elplaguister/DTS_line_datasets · Datasets at Hugging Face

topic_shift bool 2 classes	utterance stringlengths 1 7.9k	session_id stringlengths 7 14
false	It 's noisy TIMIT.	QMSum_120
false	Yeah.	QMSum_120
false	It seems quite robust to noise , so when we take we draw its parameters across time for a clean sentence and then nois the same noisy sentence , it 's very close.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Yeah. So there are there is this. There could be also the , um something like the maximum of the auto - correlation function or which	QMSum_120
false	Is this a a s a trained system ? Or is it a system where you just pick some thresholds ? Ho - how does it work ?	QMSum_120
false	Right now we just are trying to find some features. And ,	QMSum_120
false	Mm - hmm.	QMSum_120
false	uh Yeah. Hopefully , I think what we want to have is to put these features in s some kind of , um well , to to obtain a statistical model on these features and to or just to use a neural network and hopefully these features w would help	QMSum_120
false	Because it seems like what you said about the mean of the the voiced and the unvoiced that seemed pretty encouraging.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Well , yeah , except the variance was big.	QMSum_120
false	Right ?	QMSum_120
false	Yeah. Except the variance is quite high.	QMSum_120
false	Right ?	QMSum_120
false	Well , y	QMSum_120
false	Yeah.	QMSum_120
false	Well , y I I don't know that I would trust that so much because you 're doing these canonical mappings from TIMIT labellings.	QMSum_120
false	Uh - huh.	QMSum_120
false	Right ? So , really that 's sort of a cartoon picture about what 's voiced and unvoiced. So that could be giving you a lot of variance.	QMSum_120
false	Yeah.	QMSum_120
false	I mean , i it it may be that that you 're finding something good and that the variance is sort of artificial because of how you 're getting your truth.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Yeah. But another way of looking at it might be that I mean , what w we we are coming up with feature sets after all. So another way of looking at it is that um , the mel cepstru mel spectrum , mel cepstrum , any of these variants , um , give you the smooth spectrum. It 's the spectral envelope. By going back to the FFT , you 're getting something that is more like the raw data. So the question is , what characterization and you 're playing around with this another way of looking at it is what characterization of the difference between the raw data and this smooth version is something that you 're missing that could help ? So , I mean , looking at different statistical measures of that difference , coming up with some things and just trying them out and seeing if you add them onto the feature vector does that make things better or worse in noise , where you 're really just i i the way I 'm looking at it is not so much you 're trying to f find the best the world 's best voiced - unvoiced , uh , uh , classifier ,	QMSum_120
false	Mm - hmm.	QMSum_120
false	Mmm.	QMSum_120
false	but it 's more that , you know , uh , uh , try some different statistical characterizations of that difference back to the raw data	QMSum_120
false	Right.	QMSum_120
false	and and m maybe there 's something there that the system can use.	QMSum_120
false	Right.	QMSum_120
false	Yeah. Yeah , but ther more obvious is that Yeah. The the more obvious is that that well , using the th the FFT , um , you just it gives you just information about if it 's voiced or not voiced , ma mainly , I mean. But So ,	QMSum_120
false	Yeah.	QMSum_120
false	this is why we we started to look by having sort of voiced phonemes	QMSum_120
false	Well , that 's the rea w w what I 'm arguing is that 's Yeah. I mean , uh , what I 'm arguing is that that that 's givi you gives you your intuition.	QMSum_120
false	and Mm - hmm.	QMSum_120
false	But in in reality , it 's you know , there 's all of this this overlap and so forth ,	QMSum_120
false	Oh , sorry.	QMSum_120
false	and But what I 'm saying is that may be OK , because what you 're really getting is not actually voiced versus unvoiced , both for the fac the reason of the overlap and and then , uh , th you know , structural reasons , uh , uh , like the one that Chuck said , that that in fact , well , the data itself is that you 're working with is not perfect.	QMSum_120
false	Yeah. Mm - hmm.	QMSum_120
false	So , what I 'm saying is maybe that 's not a killer because you 're just getting some characterization , one that 's driven by your intuition about voiced - unvoiced certainly ,	QMSum_120
false	Mm - hmm.	QMSum_120
