Restricting Language Model to Full Sentences

[P Laurens]

Hi,

I have a specific set of commands (one to three words) which I'd like to be
able to recognise. Although the online language-model tool is great for
producing n-grams from the provided sentences, mixing and matching words from
each line, I don't actually want a model which does this.

I'd rather have the language model strictly choose from the commands which I
list in the corpus, and not match different words from different commands.

Is there an easy way to modify the language model (perhaps removing or
manually adding n-grams) in order to achieve this restriction?

Many thanks in advance,

• P

[Nickolay V. Shmyrev]

You probably want to use a grammar in JSGF format instead of language model,
don't you? Both pocketsphinx and sphinx4 support JSGF.

[P Laurens]

Hi,

A quick Google shows that JSGF sounds like a good option. I'm only familiar
with the process of producing language models and dic files through the online
tool though. Do you have any pointers on tools for generating a JSGF suitable
for PocketSphinx though?

Thanks,

• P

[Nickolay V. Shmyrev]

1. Write JSGF according to specification

http://java.sun.com/products/java-
media/speech/forDevelopers/JSGF/

1. Use it in pocketsphinx with -jsgf option. Just

   pocketsphinx_continuous -jsgf your.jsgf

isntead of

pocketsphinx_continuous -lm your.lm

1. For some advanced example see grammars at

http://wiki.freeswitch.org/wiki/Mod_pocketsphinx

[P Laurens]

Oh that's great, thanks!

[P Laurens]

One additional question - when I've generated language models from a corpus
using the web tool, I get a dictionary (.dic) file. Do I use the same .dic
file with the jsgf model as I did with the lm model?

Thanks again.

[Nickolay V. Shmyrev]

yes