Include LDA/MLLT feature adaptation

[Anurag Jain]

Hi all,

I have done acoustic model adaptation without LDA/MLLT but I also want to
include LDA/MLLT feature adaptation.
Can anyone tell me how to do that?
Any help is highly appreciated.

Regards
Anurag

[Pranav Jawale]

try http://cmusphinx.sourceforge.net/wiki/ldamllt

[Anurag Jain]

Hi pranavj,

I am not training acoustic model, rather I am doing acoustic model adaptation.
I want to include lda/mllt during adaptation itself.
The tutorial you pointed out explains lda/mllt adaptation during training
whole of acoustic model.
Please tell me if its possible to include lda/mllt adaptation during acoustic
model adaptation.

Presently I am using these three commands for acoustic model adaptation

./bw -hmmdir hub4opensrc.cd_continuous_8gau/ -moddeffn hub4opensrc.cd_continuous_8gau/mdef -ts2cbfn .cont. -cmn current -agc none -dictfn mydic.dic -ctlfn train.fileids -lsnfn train.transcription -accumdir .





./mllr_solve -meanfn ./hub4opensrc.cd_continuous_8gau/means -varfn ./hub4opensrc.cd_continuous_8gau/variances -outmllrfn mllr_matrix -accumdir .





./map_adapt -meanfn hub4opensrc.cd_continuous_8gau/means -varfn hub4opensrc.cd_continuous_8gau/variances -mixwfn hub4opensrc.cd_continuous_8gau/mixture_weights -tmatfn hub4opensrc.cd_continuous_8gau/transition_matrices -accumdir . -mapmeanfn hub4opensrc.cd_continuous_8gau/means -mapvarfn hub4opensrc.cd_continuous_8gau/variances -mapmixwfn hub4opensrc.cd_continuous_8gau/mixture_weights -maptmatfn hub4opensrc.cd_continuous_8gau/transition_matrices

Regards
Anurag

[Pranav Jawale]

Sorry, no idea.

I suppose in map_adapt you give original models + {audio + transcription} of
data to be adapted to. Is there any other parameter to map_adapt which will
accommodate mllt ?

[Anurag Jain]

Hi pranavj,

I suppose in map_adapt you give original models + {audio + transcription} of data to be adapted to

yes

Is there any other parameter  to map_adapt which will accommodate mllt ?

no, but there is -lda parameter in bw, but it asks for file containing
transformation matrix to be applied to features (single-stream features only)
and I am not understanding how can I generate this file.

Anyway thanks for your reply

Regards
Anurag

[Pranav Jawale]

Hi,

(May be I'm not getting what you need but) why don't you try training with
RunAll.pl as given in http://cmusphinx.sourceforge.net/wiki/tutorialam

and in the sphinx_train.cfg file give the LDA parameter as

$CFG_LDA_MLLT = 'yes';
$CFG_LDA_DIMENSION = 29;

as given in the lda-mllt tutorial.

[Anurag Jain]

Hi,
Acoustic model training and acoustic model adaptation are two different
things.
sphinx_train.cfg is a part of training, it is not used at all while adapting
the acoustic model.
As due to lack of database, I cannot train acoustic model, I am just adapting
a previous acoustic model.
see this null

Regards
Anurag

[Anurag Jain]

http://cmusphinx.sourceforge.net/wiki/tutorialadapt

Regards
Anurag

[Pranav Jawale]

Ok ..

I got confused because they asked to set that LDA param from sphinx_train.cfg
(which is used only during training).

[Nickolay V. Shmyrev]

You still need to explain better what do you want to accomplish. There are
many variants here:

1. You want to adapt an acoustic model which already uses feature space transformation and because of that the model folder has the file feature_transform. In that case you need to add -lda model/feature_transform parameter when you invoke bw to count transformation.

2. You wan to modify feature_transform from the model which already uses it. Update of the feature_transform is not supported. It's also not really necessary. You can use MLLR instead.

3. You want to do something else

In all cases your "include LDA/MLLT adaptation" is not really relevant.
Feature space transformation is not an adaptation.

[Anurag Jain]

Hi,
I am adapting HUB4 acoustic model and want to do lda/mllt feature space
transformation.
But its model folder doesn't contain feature_transform file. So this means
that HUB4 model doesn't uses feature space transformation. Can I still
introduce feature space transformation during acoustic model adaptation?

Regards
Anurag

[Nickolay V. Shmyrev]

Can I still introduce feature space transformation during acoustic model
adaptation?

No, you can't

[Anurag Jain]

ok
Thanks for your reply

Regards
Anurag