false	but it 's just some characterization of something back in the in the in the almost raw data , rather than the smooth version.	QMSum_120
false	Mm - hmm.	QMSum_120
false	And your intuition is driving you towards particular kinds of , uh , statistical characterizations of , um , what 's missing from the spectral envelope.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Um , obviously you have something about the excitation , um , and what is it about the excitation , and , you know and you 're not getting the excitation anyway , you know. So so I I would almost take a uh , especially if if these trainings and so forth are faster , I would almost just take a uh , a scattershot at a few different ways of look of characterizing that difference and , uh , you could have one of them but and and see , you know , which of them helps.	QMSum_120
false	Mm - hmm. OK.	QMSum_120
false	So i is the idea that you 're going to take whatever features you develop and and just add them onto the future vector ? Or , what 's the use of the the voiced - unvoiced detector ?	QMSum_120
false	Uh , I guess we don't know exactly yet. But , um Yeah. Th	QMSum_120
false	It 's not part of a VAD system that you 're doing ?	QMSum_120
false	No.	QMSum_120
false	Uh , no. No.	QMSum_120
false	Oh , OK.	QMSum_120
false	No , the idea was , I guess , to to use them as as features.	QMSum_120
false	Features. I see.	QMSum_120
false	Uh Yeah , it could be , uh it could be a neural network that does voiced and unvoiced detection ,	QMSum_120
false	Mm - hmm.	QMSum_120
false	but it could be in the also the big neural network that does phoneme classification.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Mmm. Yeah.	QMSum_120
false	But each one of the mixture components I mean , you have , uh , uh , variance only , so it 's kind of like you 're just multiplying together these , um , probabilities from the individual features within each mixture. So it 's so , uh , it seems l you know	QMSum_120
false	I think it 's a neat thing. Uh , it seems like a good idea.	QMSum_120
false	Yeah. Um. Yeah. I mean , I know that , um , people doing some robustness things a ways back were were just doing just being gross and just throwing in the FFT and actually it wasn't wasn't wasn't so bad. Uh , so it would s and and you know that i it 's gotta hurt you a little bit to not have a a spectral , uh a s a smooth spectral envelope , so there must be something else that you get in return for that	QMSum_120
false	Mm - hmm.	QMSum_120
false	that , uh uh So.	QMSum_120
false	So how does uh , maybe I 'm going in too much detail , but how exactly do you make the difference between the FFT and the smoothed spectral envelope ? Wha - wh i i uh , how is that , uh ?	QMSum_120
false	Um , we just How did we do it up again ?	QMSum_120
false	Uh , we distend the we have the twenty - three coefficient af after the mel f filter ,	QMSum_120
false	Mm - hmm.	QMSum_120
false	and we extend these coefficient between the all the frequency range.	QMSum_120
false	Mm - hmm.	QMSum_120
false	And i the interpolation i between the point is give for the triang triangular filter , the value of the triangular filter and of this way we obtained this mode this model speech.	QMSum_120
false	So you essentially take the values that th that you get from the triangular filter and extend them to sor sort of like a rectangle , that 's at that m value.	QMSum_120
false	Yeah.	QMSum_120
false	Yeah. I think we have linear interpolation.	QMSum_120
false	Mm - hmm.	QMSum_120
false	So we have we have one point for one energy for each filter bank ,	QMSum_120
false	mmm Yeah , it 's linear.	QMSum_120
false	Mmm.	QMSum_120
false	Oh.	QMSum_120
false	which is the energy that 's centered on on on the triangle	QMSum_120
false	Yeah. At the n at the center of the filter	QMSum_120
false	So you you end up with a vector that 's the same length as the FFT vector ?	QMSum_120
false	Yeah. That 's right.	QMSum_120
false	Yeah.	QMSum_120
false	And then you just , uh , compute differences	QMSum_120
false	Yeah. I have here one example if you if you want see something like that.	QMSum_120
false	Then we compute the difference.	QMSum_120
false	and ,	QMSum_120
false	Yeah. Uh - huh.	QMSum_120
false	OK.	QMSum_120
false	uh , sum the differences ?	QMSum_120
false	So. And I think the variance is computed only from , like , two hundred hertz to one to fifteen hundred.	QMSum_120
false	Oh ! OK.	QMSum_120
false	Mm - hmm.	QMSum_120
false	Two thou two fifteen hundred ?	QMSum_120
false	Mm - hmm.	QMSum_120
false	Because	QMSum_120
false	No.	QMSum_120
false	Right.	QMSum_120

false

It 's noisy TIMIT.

