topic_shift
bool 2
classes | utterance
stringlengths 1
7.9k
| session_id
stringlengths 7
14
|
|---|---|---|
false |
Two hundred and fifty thousand.
|
QMSum_120
|
false |
Fifteen hundred. Because Yeah.
|
QMSum_120
|
false |
Yeah. Two thousand and fifteen hundred.
|
QMSum_120
|
false |
Above , um it seems that Well , some voiced sound can have also , like , a noisy part on high frequencies , and But
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Well , it 's just
|
QMSum_120
|
false |
No , it 's makes sense to look at low frequencies.
|
QMSum_120
|
false |
So this is uh , basically this is comparing an original version of the signal to a smoothed version of the same signal ?
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Right. So i so i i this is I mean , i you could argue about whether it should be linear interpolation or or or or zeroeth order , but but
|
QMSum_120
|
false |
Uh - huh.
|
QMSum_120
|
false |
at any rate something like this is what you 're feeding your recognizer , typically.
|
QMSum_120
|
false |
Like which of the ?
|
QMSum_120
|
false |
No. Uh , so the mel cepstrum is the is the is the cepstrum of this this , uh , spectrum or log spectrum ,
|
QMSum_120
|
false |
So this is Yeah.
|
QMSum_120
|
false |
Yeah. Right , right.
|
QMSum_120
|
false |
whatever it You - you 're subtracting in in in power domain or log domain ?
|
QMSum_120
|
false |
In log domain. Yeah.
|
QMSum_120
|
false |
Log domain.
|
QMSum_120
|
false |
OK. So it 's sort of like division , when you do the yeah , the spectra.
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Uh , yeah.
|
QMSum_120
|
false |
It 's the ratio.
|
QMSum_120
|
false |
Um. Yeah. But , anyway , um and that 's
|
QMSum_120
|
false |
So what 's th uh , what 's the intuition behind this kind of a thing ? I I don't know really know the signal - processing well enough to understand what what is that doing.
|
QMSum_120
|
false |
So. Yeah. What happen if what we have have what we would like to have is some spectrum of the excitation signal ,
|
QMSum_120
|
false |
Yeah. I guess that makes sense. Yeah.
|
QMSum_120
|
false |
which is for voiced sound ideally a a pulse train
|
QMSum_120
|
false |
Uh - huh.
|
QMSum_120
|
false |
and for unvoiced it 's something that 's more flat.
|
QMSum_120
|
false |
Uh - huh. Right.
|
QMSum_120
|
false |
And the way to do this is that well , we have the we have the FFT because it 's computed in in the in the system , and we have the mel filter banks ,
|
QMSum_120
|
false |
Mm - hmm. Mm - hmm.
|
QMSum_120
|
false |
and so if we if we , like , remove the mel filter bank from the FFT , we have something that 's close to the excitation signal.
|
QMSum_120
|
false |
Oh.
|
QMSum_120
|
false |
It 's something that 's like a a a train of p a pulse train for voiced sound
|
QMSum_120
|
false |
OK.
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Oh ! OK. Yeah.
|
QMSum_120
|
false |
and that 's that should be flat for
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
I see. So do you have a picture that sh ?
|
QMSum_120
|
false |
So - It 's Y
|
QMSum_120
|
false |
Is this for a voiced segment ,
|
QMSum_120
|
false |
yeah.
|
QMSum_120
|
false |
this picture ? What does it look like for unvoiced ?
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
You have several some unvoiced ?
|
QMSum_120
|
false |
The dif No. Unvoiced , I don't have
|
QMSum_120
|
false |
Oh.
|
QMSum_120
|
false |
for unvoiced.
|
QMSum_120
|
false |
Yeah. So , you know , all
|
QMSum_120
|
false |
I 'm sorry.
|
QMSum_120
|
false |
But Yeah.
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Yeah. This is the between
|
QMSum_120
|
false |
This is another voiced example. Yeah.
|
QMSum_120
|
false |
No. But it 's this ,
|
QMSum_120
|
false |
Oh , yeah. This is
|
QMSum_120
|
false |
but between the frequency that we are considered for the excitation
|
QMSum_120
|
false |
Right. Mm - hmm.
|
QMSum_120
|
false |
for the difference and this is the difference.
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
This is the difference. OK.
|
QMSum_120
|
false |
So , of course , it 's around zero ,
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
Sure looks
|
QMSum_120
|
false |
but
|
QMSum_120
|
false |
Hmm.
|
QMSum_120
|
false |
Well , no.
|
QMSum_120
|
false |
Hmm.
|
QMSum_120
|
false |
It is
|
QMSum_120
|
false |
Yeah. Because we begin , uh , in fifteen point the fifteen point.
|
QMSum_120
|
false |
So , does does the periodicity of this signal say something about the the
|
QMSum_120
|
false |
Fifteen p
|
QMSum_120
|
false |
So it 's Yeah.
|
QMSum_120
|
false |
Pitch.
|
QMSum_120
|
false |
It 's the pitch.
|
QMSum_120
|
false |
the pitch ?
|
QMSum_120
|
false |
Yeah. Mm - hmm.
|
QMSum_120
|
false |
Yeah.
|
QMSum_120
|
false |
OK.
|
QMSum_120
|
false |
That 's like fundamental frequency.
|
QMSum_120
|
false |
Mm - hmm.
|
QMSum_120
|
false |
So , I mean , i t t
|
QMSum_120
|
false |
OK. I see.
|
QMSum_120
|
false |
I mean , to first order what you 'd what you 're doing I mean , ignore all the details and all the ways which is that these are complete lies. Uh , the the you know , what you 're doing in feature extraction for speech recognition is you have , uh , in your head a a a a simplified production model for speech ,
|
QMSum_120
|
false |
Mm - hmm.
|
QMSum_120
|
false |
in which you have a periodic or aperiodic source that 's driving some filters.
|
QMSum_120
|
false |
Mm - hmm.
|
QMSum_120
|
false |
Yeah. This is the the auto - correlation the R - zero energy.
|
QMSum_120
|
false |
Do you have the mean do you have the mean for the auto - correlation ?
|
QMSum_120
|
false |
Uh , first order for speech recognition , you say " I don't care about the source ".
|
QMSum_120
|
false |
For Yeah.
|
QMSum_120
|
false |
Well , I mean for the the energy.
|
QMSum_120
|
false |
I have the mean.
|
QMSum_120
|
false |
Right ?
|
QMSum_120
|
false |
Right.
|
QMSum_120
|
false |
And so you just want to find out what the filters are.
|
QMSum_120
|
false |
Right.
|
QMSum_120
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.