topic_shift
bool 2
classes | utterance
stringlengths 1
7.9k
| session_id
stringlengths 7
14
|
---|---|---|
false |
I beg your pardon ?
|
QMSum_78
|
false |
Well
|
QMSum_78
|
false |
Which is ?
|
QMSum_78
|
false |
What 's the interesting stuff ?
|
QMSum_78
|
false |
I beg your pardon ?
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Yeah. Th - now you get to tell us what 's the interesting part.
|
QMSum_78
|
false |
Please specify.
|
QMSum_78
|
false |
But
|
QMSum_78
|
false |
Well , uh , I guess the work that 's been done on segmentation would be most
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
I think that would be a good thing to start with.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
OK. Um , and , um , the other thing , uh , which I 'll just say very briefly that maybe relates to that a little bit , which is that , um , uh , one of the suggestions that came up in a brief meeting I had the other day when I was in Spain with , uh , Manolo Pardo and Javier , uh , Ferreiros , who was here before , was , um , why not start with what they had before but add in the non - silence boundaries. So , in what Javier did before when they were doing , um h he was looking for , uh , speaker change points.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
Um. As a simplification , he originally did this only using silence as , uh , a putative , uh , speaker change point.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
And , uh , he did not , say , look at points where you were changing broad sp uh , phonetic class , for instance. And for Broadcast News , that was fine. Here obviously it 's not.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
And , um , so one of the things that they were pushing in d in discussing with me is , um , w why are you spending so much time , uh , on the , uh , feature issue , uh , when perhaps if you sort of deal with what you were using before
|
QMSum_78
|
false |
Uh - huh.
|
QMSum_78
|
false |
and then just broadened it a bit , instead of just ta using silence as putative change point also ?
|
QMSum_78
|
false |
Nnn , yeah.
|
QMSum_78
|
false |
So then you 've got you already have the super - structure with Gaussians and H - you know , simple H M Ms and so forth. And you you might So there was a there was a little bit of a a a a difference of opinion because I I thought that it was it 's interesting to look at what features are useful.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
But , uh , on the other hand I saw that the they had a good point that , uh , if we had something that worked for many cases before , maybe starting from there a little bit Because ultimately we 're gonna end up with some s su kind of structure like that ,
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
where you have some kind of simple HMM and you 're testing the hypothesis that , uh , there is a change.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
So so anyway , I just reporting that.
|
QMSum_78
|
false |
OK.
|
QMSum_78
|
false |
But , uh , uh So. Yeah , why don't we do the speech - nonspeech discussion ?
|
QMSum_78
|
false |
Yeah. Do I I hear you you didn't
|
QMSum_78
|
false |
Speech - nonspeech ? OK.
|
QMSum_78
|
false |
Uh - huh. Yeah.
|
QMSum_78
|
false |
Um , so , uh , what we basically did so far was using the mixed file to to detect s speech or nonspeech portions in that.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
And what I did so far is I just used our old Munich system , which is an HMM - ba based system with Gaussian mixtures for s speech and nonspeech. And it was a system which used only one Gaussian for silence and one Gaussian for speech. And now I added , uh , multi - mixture possibility for for speech and nonspeech.
|
QMSum_78
|
false |
Mm - hmm. Mm - hmm.
|
QMSum_78
|
false |
And I did some training on on one dialogue , which was transcribed by Yeah. We we did a nons s speech - nonspeech transcription.
|
QMSum_78
|
false |
Jose.
|
QMSum_78
|
false |
Adam , Dave , and I , we did , for that dialogue and I trained it on that. And I did some pre - segmentations for for Jane. And I 'm not sure how good they are or what what the transcribers say. They they can use it or ?
|
QMSum_78
|
false |
Uh , they they think it 's a terrific improvement. And , um , it real it just makes a a world of difference.
|
QMSum_78
|
false |
Hmm.
|
QMSum_78
|
false |
And , um , y you also did some something in addition which was , um , for those in which there was , uh , quiet speakers in the mix.
|
QMSum_78
|
false |
Yeah. Uh , yeah. That that was one one one thing , uh , why I added more mixtures for for the speech. So I saw that there were loud loudly speaking speakers and quietly speaking speakers.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
And so I did two mixtures , one for the loud speakers and one for the quiet speakers.
|
QMSum_78
|
false |
And did you hand - label who was loud and who was quiet , or did you just ?
|
QMSum_78
|
false |
I did that for for five minutes of one dialogue
|
QMSum_78
|
false |
Right.
|
QMSum_78
|
false |
and that was enough to to train the system.
|
QMSum_78
|
false |
W What ?
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
And so it it adapts , uh , on while running. So.
|
QMSum_78
|
false |
What kind of , uh , front - end processing did you do ?
|
QMSum_78
|
false |
Hopefully.
|
QMSum_78
|
false |
OK.
|
QMSum_78
|
false |
It 's just our our old Munich , uh , loudness - based spectrum on mel scale twenty twenty critical bands and then loudness.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
And four additional features , which is energy , loudness , modified loudness , and zero crossing rate. So it 's twenty - four twenty - four features.
|
QMSum_78
|
false |
Mmm.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
And you also provided me with several different versions ,
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
which I compared.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
And so you change parameters. What do you wanna say something about the parameters that you change ?
|
QMSum_78
|
false |
Yeah. You can specify the minimum length of speech or and silence portions which you want. And so I did some some modifications in those parameters , basically changing the minimum minimum length for s for silence to have , er to have , um yeah to have more or less , uh , silence portions in inserted. So.
|
QMSum_78
|
false |
Right. So this would work well for , uh , pauses and utterance boundaries and things like that.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Yeah. Yeah.
|
QMSum_78
|
false |
But for overlap I imagine that doesn't work at all ,
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
that you 'll have plenty of s sections that are
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
That 's it. Yeah.
|
QMSum_78
|
false |
Mm - hmm , mm - hmm.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
But
|
QMSum_78
|
false |
That 's true. But it it saves so much time the the transcribers
|
QMSum_78
|
false |
Um
|
QMSum_78
|
false |
Yep.
|
QMSum_78
|
false |
just enormous , enormous savings. Fantastic.
|
QMSum_78
|
false |
That 's great. Um , just qu one quickly , uh , still on the features. So you have these twenty - four features.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Uh , a lot of them are spectral features. Is there a a transformation , uh , like principal components transformation or something ?
|
QMSum_78
|
false |
No.
|
QMSum_78
|
false |
Yeah. It was IS two.
|
QMSum_78
|
false |
No. W w we originally we did that
|
QMSum_78
|
false |
Just
|
QMSum_78
|
false |
but we saw , uh , when we used it , uh , f for our close - talking microphone , which yeah , for our for our recognizer in Munich we saw that w it 's it 's not it 's not so necessary. It it works as well f with with without , uh , a LDA or something.
|
QMSum_78
|
false |
OK. OK. No , I was j curious.
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Mm - hmm.
|
QMSum_78
|
false |
Yeah , I don't think it 's a big deal for this application ,
|
QMSum_78
|
false |
Yeah.
|
QMSum_78
|
false |
Right.
|
QMSum_78
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.