Dumping bigram & trigram LM after lo...

[Joseph S. Wisniewski]

Obviously, since the first two search passes are a bigram and trigram search,
if you give Pocketsphinx an FSG or JSFG grammar, it must immediately generate
the appropriate bigram and trigram language models. Is there any way to have
it dump those to console or file immediately after generating them?

Thanks.

[Nickolay V. Shmyrev]

Obviously, since the first two search passes are a bigram and trigram search

Are you sure it's true? I don't think so if you are talking about
pocketsphinx. Both are using trigram scores but fwdtree uses approximate
single lextree search so that only one trigram history per state survives.

you give Pocketsphinx an FSG or JSFG grammar, it must immediately generate
the appropriate bigram and trigram language models.

FSG search is separate, it uses FSG as a search space it doesn't convert it to
n-gram.

[Joseph S. Wisniewski]

Sometimes, it's hard to be sure of anything in Sphinx ;)

But I'm going by the Hieroglyphs and the architecture diagrams in David
Huggins Daines "An Architecture for Scalable, Universal Speech Recognition",
which I find to be the most comprehensible piece of Sphinx documentation.

[Joseph S. Wisniewski]

And I'm not at all sure about the "poor man's trigram" part. I'd settle for a
bigram. ;)

[Nickolay V. Shmyrev]

But I'm going by the Hieroglyphs and the architecture diagrams in David
Huggins Daines "An Architecture for Scalable, Universal Speech Recognition",
which I find to be the most comprehensible piece of Sphinx documentation.

I would drop out Hieroglyphs, it's not finished. Instead, I would recommend to
start with Ravishankar Mosur thesis "Efficient Algorithms for Speech
Recognition", then go to David's thesis.