Datasets:
elplaguister
/
DTS_line_datasets

Dataset card Data Studio Files Community
topic_shift
bool
2 classes
utterance
stringlengths
1
7.9k
session_id
stringlengths
7
14
false
Mm - hmm.
QMSum_86
false
But , um , they 're probably not at all set right for these things , particularly these things that look over , uh , larger time windows , in one way or another with with LDA and KLT and neural nets and all these things. In the fa past we 've always found that we had to increase the insertion penalty to to correspond to such things. So , I think that 's , uh , @ @ that 's kind of a first - order thing that that we should try.
QMSum_86
false
So for th so the experiment is to , um , run our front - end like normal , with the default , uh , insertion penalties and so forth , and then tweak that a little bit and see how much of a difference it makes
QMSum_86
false
So by " our front - end " I mean take , you know , the Aurora - two s take some version that Stephane has that is , you know , our current best version of something.
QMSum_86
false
if we were Mm - hmm.
QMSum_86
false
Um. I mean , y don't wanna do this over a hundred different things that they 've tried but , you know , for some version that you say is a good one. You know ? Um. How how much , uh , does it improve if you actually adjust that ?
QMSum_86
false
OK.
QMSum_86
false
But it is interesting. You say you you have for the noisy How about for the for the mismatched or or or or the or the medium mismatched conditions ? Have you ? When you adjusted those numbers for mel cepstrum , did it ?
QMSum_86
false
Uh , I I don't remember off the top of my head. Um. Yeah. I didn't even write them down. I I I don't remember. I would need to Well , I did write down , um So , when I was doing I just wrote down some numbers for the well - matched case.
QMSum_86
false
Yeah.
QMSum_86
false
Um. Looking at the I wrote down what the deletions , substitutions , and insertions were , uh , for different numbers of states per phone.
QMSum_86
false
Yeah.
QMSum_86
false
Um , but , uh , that that 's all I wrote down.
QMSum_86
false
OK.
QMSum_86
false
So. I I would Yeah. I would need to do that.
QMSum_86
false
OK. So
QMSum_86
false
I can do that for next week.
QMSum_86
false
Yeah. And , um Yeah. Also , eh , eh , sometimes if you run behind on some of these things , maybe we can get someone else to do it and you can supervise or something. But but I think it would be it 'd be good to know that.
QMSum_86
false
OK. I just need to get , um , front - end , uh , stuff from you
QMSum_86
false
Hmm.
QMSum_86
false
or you point me to some files that you 've already calculated.
QMSum_86
false
Yeah. Alright.
QMSum_86
false
OK. Uh.
QMSum_86
false
I probably will have time to do that and time to play a little bit with the silence model.
QMSum_86
false
Mm - hmm.
QMSum_86
false
So maybe I can have that for next week when Hynek 's here.
QMSum_86
false
Yeah.
QMSum_86
false
Mm - hmm.
QMSum_86
false
Yeah. Cuz , I mean , the the other That , in fact , might have been part of what , uh , the difference was at least part of it that that we were seeing. Remember we were seeing the SRI system was so much better than the tandem system.
QMSum_86
false
Hmm.
QMSum_86
false
Part of it might just be that the SRI system , they they they always adjust these things to be sort of optimized ,
QMSum_86
false
Is there ?
QMSum_86
false
and
QMSum_86
false
I wonder if there 's anything that we could do to the front - end that would affect the insertion
QMSum_86
false
Yes. I think you can.
QMSum_86
false
What could you do ?
QMSum_86
false
Well , um uh , part of what 's going on , um , is the , uh , the range of values. So , if you have something that has a much smaller range or a much larger range , and taking the appropriate root.
QMSum_86
false
Oh. Mm - hmm.
QMSum_86
false
You know ? If something is kind of like the equivalent of a bunch of probabilities multiplied together , you can take a root of some sort. If it 's like seven probabilities together , you can take the seventh root of it or something , or if it 's in the log domain , divide it by seven.
QMSum_86
false
Mm - hmm.
QMSum_86
false
But but , um , that has a similar effect because it changes the scale of the numbers of the differences between different candidates from the acoustic model
QMSum_86
false
Oh , right.
QMSum_86
false
as opposed to what 's coming from the language model.
QMSum_86
false
So that w Right. So , in effect , that 's changing the value of your insertion penalty.
QMSum_86
false
Yeah. I mean , it 's more directly like the the language scaling or the , uh the model scaling or acoustic scaling ,
QMSum_86
false
That 's interesting.
QMSum_86
false
but you know that those things have kind of a similar effect to the insertion penalty
QMSum_86
false
Mm - hmm.
QMSum_86
false
anyway. They 're a slightly different way of of handling it.
QMSum_86
false
Right.
QMSum_86
false
So , um
QMSum_86
false
So if we know what the insertion penalty is , then we can get an idea about what range our number should be in ,
QMSum_86
false
I think so.
QMSum_86
false
so that they match with that.
