Different results on windows cmd and on Android demo (Pocketsphinx)

[Anais Ramirez]

Hello, I am currently developing a game using pocketsphinx and I need words to be spelled in the game.

I am using Unity and I based my solution on Android Demo but I am having problems with the use of grams.

I ran pocketsphinx from my Windows cmd using a dictionary that only contains the letters of the alphabet, and as you can see the result, is not always perfect (sometimes it confuses letters like b and d), but it's very clean. On the other hand, I did the same on android using the same dictionary and the same "en-us" language model using the same "feat.params" and with the same background noise. The only difference is that here I used a gram and it seems to be printing every single noise it hears and matching it to the gram options.

I need my android game to print a clean result as on my Windows cmd and I am guessing that the gram is the one messing with with.
I checked the android tutorial on the CMUSphinx page but even though it says that the use of grammars is optional I always have to give the StartListening method something to search:
recognizer.startListening(searchName);
And if I don't set addKeyphraseSearch, addGrammarSearch, or addNgramSearch it crashes.

Is it possible to have a cleaner result as on my windows cmd and if it is, what are the steps to achieve it?

[Nickolay V. Shmyrev]

using a dictionary that only contains the letters of the alphabet

The recommended way is to restrict the language model vocabulary instead

same "en-us" language model using the same "feat.params"

It is acoustic model, not language model

here I used a gram

You can try to use the same language model on Android as you used on Windows.

[Anais Ramirez]

Thanks for the quick response!
You're right I was confusing acoustic model with language model, it worked. Still, the accuracy is quite low for spelling, I'll try your suggestion and restrict the LM instead of the dictionary