I was told that the sphinx trainer needed to be slightly modified to work for pocketsphinx. I was just wondering if anyone could provide me with some info on training for pocketsphinx.
Thanks,
Nate
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Ahh, this is an annoying problem that I had forgotten about. What you can do is run pocketsphinx on your desktop using the models and the -sendumpfn flag pointing to a file that doesn't exist - it will then create a senone dump in the correct format.
In the meantime I'll update SphinxTrain to be able to create the proper format for pocketsphinx.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Okay if you use the latest version of SphinxTrain from Subversion, you can now pass the argument "-pocketsphinx yes" to mk_s2sendump, and it will produce a file compatible with PocketSphinx.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Training for PocketSphinx is much like training for Sphinx2, except that the final step ("make Sphinx2 models") is not used. You can also use the pocketsphinx_mdef_convert tool (from PocketSphinx) to make the model definition file smaller and quicker to load, and the mk_s2sendump tool (from SphinxTrain) to create a "senone dump file" (you should name this "sendump" in the acoustic model directory) that is more compact and faster to load than the normal mixture_weights file. Finally you can also use the kdtree tool (from SphinxTrain) to create a "kdtrees" file in the acoustic model directory which will also speed it up a bit.
We definitely need to write some more documentation about this.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I was told that the sphinx trainer needed to be slightly modified to work for pocketsphinx. I was just wondering if anyone could provide me with some info on training for pocketsphinx.
Thanks,
Nate
Ahh, this is an annoying problem that I had forgotten about. What you can do is run pocketsphinx on your desktop using the models and the -sendumpfn flag pointing to a file that doesn't exist - it will then create a senone dump in the correct format.
In the meantime I'll update SphinxTrain to be able to create the proper format for pocketsphinx.
Okay if you use the latest version of SphinxTrain from Subversion, you can now pass the argument "-pocketsphinx yes" to mk_s2sendump, and it will produce a file compatible with PocketSphinx.
And the SphinxTrain scripts have been updated so that this sendump file will be created automatically in the last stage of training.
Training for PocketSphinx is much like training for Sphinx2, except that the final step ("make Sphinx2 models") is not used. You can also use the pocketsphinx_mdef_convert tool (from PocketSphinx) to make the model definition file smaller and quicker to load, and the mk_s2sendump tool (from SphinxTrain) to create a "senone dump file" (you should name this "sendump" in the acoustic model directory) that is more compact and faster to load than the normal mixture_weights file. Finally you can also use the kdtree tool (from SphinxTrain) to create a "kdtrees" file in the acoustic model directory which will also speed it up a bit.
We definitely need to write some more documentation about this.
I created a sendump file with mk_2s2sendump, but, after loading it, I get the error message: Dump file is incompatible with PocketSphinx...
Could this be the 8-bit vs. 32-bit issue? If so, how do I build the 8-bit file?
CB