elplaguister/DTS_line_datasets · Datasets at Hugging Face

topic_shift bool 2 classes	utterance stringlengths 1 7.9k	session_id stringlengths 7 14
false	Hmm. OK. Y you 're you were saying before ?	QMSum_86
false	Uh. Yeah. So , this noise , um Yeah. The MSG Um. Mmm. There is something perhaps , I could spend some days to look at this thing , cuz it seems that when we train networks on let 's say , on TIMIT with MSG features , they they look as good as networks trained on PLP. But , um , when they are used on on the SpeechDat - Car data , it 's not the case oh , well. The MSG features are much worse , and so maybe they 're , um , less more sensitive to different recording conditions , or Shou	QMSum_86
false	Shouldn't be. They should be less so.	QMSum_86
false	Yeah. But	QMSum_86
false	R right ?	QMSum_86
false	Mmm.	QMSum_86
false	Wh - ? But let me ask you this. What what 's the , um ? Do you kno recall if the insertions were were higher with MSG ?	QMSum_86
false	I don't know. I cannot tell. But It 's it the the error rate is higher. So , I don	QMSum_86
false	Yeah. But you should always look at insertions , deletions , and substitutions.	QMSum_86
false	Yeah. Mm - hmm.	QMSum_86
false	So	QMSum_86
false	Mm - hmm.	QMSum_86
false	so , uh MSG is very , very dif Eh , PLP is very much like mel cepstrum. MSG is very different from both of them.	QMSum_86
false	Mm - hmm.	QMSum_86
false	So , if it 's very different , then this is the sort of thing I mean I 'm really glad Andreas brought this point up. I sort of had forgotten to discuss it. Um. You always have to look at how this uh , these adjustments , uh , affect things. And even though we 're not allowed to do that , again we maybe could reflect that back to our use of the features.	QMSum_86
false	Mm - hmm.	QMSum_86
false	So if it if in fact , uh The problem might be that the range of the MSG features is quite different than the range of the PLP or mel cepstrum.	QMSum_86
false	Mm - hmm. Mm - hmm.	QMSum_86
false	And you might wanna change that.	QMSum_86
false	But Yeah. But , it 's d it 's after Well , it 's tandem features , so Mmm.	QMSum_86
false	Yeah.	QMSum_86
false	Yeah. We we have estimation of post posteriors with PLP and with MSG as input ,	QMSum_86
false	Yeah.	QMSum_86
false	so I don Well. I don't know.	QMSum_86
false	That means they 're between zero and one.	QMSum_86
false	Mm - hmm.	QMSum_86
false	But i it it it it doesn't necessarily You know , they could be , um Do - doesn't tell you what the variance of the things is.	QMSum_86
false	Mmm. Mm - hmm.	QMSum_86
false	Right ? Cuz if you 're taking the log of these things , it could be , uh Knowing what the sum of the probabilities are , doesn't tell you what the sum of the logs are.	QMSum_86
false	Mm - hmm. Yeah.	QMSum_86
false	So.	QMSum_86
false	Yeah. So we should look at the likelihood , or or what ? Or well , at the log , perhaps , and	QMSum_86
false	Yeah. Yeah.	QMSum_86
false	Mm - hmm.	QMSum_86
false	Or what you know , what you 're uh the thing you 're actually looking at.	QMSum_86
false	Mm - hmm.	QMSum_86
false	So your your the values that are are actually being fed into HTK.	QMSum_86
false	Mm - hmm. But	QMSum_86
false	What do they look like ?	QMSum_86
false	No And so th the , uh for the tandem system , the values that come out of the net don't go through the sigmoid. Right ? They 're sort of the pre - nonlinearity values ?	QMSum_86
false	Yes.	QMSum_86
false	Right. So they 're kinda like log probabilities is what I was saying.	QMSum_86
false	And those OK. And tho that 's what goes into HTK ?	QMSum_86
false	Uh , almost. But then you actually do a KLT on them.	QMSum_86
false	OK.	QMSum_86
false	Um. They aren't normalized after that , are they ?	QMSum_86
false	Mmm. No , they are not no.	QMSum_86
false	No. OK. So , um. Right. So the question is Yeah. Whatever they are at that point , um , are they something for which taking a square root or cube root or fourth root or something like that is is gonna be a good or a bad thing ? So.	QMSum_86
false	Mm - hmm.	QMSum_86
false	Uh , and that 's something that nothing nothing else after that is gonna Uh , things are gonna scale it Uh , you know , subtract things from it , scale it from it , but nothing will have that same effect. Um. So. Um. Anyway , eh	QMSum_86
false	Yeah. Cuz if if the log probs that are coming out of the MSG are really big , the standard insertion penalty is gonna have very little effect	QMSum_86