QMSum_120

false

Yeah.

QMSum_120

false

It seems quite robust to noise , so when we take we draw its parameters across time for a clean sentence and then nois the same noisy sentence , it 's very close.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Yeah. So there are there is this. There could be also the , um something like the maximum of the auto - correlation function or which

QMSum_120

false

Is this a a s a trained system ? Or is it a system where you just pick some thresholds ? Ho - how does it work ?

QMSum_120

false

Right now we just are trying to find some features. And ,

QMSum_120

false

Mm - hmm.

QMSum_120

false

uh Yeah. Hopefully , I think what we want to have is to put these features in s some kind of , um well , to to obtain a statistical model on these features and to or just to use a neural network and hopefully these features w would help

QMSum_120

false

Because it seems like what you said about the mean of the the voiced and the unvoiced that seemed pretty encouraging.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Well , yeah , except the variance was big.

QMSum_120

false

Right ?

QMSum_120

false

Yeah. Except the variance is quite high.

QMSum_120

false

Right ?

QMSum_120

false

Well , y

QMSum_120

false

Yeah.

QMSum_120

false

Well , y I I don't know that I would trust that so much because you 're doing these canonical mappings from TIMIT labellings.

QMSum_120

false

Uh - huh.

QMSum_120

false

Right ? So , really that 's sort of a cartoon picture about what 's voiced and unvoiced. So that could be giving you a lot of variance.

QMSum_120

false

Yeah.

QMSum_120

false

I mean , i it it may be that that you 're finding something good and that the variance is sort of artificial because of how you 're getting your truth.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Yeah. But another way of looking at it might be that I mean , what w we we are coming up with feature sets after all. So another way of looking at it is that um , the mel cepstru mel spectrum , mel cepstrum , any of these variants , um , give you the smooth spectrum. It 's the spectral envelope. By going back to the FFT , you 're getting something that is more like the raw data. So the question is , what characterization and you 're playing around with this another way of looking at it is what characterization of the difference between the raw data and this smooth version is something that you 're missing that could help ? So , I mean , looking at different statistical measures of that difference , coming up with some things and just trying them out and seeing if you add them onto the feature vector does that make things better or worse in noise , where you 're really just i i the way I 'm looking at it is not so much you 're trying to f find the best the world 's best voiced - unvoiced , uh , uh , classifier ,

QMSum_120

false

Mm - hmm.

QMSum_120

false

Mmm.

QMSum_120

false

but it 's more that , you know , uh , uh , try some different statistical characterizations of that difference back to the raw data

QMSum_120

false

Right.

QMSum_120

false

and and m maybe there 's something there that the system can use.

QMSum_120

false

Right.

QMSum_120

false

Yeah. Yeah , but ther more obvious is that Yeah. The the more obvious is that that well , using the th the FFT , um , you just it gives you just information about if it 's voiced or not voiced , ma mainly , I mean. But So ,

QMSum_120

false

Yeah.

QMSum_120

false

this is why we we started to look by having sort of voiced phonemes

QMSum_120

false

Well , that 's the rea w w what I 'm arguing is that 's Yeah. I mean , uh , what I 'm arguing is that that that 's givi you gives you your intuition.

QMSum_120

false

and Mm - hmm.

QMSum_120

false

But in in reality , it 's you know , there 's all of this this overlap and so forth ,

QMSum_120

false

Oh , sorry.

QMSum_120

false

and But what I 'm saying is that may be OK , because what you 're really getting is not actually voiced versus unvoiced , both for the fac the reason of the overlap and and then , uh , th you know , structural reasons , uh , uh , like the one that Chuck said , that that in fact , well , the data itself is that you 're working with is not perfect.

QMSum_120

false

Yeah. Mm - hmm.

QMSum_120

false

So , what I 'm saying is maybe that 's not a killer because you 're just getting some characterization , one that 's driven by your intuition about voiced - unvoiced certainly ,

QMSum_120

false

Mm - hmm.

QMSum_120

false

but it 's just some characterization of something back in the in the in the almost raw data , rather than the smooth version.

QMSum_120

false

Mm - hmm.

