Decoding with intermediary acoustic models

[Horia Cucu]

Hi,

The AM tutorial states that "At the end of any stage you may use the models
for recognition", but the models I have in the folders:
spk06_rec00-04.cd_cont_1000_1
spk06_rec00-04.cd_cont_1000_2
*spk06_rec00-04.cd_cont_1000_4

lack these files:
feat.params
mdef
*noisedict

which are only found in the folders:
spk06_rec00-04.ci_cont
spk06_rec00-04.cd_cont_1000
*spk06_rec00-04.cd_cont_untied

So... is it possible when training 1000 senones with 8 gaussian mixtures to
decode with the model that has 1000 senones with 4 gaussian mixtures (obtained
before the last splitting and training phase)?

And a more general question:
Are there any other documentation resources except the tutorials (http://cmus
phinx.sourceforge.net/wiki/)? How
can I learn more about Sphinx after I went through all these tutorials?
I'm asking this because I find myself asking all sorts of (maybe) trivial
questions on this forum and I guess you don't have time to answer everybody's
newbie questions...

Yet another more general question:
Is there a way to search for a full phrase like "using intermediary acoustic
models" on this forum? If I try searching for this phrase I'm getting a lot of
results that contain only a subset of these words... I'm trying to avoid
posing questions that have already been answered, but I'm just not able to
find the appropriate topics.

Sorry for wasting your time,
Horia

[Nickolay V. Shmyrev]

So... is it possible when training 1000 senones with 8 gaussian mixtures to
decode with the model that has 1000 senones with 4 gaussian mixtures (obtained
before the last splitting and training phase)?

Yes, you can modify decoding script for that.

I'm asking this because I find myself asking all sorts of (maybe) trivial
questions on this forum and I guess you don't have time to answer everybody's
newbie questions...

It depends on the direction you'd like to pursue. If you want to study the
code or algorithms. You can ask on forum first.

Is there a way to search for a full phrase like "using intermediary acoustic
models" on this forum? If I try searching for this phrase I'm getting a lot of
results that contain only a subset of these words... I'm trying to avoid
posing questions that have already been answered, but I'm just not able to
find the appropriate topics

This is a question to sourceforge support. I'm not happy with the forum
functionality and search either.

[Horia Cucu]

So... is it possible when training 1000 senones with 8 gaussian mixtures
to decode with the model that has 1000 senones with 4 gaussian mixtures
(obtained before the last splitting and training phase)?

Yes, you can modify decoding script for that.

Well, I've modified it:

$DEC_CFG_MODEL_NAME =

"$CFG_EXPTNAME.cd_${CFG_DIRLABEL}${CFG_N_TIED_STATES}";
$DEC_CFG_MODEL_NAME = "$CFG_EXPTNAME.cd${CFG_DIRLABEL}_1000_2";

And the missing files mentioned above are needed:
cucu@argos:~/projectHoriaRoot/project/spk06_rec00-04$
./scripts_pl/decode/slave.pl
MODULE: DECODE Decoding using models previously trained
Decoding 150 segments starting at 0 (part 1 of 1)
0% FATAL_ERROR: "mdef.c", line 679: No mdef-file

Should I just copy them from one folder to another?

[Nickolay V. Shmyrev]

Well, I've modified it:

Decoding script is ./scripts_pl/decode/s3decode.pl or psdecode.pl perl script
where mdef and hmm path are specified. You need to add a separate mdef
argument to sphinx3 or pocketsphinx arguments to that perl script.