[pocketsphinx] pocketsphinx training problems

[Luis Uebel]

Hello,

I am trying to build some acoustic models for Brazilian Portuguese. These
models will work with pocketsphinx.
I read sphinx train tutorial (http://www.speech.cs.cmu.edu/sphinxman/scriptma
n1.html) and
follow training steps.

Some questions about training:
1. Is it the correct to use $CFG_HMM_TYPE = '.semi.'; for pocketsphinx training or it is better to train for Sphinx3 and later covert to sphinx2 models?

1. Scripts manage to build acoustic models in the end using .semi, but I had a ERROR when I used pocketsphinx:
   ./pocketsphinx_continuous -hmm $HMM_MODELS -lm $LM_MODELS

INFO: acmod.c(238): Parsed model-specific feature parameters from
./model_parameters/cd_semi_1000/feat.params
INFO: fe_interface.c(288): You are using the internal mechanism to generate
the seed.
INFO: feat.c(848): Initializing feature stream to type: 's2_4x', ceplen=13,
CMN='current', VARNORM='no', AGC='none'
INFO: cmn.c(142): mean= 12.00, mean= 0.0
INFO: mdef.c(520): Reading model definition:
./model_parameters/cd_semi_1000/mdef
FATAL_ERROR: "mdef.c", line 337: Duplicate base phone: e - - - n/a 14 70 71 72
73 74 N

What am I doing wrong?

1. In file ./logdir/01.vector_quantize/an4.kmeans.log, there is the following message:
   INFO: main.c(1464): No mixing weight file given; none written

Where do I specify mixing weight file?

1. In file ./logdir/20.ci_hmm/an4.1.1-1.bw.log, there is the following warning:
   utt> 15185 br-00-spk-0032-0034 353 0 162 143 61 158 6.053964e-11 -8.745028e+00
   -3.086995e+03
   utt> 15186 br-00-spk-0032-0035 426 0 186 165 66 173 6.739281e-11 -1.054593e+01
   -4.492564e+03
   overall> localhost 4338168 (-0) -1.018271e+01 -4.417433e+07
   WARNING: "accum.c", line 626: The following seno never occur in the input data
   30 31 32 33 34

Is this related to error message when I try to run pocketsphinx?
How can I solve this?

Thanks,

Luis Felipe Uebel

[Nickolay V. Shmyrev]

I am trying to build some acoustic models for Brazilian Portuguese. These
models will work with pocketsphinx. I read sphinx train tutorial (http://www.
speech.cs.cmu.edu/sphinxman/scriptman1.html) and follow training steps.

This is an old document we don't recommend you to read. The training tutorial
is here:

http://cmusphinx.sourceforge.net/html/tutorial.html

We also recommend you to use latest sphinxtrain version and try to search
information from cmusphinx website;

http://cmusphinx.sourceforge.net

1. Is it the correct to use $CFG_HMM_TYPE = '.semi.'; for pocketsphinx
   training

Yes

FATAL_ERROR: "mdef.c", line 337: Duplicate base phone: e - - - n/a 14 70 71
72 73 74 N What am I doing wrong?

Your phoneset have duplicate phones. It's better to use phoneset which is not
case sensitive.

Where do I specify mixing weight file?

You shouldn't specify any file

WARNING: "accum.c", line 626: The following seno never occur in the input
data 30 31 32 33 34 Is this related to error message when I try to run
pocketsphinx?

No

How can I solve this?

Add more data to contain mentioned senones

[Luis Uebel]

Nickolay,

Thanks for answer my questions.
I believe that problem is related to case sensitive phonemes that I am using.
I will try to change this.
In the end of this message there are phoneset used.

These numbers "WARNING: "accum.c", line 626: The following seno never occur in
the input data 30 31 32 33 34 "
are related to number of phoneme in phoneset?

What are the mininum number of samples to robust estimate a model? 400
samples?

I think that problem is not related to data since I train without problems
acoustic models for HTK and using without problems with Julius.

Thanks,

Luis

---

1 a
2 a~
3 b
4 d
5 dZ
6 e
7 e~
8 E
9 f
10 g
11 i
12 i~
13 j
14 j~
15 J
16 k
17 l
18 L
19 m
20 n
21 o
22 o~
23 O
24 p
25 r
26 R
27 s
28 S
29 SIL
30 t
31 tS
32 u
33 u~
34 v
35 w
36 w~
37 X
38 z
39 Z

[Nickolay V. Shmyrev]

These numbers "WARNING: "accum.c", line 626: The following seno never occur
in the input data 30 31 32 33 34 " are related to number of phoneme in
phoneset?

Those numbers are numbers of senones (GMMs in the model). You can find the
mapping between senones and phones in mdef file in model_architecture folder.

What are the mininum number of samples to robust estimate a model? 400
samples?

There is no minimum. Warning says you have no samples at all

I think that problem is not related to data since I train without problems
acoustic models for HTK and using without problems with Julius.

Of course the problem is in e and E in your phoneset, that's what warning said
you before.