Problems with custom .lm with sphinx4/Java

[Jake Nix]

Hello,

I'm really new to all this and have tried to read/search as much as possible
before I posted with this problem but I'm stuck.

I'm trying to create a very simple voice activated radio as a personal
project. I followed the tutorial with the web suggestion in creating your own
language model to create my own corpus.txt using the
lmtool. Got the .dic and .lm
files and when I pass those to pocketsphinx_continuous, I get great results -
no problems, very accurate.

However, I want to code this Java and tried following the docs on the wiki for
sphinx4 dev. I've modified the HelloWorld demo config slightly to accept the
new language model and dictionary file.

Problem: the recognition of utterances is significantly lower (borderline
crap) when using the Java app - I assume this is a problem with the XML
config, but I have no idea where to start.

I had a try and read through the
tutorial, specifically
the sections mentioning the vocab size - this is very small so I changed the
variables accordingly to "Small (80 words - AN4)" in the "frequently tuned
properties" section of the config, but still no luck :(

My grammar file is:

jake@monkey ~/Code/Java/RadioPi/bin $ cat radio.gram 
#JSGF V1.0;

grammar radio;

public <command> = ( <radio> | <station> | <volume> );

<radio> = radio ( on | off );
<station> =  ( the bay | radio two | jazz f m | planet rock | radio lancashire );
<volume> = volume ( up | down | mute );

Any help would be really appreciated!

Thanks :)

[Jake Nix]

Sorry this should have gone in the Sphinx4 Help section - apologies.

[Nickolay V. Shmyrev]

Please use Sphinx4 help to ask for help

Please provide samples of the audio you are trying to recognize

Please provide the modifications you have made, you can pack everything into
archive and upload to a public file sharing resource.