I train a mllr matrix and wish to adapt on the same model (hub4wsj_sc_8k) in
this tutorial. No matter I use -mllr or not, it has identical result which I
think is strange. Even if the model change not much on mllr, it should have
some difference? In the log there are also lines indicate that it loads the
mllr:
I tried to modify the mllr matrix, make some value in it become 999999(So it
"should" have effects on the model, since this modification is huge), but the
result is still the same. Does anyone has idea about it?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
May I ask the reason why it has no effect? I check the code and I see it
modifies the mean and variance for the Gaussians, I know semi-continuous HMM
uses the same set of Gaussians with different weight for different state. If
we modified the Gaussian itself, it should at least change the hypothesis
score? I don't know why the result is exactly the same.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Semi-continuous models are built from a very small amount of gaussians (256 *
3) which are combined with different mixture weights to get senones. No matter
how you change gaussians it's important how do you mix them not their values.
That's why mixture weights adaptation is more important than MLLR
Because MLLR doesn't make sense it's not even implemented in s2_semi_mgau.c.
Only generic continuous model computation in ms_mgau actually uses MLLR data.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
So the Semi-continuous HMM did not use ms_mgau.c？Cause the log file I posted
on the first post indicates it loaded MLLR Matrix although I am using semi-
continuous model. So the code still call ms_mgau although it only use the
information from s2_semi_mgau? Is that the case?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
All of there are stored in the acmod, but where did other codes uses these
values? Maybe it's because the mean and variance in modified by MLLR after
they get these values, so it makes no difference whether we use MLLR or not...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
According to http://cmusphinx.sourceforge.net/wiki/tutorialadapt
I train a mllr matrix and wish to adapt on the same model (hub4wsj_sc_8k) in
this tutorial. No matter I use -mllr or not, it has identical result which I
think is strange. Even if the model change not much on mllr, it should have
some difference? In the log there are also lines indicate that it loads the
mllr:
INFO: ps_mllr.c(66): Reading MLLR transformation file
/net/dogwood/usr1/jchiu1/pocketsphinx/Justin_hub_mllr
INFO: ms_gauden.c(198): Reading mixture gaussian parameter:
/net/dogwood/usr1/jchiu1/pocketsphinx/hub4wsj_sc_8k/means
INFO: ms_gauden.c(292): 1 codebook, 3 feature, size:
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(198): Reading mixture gaussian parameter:
/net/dogwood/usr1/jchiu1/pocketsphinx/hub4wsj_sc_8k/variances
INFO: ms_gauden.c(292): 1 codebook, 3 feature, size:
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(294): 256x13
INFO: ms_gauden.c(354): 0 variance values floored
I tried to modify the mllr matrix, make some value in it become 999999(So it
"should" have effects on the model, since this modification is huge), but the
result is still the same. Does anyone has idea about it?
MLLR has no effect with semi-continuous models. You can try continuous model
to test it.
May I ask the reason why it has no effect? I check the code and I see it
modifies the mean and variance for the Gaussians, I know semi-continuous HMM
uses the same set of Gaussians with different weight for different state. If
we modified the Gaussian itself, it should at least change the hypothesis
score? I don't know why the result is exactly the same.
Hello
Semi-continuous models are built from a very small amount of gaussians (256 *
3) which are combined with different mixture weights to get senones. No matter
how you change gaussians it's important how do you mix them not their values.
That's why mixture weights adaptation is more important than MLLR
Because MLLR doesn't make sense it's not even implemented in s2_semi_mgau.c.
Only generic continuous model computation in ms_mgau actually uses MLLR data.
So the Semi-continuous HMM did not use ms_mgau.c？Cause the log file I posted
on the first post indicates it loaded MLLR Matrix although I am using semi-
continuous model. So the code still call ms_mgau although it only use the
information from s2_semi_mgau? Is that the case?
Hmm, I've check the s2_semi_mgau.c
Now I am thinking that they stored some value like
s->topn_beam
s->max_topn
s->n_topn_hist
s->topn_hist
All of there are stored in the acmod, but where did other codes uses these
values? Maybe it's because the mean and variance in modified by MLLR after
they get these values, so it makes no difference whether we use MLLR or not...