false	Well , the Right.	QMSum_86
false	compared to , you know , a smaller set of log probs.	QMSum_86
false	Yeah. No. Again you don't really look at that. It 's something that , and then it 's going through this transformation that 's probably pretty close to It 's , eh , whatever the KLT is doing. But it 's probably pretty close to what a a a discrete cosine transformation is doing.	QMSum_86
false	Yeah.	QMSum_86
false	But still it 's it 's not gonna probably radically change the scale of things. I would think. And , uh Yeah. It may be entirely off and and it may be at the very least it may be quite different for MSG than it is for mel cepstrum or PLP. So that would be So the first thing I 'd look at without adjusting anything would just be to go back to the experiment and look at the , uh , substitutions , insertions , and deletions. And if the if the , uh i if there 's a fairly large effect of the difference , say , uh , uh , the r ratio between insertions and deletions for the two cases then that would be , uh , an indicator that it might might be in that direction.	QMSum_86
false	Mm - hmm. Mm - hmm. Yeah. But ,	QMSum_86
false	Anything else ?	QMSum_86
false	my my point was more that it it works sometimes and but sometimes it doesn't work.	QMSum_86
false	Yeah.	QMSum_86
false	So.	QMSum_86
false	Well.	QMSum_86
false	And it works on TI - digits and on SpeechDat - Car it doesn't work , and	QMSum_86
false	Yeah.	QMSum_86
false	Mm - hmm. Yeah. Well.	QMSum_86
false	But , you know , some problems are harder than others ,	QMSum_86
false	Mm - hmm. Yeah.	QMSum_86
false	and And , uh , sometimes , you know , there 's enough evidence for something to work and then it 's harder , it breaks. You know ,	QMSum_86
false	Mm - hmm.	QMSum_86
false	so it 's But it but , um , i it it could be that when you say it works maybe we could be doing much better , even in TI - digits. Right ?	QMSum_86
false	Yeah. Yeah , sure.	QMSum_86
false	So.	QMSum_86
false	Uh.	QMSum_86
false	Hmm ? Yeah.	QMSum_86
false	Yeah. Well , there is also the spectral subtraction , which , um I think maybe we should , uh , try to integrate it in in our system.	QMSum_86
false	Yeah.	QMSum_86
false	Mmm. Mm - hmm.	QMSum_86
false	Right.	QMSum_86
false	But ,	QMSum_86
false	I think that would involve to to mmm use a big a al already a big bunch of the system of Ericsson. Because he has spectral subtraction , then it 's followed by , um , other kind of processing that 's are dependent on the uh , if it 's speech or noi or silence.	QMSum_86
false	Mm - hmm.	QMSum_86
false	And there is this kind of spectral flattening after if it 's silence , and and s I I think it 's important , um , to reduce this musical noise and this this increase of variance during silence portions. So. Well. This was in this would involve to take almost everything from from the this proposal and and then just add some kind of on - line normalization in in the neural network. Mmm.	QMSum_86
false	OK. Well , this 'll be , I think , something for discussion with Hynek next week.	QMSum_86
false	Yeah. Mm - hmm.	QMSum_86
true	Yeah. OK. Right. So. How are , uh , uh how are things going with what you 're doing ?	QMSum_86
false	Oh. Well , um , I took a lot of time just getting my taxes out of the way multi - national taxes. So , I 'm I 'm starting to write code now for my work but I don't have any results yet. Um , i it would be good for me to talk to Hynek , I think , when he 's here.	QMSum_86
false	Yeah.	QMSum_86
false	Do you know what his schedule will be like ?	QMSum_86
false	Uh , he 'll be around for three days.	QMSum_86
false	OK. So , y	QMSum_86
false	Uh , we 'll have a lot of time.	QMSum_86
false	OK.	QMSum_86
false	So , uh Um. I 'll , uh You know , he 's he 'll he 'll be talking with everybody in this room So.	QMSum_86
false	But you said you won't you won't be here next Thursday ?	QMSum_86
false	Not Thursday and Friday. Yeah. Cuz I will be at faculty retreat.	QMSum_86
false	Hmm.	QMSum_86
false	So. I 'll try to connect with him and people as as I can on on Wednesday. But Um. Oh , how 'd taxes go ? Taxes go OK ?	QMSum_86
false	Mmm. Yeah.	QMSum_86
false	Yeah. Oh , good. Yeah. Yeah. That 's just that 's that 's one of the big advantages of not making much money is the taxes are easier. Yeah.	QMSum_86
false	Unless you 're getting money in two countries.	QMSum_86

false

Hmm. OK. Y you 're you were saying before ?

