You only need etc/voxforge.dic and files in model_parameters/voxforge.cd_cont_3000. Total size is 16Mb. You can also convert mdef to binary format with pocketsphinx_mdef_convert to save space.

Unfortunately Günther trained continuous model for accurate recognition, it is too big for mobile. You need to train PTM model yourself or ask Günther to do that. It's enough to train a model from the downloadable voxforge data, accuracy for few words shoudl be enough. You can add german-speechdata-v2 too.

Hi nsh,

thank you very much for  your answer. i will try to follow your instructions , otherwise  i will contact Guenther.

Have a  nice weekend

Hi there!

I am currently preparing the next release of my german model which will include a much larger portion of german-speechdata-v2 besides the voxforge material (all in all >200h of speech!).

I had never heard of PTM models before, I can definitely try to train one of those. Is there a sphinx_train.cfg file available somewhere that I could use as a starting point?

Thanks,

   Guenter

Hello Guenter

To learn about model types you can check

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

to train PTM model you could try to set the following in config:

$CFG_HMM_TYPE  = '.ptm.'; # PocketSphinx (larger data sets)

} elsif ($CFG_HMM_TYPE eq '.ptm.') {

  $CFG_DIRLABEL = 'ptm';

  $CFG_FEATURE = "1s_c_d_dd";

  $CFG_NUM_STREAMS = 3;

  $CFG_SVSPEC = "0-12/13-25/26-38";

  $CFG_INITIAL_NUM_DENSITIES = 128;

  $CFG_FINAL_NUM_DENSITIES = 128;

  die "For phonetically tied models, the initial and final models have the same density"

    if ($CFG_INITIAL_NUM_DENSITIES != $CFG_FINAL_NUM_DENSITIES);

$CFG_N_TIED_STATES = 5000; 

The resulting model will be less accurate but more suitable for mobile. 

US English models is distributed in both PTM and continuous version for example.

nsh, thanks for the quick response! I have built a PTM model using these settings which seems to have worked without any issues, very nice work :)

stats for these models (trained on 200 hours):

cmusphinx cont model: SENTENCE ERROR: 25.4% (1523/6006)   WORD ERROR RATE: 5.7% (4147/73022)
cmusphinx ptm model: SENTENCE ERROR: 24.7% (1481/6006)   WORD ERROR RATE: 7.1% (5199/73022)

Dear Guenter,

THANK YOU VERY MUCH for your PTM Model.

i have  tested it under PocketSphinx / Android and it works great.

and getting  results  nearly  the english PTM Model.