QMSum_120

false

And your intuition is driving you towards particular kinds of , uh , statistical characterizations of , um , what 's missing from the spectral envelope.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Um , obviously you have something about the excitation , um , and what is it about the excitation , and , you know and you 're not getting the excitation anyway , you know. So so I I would almost take a uh , especially if if these trainings and so forth are faster , I would almost just take a uh , a scattershot at a few different ways of look of characterizing that difference and , uh , you could have one of them but and and see , you know , which of them helps.

QMSum_120

false

Mm - hmm. OK.

QMSum_120

false

So i is the idea that you 're going to take whatever features you develop and and just add them onto the future vector ? Or , what 's the use of the the voiced - unvoiced detector ?

QMSum_120

false

Uh , I guess we don't know exactly yet. But , um Yeah. Th

QMSum_120

false

It 's not part of a VAD system that you 're doing ?

QMSum_120

false

No.

QMSum_120

false

Uh , no. No.

QMSum_120

false

Oh , OK.

QMSum_120

false

No , the idea was , I guess , to to use them as as features.

QMSum_120

false

Features. I see.

QMSum_120

false

Uh Yeah , it could be , uh it could be a neural network that does voiced and unvoiced detection ,

QMSum_120

false

Mm - hmm.

QMSum_120

false

but it could be in the also the big neural network that does phoneme classification.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Mmm. Yeah.

QMSum_120

false

But each one of the mixture components I mean , you have , uh , uh , variance only , so it 's kind of like you 're just multiplying together these , um , probabilities from the individual features within each mixture. So it 's so , uh , it seems l you know

QMSum_120

false

I think it 's a neat thing. Uh , it seems like a good idea.

QMSum_120

false

Yeah. Um. Yeah. I mean , I know that , um , people doing some robustness things a ways back were were just doing just being gross and just throwing in the FFT and actually it wasn't wasn't wasn't so bad. Uh , so it would s and and you know that i it 's gotta hurt you a little bit to not have a a spectral , uh a s a smooth spectral envelope , so there must be something else that you get in return for that

QMSum_120

false

Mm - hmm.

QMSum_120

false

that , uh uh So.

QMSum_120

false

So how does uh , maybe I 'm going in too much detail , but how exactly do you make the difference between the FFT and the smoothed spectral envelope ? Wha - wh i i uh , how is that , uh ?

QMSum_120

false

Um , we just How did we do it up again ?

QMSum_120

false

Uh , we distend the we have the twenty - three coefficient af after the mel f filter ,

QMSum_120

false

Mm - hmm.

QMSum_120

false

and we extend these coefficient between the all the frequency range.

QMSum_120

false

Mm - hmm.

QMSum_120

false

And i the interpolation i between the point is give for the triang triangular filter , the value of the triangular filter and of this way we obtained this mode this model speech.

QMSum_120

false

So you essentially take the values that th that you get from the triangular filter and extend them to sor sort of like a rectangle , that 's at that m value.

QMSum_120

false

Yeah.

QMSum_120

false

Yeah. I think we have linear interpolation.

QMSum_120

false

Mm - hmm.

QMSum_120

false

So we have we have one point for one energy for each filter bank ,

QMSum_120

false

mmm Yeah , it 's linear.

QMSum_120

false

Mmm.

QMSum_120

false

Oh.

QMSum_120

false

which is the energy that 's centered on on on the triangle

QMSum_120

false

Yeah. At the n at the center of the filter

QMSum_120

false

So you you end up with a vector that 's the same length as the FFT vector ?

QMSum_120

false

Yeah. That 's right.

QMSum_120

false

Yeah.

QMSum_120

false

And then you just , uh , compute differences

QMSum_120

false

Yeah. I have here one example if you if you want see something like that.

QMSum_120

false

Then we compute the difference.

QMSum_120

false

and ,

QMSum_120

false

Yeah. Uh - huh.

QMSum_120

false

OK.

QMSum_120

false

uh , sum the differences ?

QMSum_120

false

So. And I think the variance is computed only from , like , two hundred hertz to one to fifteen hundred.

QMSum_120

false

Oh ! OK.

QMSum_120

false

Mm - hmm.

QMSum_120

false

Two thou two fifteen hundred ?

QMSum_120

false

Mm - hmm.

QMSum_120

false

Because

QMSum_120

false

No.

QMSum_120

false

Right.

QMSum_120