QMSum_86
false
Yeah. Yeah. So that 's why I think that 's another reason other than curiosity as to why i it would in fact be kinda neat to find out if we 're way off. I mean , the other thing is , are aren't we seeing ? Y y
QMSum_86
false
Mm - hmm.
QMSum_86
false
I 'm sure you 've already looked at this bu in these noisy cases , are ? We are seeing lots of insertions. Right ? The insertion number is quite high ?
QMSum_86
false
Yeah.
QMSum_86
false
I know the VAD takes pre care of part of that ,
QMSum_86
false
Yeah.
QMSum_86
false
Yeah.
QMSum_86
false
but
QMSum_86
false
I 've seen that with the mel cepstrum. I don't I don't know about the Aurora front - end , but
QMSum_86
false
I think it 's much more balanced with , uh when the front - end is more robust. Yeah. I could look at it at this. Yeah. Mm - hmm.
QMSum_86
false
Yeah. Wha - what 's a typical number ?
QMSum_86
false
I don't I don't know.
QMSum_86
false
Do we ? Oh , you oh , you don't know.
QMSum_86
false
I don't have this in
QMSum_86
false
OK. I 'm sure it 's more balanced ,
QMSum_86
false
Mm - hmm.
QMSum_86
false
but it it it wouldn't surprise me if there 's still
QMSum_86
false
Mm - hmm.
QMSum_86
false
I mean , in in the the the old systems we used to do , I I uh , I remember numbers kind of like insertions being half the number of deletions , as being and both numbers being tend to be on the small side comparing to to , uh , substitutions.
QMSum_86
false
Mm - hmm.
QMSum_86
false
Well , this the whole problem with insertions was what I think , um , we talked about when the guy from OGI came down that one time and and that was when people were saying , well we should have a , uh , uh , voice activity detector
QMSum_86
false
Right.
QMSum_86
false
that , because all that stuff that we 're getting thr the silence that 's getting through is causing insertions. So.
QMSum_86
false
Mmm.
QMSum_86
false
Right.
QMSum_86
false
I 'll bet you there 's still a lot of insertions.
QMSum_86
false
Mm - hmm.
QMSum_86
false
Yeah. And it may be less of a critical thing. I mean , the fact that some get by may be less of a critical thing if you , uh , get things in the right range.
QMSum_86
false
Mm - hmm.
QMSum_86
false
So , I mean , the insertions is is a symptom. It 's a symptom that there 's something , uh , wrong with the range.
QMSum_86
false
Right.
QMSum_86
false
But there 's uh , your your your substitutions tend to go up as well. So , uh , I I I think that ,
QMSum_86
false
Mm - hmm.
QMSum_86
false
uh , the most obvious thing is just the insertions , @ @. But Uh um. If you 're operating in the wrong range I mean , that 's why just in general , if you change what these these penalties and scaling factors are , you reach some point that 's a that 's a minimum. So. Um. Um. We do have to do well over a range of different conditions , some of which are noisier than others. Um. But , um , I think we may get a better handle on that if we if we see Um , I mean we ca it 's if we actually could pick a a a more stable value for the range of these features , it , um , uh , could Uh Even though it 's it 's it 's true that in a real situation you can in fact adjust the these these scaling factors in the back - end , and it 's ar artificial here that we 're not adjusting those , you certainly don't wanna be adjusting those all the time. And if you have a nice front - end that 's in roughly the right range
QMSum_86
false
Hmm.
QMSum_86
false
I remember after we got our stuff more or less together in the previous systems we built , that we tended to set those scaling factors at kind of a standard level , and we would rarely adjust them again , even though you could get a
QMSum_86
false
Mm - hmm.
QMSum_86
false
for an evaluation you can get an extra point or something if you tweaked it a little bit. But , once we knew what rou roughly the right operating range was , it was pretty stable , and Uh , we might just not even be in the right operating range.
QMSum_86
false
So , would the ? Uh , would a good idea be to try to map it into the same range that you get in the well - matched case ? So , if we computed what the range was in well - matched , and then when we get our noisy conditions out we try to make it have the same range as ?
QMSum_86
false
No. You don't wanna change it for different conditions. No. No. I I I What what I 'm saying
QMSum_86
false
Oh , I wasn't suggesting change it for different conditions. I was just saying that when we pick a range , we we wanna pick a range that we map our numbers into
QMSum_86
false
Yeah.
QMSum_86
false
we should probably pick it based on the range that we get in the well - matched case. Otherwise , I mean , what range are we gonna choose to to map everything into ?
QMSum_86
false
Well. It depends how much we wanna do gamesmanship and how much we wanna do I mean , i if he it to me , actually , even if you wanna be play on the gamesmanship side , it can be kinda tricky. So , I mean , what you would do is set the set the scaling factors , uh , so that you got the best number for this point four five times the you know , and so on.
QMSum_86
false
Mm - hmm.
QMSum_86
false
But they might change that those weightings.
QMSum_86