m s Yeah , but Well , let 's say for ins see , MFCC for instance doesn't have anything in it , uh , related to the pitch. So just just for example. So suppose you 've that what you really wanna do is put a good pitch detector on there and if it gets an unambiguous
QMSum_120
false
Oh , oh. I see.
QMSum_120
false
Mm - hmm.
QMSum_120
false
if it gets an unambiguous result then you 're definitely in a in a in a voice in a , uh , s region with speech. Uh.
QMSum_120
false
So there 's this assumption that the v the voice activity detector can only use the MFCC ?
QMSum_120
false
That 's not clear , but this e
QMSum_120
false
Well , for the baseline.
QMSum_120
false
Yeah.
QMSum_120
false
So so if you use other features then y But it 's just a question of what is your baseline. Right ? What is it that you 're supposed to do better than ?
QMSum_120
false
I g Yeah.
QMSum_120
false
And so having the baseline be the MFCC 's means that people could choose to pour their ener their effort into trying to do a really good VAD
QMSum_120
false
I don't s But they seem like two separate issues.
QMSum_120
false
or tryi They 're sort of separate.
QMSum_120
false
Right ? I mean
QMSum_120
false
Unfortunately there 's coupling between them , which is part of what I think Stephane is getting to , is that you can choose your features in such a way as to improve the VAD.
QMSum_120
false
Yeah.
QMSum_120
false
And you also can choose your features in such a way as to prove improve recognition. They may not be the same thing.
QMSum_120
false
But it seems like you should do both.
QMSum_120
false
You should do both
QMSum_120
false
Right ?
QMSum_120
false
and and I I think that this still makes I still think this makes sense as a baseline. It 's just saying , as a baseline , we know
QMSum_120
false
Mmm.
QMSum_120
false
you know , we had the MFCC 's before , lots of people have done voice activity detectors ,
QMSum_120
false
Mm - hmm.
QMSum_120
false
you might as well pick some voice activity detector and make that the baseline , just like you picked some version of HTK and made that the baseline.
QMSum_120
false
Yeah. Right.
QMSum_120
false
And then let 's try and make everything better. Um , and if one of the ways you make it better is by having your features be better features for the VAD then that 's so be it.
QMSum_120
false
Mm - hmm.
QMSum_120
false
But , uh , uh , uh , at least you have a starting point that 's um , cuz i i some of the some of the people didn't have a VAD at all , I guess. Right ? And and
QMSum_120
false
Yeah.
QMSum_120
false
then they they looked pretty bad and and in fact what they were doing wasn't so bad at all.
QMSum_120
false
Mm - hmm. Mm - hmm.
QMSum_120
false
But , um.
QMSum_120
false
Yeah. It seems like you should try to make your baseline as good as possible. And if it turns out that you can't improve on that , well , I mean , then , you know , nobody wins and you just use MFCC. Right ?
QMSum_120
false
Yeah. I mean , it seems like , uh , it should include sort of the current state of the art that you want are trying to improve , and MFCC 's , you know , or PLP or something it seems like reasonable baseline for the features , and anybody doing this task , uh , is gonna have some sort of voice activity detection at some level , in some way. They might use the whole recognizer to do it but rather than a separate thing , but but they 'll have it on some level. So , um.
QMSum_120