QMSum_86

false

Uh. Yeah. So , this noise , um Yeah. The MSG Um. Mmm. There is something perhaps , I could spend some days to look at this thing , cuz it seems that when we train networks on let 's say , on TIMIT with MSG features , they they look as good as networks trained on PLP. But , um , when they are used on on the SpeechDat - Car data , it 's not the case oh , well. The MSG features are much worse , and so maybe they 're , um , less more sensitive to different recording conditions , or Shou

QMSum_86

false

Shouldn't be. They should be less so.

QMSum_86

false

Yeah. But

QMSum_86

false

R right ?

QMSum_86

false

Mmm.

QMSum_86

false

Wh - ? But let me ask you this. What what 's the , um ? Do you kno recall if the insertions were were higher with MSG ?

QMSum_86

false

I don't know. I cannot tell. But It 's it the the error rate is higher. So , I don

QMSum_86

false

Yeah. But you should always look at insertions , deletions , and substitutions.

QMSum_86

false

Yeah. Mm - hmm.

QMSum_86

false

So

QMSum_86

false

Mm - hmm.

QMSum_86

false

so , uh MSG is very , very dif Eh , PLP is very much like mel cepstrum. MSG is very different from both of them.

QMSum_86

false

Mm - hmm.

QMSum_86

false

So , if it 's very different , then this is the sort of thing I mean I 'm really glad Andreas brought this point up. I sort of had forgotten to discuss it. Um. You always have to look at how this uh , these adjustments , uh , affect things. And even though we 're not allowed to do that , again we maybe could reflect that back to our use of the features.

QMSum_86

false

Mm - hmm.

QMSum_86

false

So if it if in fact , uh The problem might be that the range of the MSG features is quite different than the range of the PLP or mel cepstrum.

QMSum_86

false

Mm - hmm. Mm - hmm.

QMSum_86

false

And you might wanna change that.

QMSum_86

false

But Yeah. But , it 's d it 's after Well , it 's tandem features , so Mmm.

QMSum_86

false

Yeah.

QMSum_86

false

Yeah. We we have estimation of post posteriors with PLP and with MSG as input ,

QMSum_86

false

Yeah.

QMSum_86

false

so I don Well. I don't know.

QMSum_86

false

That means they 're between zero and one.

QMSum_86

false

Mm - hmm.

QMSum_86

false

But i it it it it doesn't necessarily You know , they could be , um Do - doesn't tell you what the variance of the things is.

QMSum_86

false

Mmm. Mm - hmm.

QMSum_86

false

Right ? Cuz if you 're taking the log of these things , it could be , uh Knowing what the sum of the probabilities are , doesn't tell you what the sum of the logs are.

QMSum_86

false

Mm - hmm. Yeah.

QMSum_86

false

So.

QMSum_86

false

Yeah. So we should look at the likelihood , or or what ? Or well , at the log , perhaps , and

QMSum_86

false

Yeah. Yeah.

QMSum_86

false

Mm - hmm.

QMSum_86

false

Or what you know , what you 're uh the thing you 're actually looking at.

QMSum_86

false

Mm - hmm.

QMSum_86

false

So your your the values that are are actually being fed into HTK.

QMSum_86

false

Mm - hmm. But

QMSum_86

false

What do they look like ?

QMSum_86

false

No And so th the , uh for the tandem system , the values that come out of the net don't go through the sigmoid. Right ? They 're sort of the pre - nonlinearity values ?

QMSum_86

false

Yes.

QMSum_86

false

Right. So they 're kinda like log probabilities is what I was saying.

QMSum_86

false

And those OK. And tho that 's what goes into HTK ?

QMSum_86

false

Uh , almost. But then you actually do a KLT on them.

QMSum_86

false

OK.

QMSum_86

false

Um. They aren't normalized after that , are they ?

QMSum_86

false

Mmm. No , they are not no.

QMSum_86

false

No. OK. So , um. Right. So the question is Yeah. Whatever they are at that point , um , are they something for which taking a square root or cube root or fourth root or something like that is is gonna be a good or a bad thing ? So.

QMSum_86

false

Mm - hmm.

QMSum_86

false

Uh , and that 's something that nothing nothing else after that is gonna Uh , things are gonna scale it Uh , you know , subtract things from it , scale it from it , but nothing will have that same effect. Um. So. Um. Anyway , eh

QMSum_86

false

Yeah. Cuz if if the log probs that are coming out of the MSG are really big , the standard insertion penalty is gonna have very little effect

QMSum_86

false

Well , the Right.

QMSum_86

false

compared to , you know , a smaller set of log probs.

QMSum_86

false

Yeah. No. Again you don't really look at that. It 's something that , and then it 's going through this transformation that 's probably pretty close to It 's , eh , whatever the KLT is doing. But it 's probably pretty close to what a a a discrete cosine transformation is doing.

QMSum_86

false

Yeah.

QMSum_86

false

But still it 's it 's not gonna probably radically change the scale of things. I would think. And , uh Yeah. It may be entirely off and and it may be at the very least it may be quite different for MSG than it is for mel cepstrum or PLP. So that would be So the first thing I 'd look at without adjusting anything would just be to go back to the experiment and look at the , uh , substitutions , insertions , and deletions. And if the if the , uh i if there 's a fairly large effect of the difference , say , uh , uh , the r ratio between insertions and deletions for the two cases then that would be , uh , an indicator that it might might be in that direction.

QMSum_86

false

Mm - hmm. Mm - hmm. Yeah. But ,

QMSum_86

false

Anything else ?

QMSum_86

false

my my point was more that it it works sometimes and but sometimes it doesn't work.

QMSum_86

false

Yeah.

QMSum_86

false

So.

QMSum_86

false

Well.

QMSum_86

false

And it works on TI - digits and on SpeechDat - Car it doesn't work , and

QMSum_86

false

Yeah.

QMSum_86

false

Mm - hmm. Yeah. Well.

QMSum_86

false

But , you know , some problems are harder than others ,

QMSum_86

false

Mm - hmm. Yeah.

QMSum_86

false

and And , uh , sometimes , you know , there 's enough evidence for something to work and then it 's harder , it breaks. You know ,

QMSum_86

false

Mm - hmm.

QMSum_86

false

so it 's But it but , um , i it it could be that when you say it works maybe we could be doing much better , even in TI - digits. Right ?

QMSum_86

false

Yeah. Yeah , sure.

QMSum_86

false

So.

QMSum_86

false

Uh.

QMSum_86

false

Hmm ? Yeah.

QMSum_86

false

Yeah. Well , there is also the spectral subtraction , which , um I think maybe we should , uh , try to integrate it in in our system.

QMSum_86

false

Yeah.

QMSum_86

false

Mmm. Mm - hmm.

QMSum_86

false

Right.

QMSum_86

false

But ,

QMSum_86

false

I think that would involve to to mmm use a big a al already a big bunch of the system of Ericsson. Because he has spectral subtraction , then it 's followed by , um , other kind of processing that 's are dependent on the uh , if it 's speech or noi or silence.

QMSum_86

false

Mm - hmm.

QMSum_86

false

And there is this kind of spectral flattening after if it 's silence , and and s I I think it 's important , um , to reduce this musical noise and this this increase of variance during silence portions. So. Well. This was in this would involve to take almost everything from from the this proposal and and then just add some kind of on - line normalization in in the neural network. Mmm.

QMSum_86

false

OK. Well , this 'll be , I think , something for discussion with Hynek next week.

QMSum_86

false

Yeah. Mm - hmm.

QMSum_86

true

Yeah. OK. Right. So. How are , uh , uh how are things going with what you 're doing ?

QMSum_86

false

Oh. Well , um , I took a lot of time just getting my taxes out of the way multi - national taxes. So , I 'm I 'm starting to write code now for my work but I don't have any results yet. Um , i it would be good for me to talk to Hynek , I think , when he 's here.

QMSum_86

false

Yeah.

QMSum_86

false

Do you know what his schedule will be like ?

QMSum_86

false

Uh , he 'll be around for three days.

QMSum_86

false

OK. So , y

QMSum_86

false

Uh , we 'll have a lot of time.

QMSum_86

false

OK.

QMSum_86

false

So , uh Um. I 'll , uh You know , he 's he 'll he 'll be talking with everybody in this room So.

QMSum_86

false

But you said you won't you won't be here next Thursday ?

QMSum_86

false

Not Thursday and Friday. Yeah. Cuz I will be at faculty retreat.

QMSum_86

false

Hmm.

QMSum_86

false

So. I 'll try to connect with him and people as as I can on on Wednesday. But Um. Oh , how 'd taxes go ? Taxes go OK ?

QMSum_86

false

Mmm. Yeah.

QMSum_86

false

Yeah. Oh , good. Yeah. Yeah. That 's just that 's that 's one of the big advantages of not making much money is the taxes are easier. Yeah.

QMSum_86

false

Unless you 're getting money in two countries.

QMSum_86