Make fe_process_utt() allocate a LOT less memory, and sanitize the
internal APIs a bit further.
Add a declaration for what will become the new recommended function
for processing cepstra. It doesn't allocate any memory at all.
Remove other brain damage here and there.
The new code processes the last frame a bit differently. I believe
that it's actually correct and the old code was wrong. In any case
the difference is quite small.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7657 94700074-3cef-4d97-a70e-9c8c206c02f5
make the fe_t structure opaque, and remove the useless FB_TYPE parameter.
Also, don't allocate temporary buffers on every frame (!!!)
Finally, some minor updates to the Q15 code, which doesn't quite work
right still but will soon.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7644 94700074-3cef-4d97-a70e-9c8c206c02f5
language models (makes life sooo much easier for the decoder).
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7629 94700074-3cef-4d97-a70e-9c8c206c02f5
PocketSphinx, and enable it in libsphinxbase.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7623 94700074-3cef-4d97-a70e-9c8c206c02f5
both closed and open vocabulary models (closed vocab still has tobe
tested)
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7597 94700074-3cef-4d97-a70e-9c8c206c02f5
adding class words and interpolating language models.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7492 94700074-3cef-4d97-a70e-9c8c206c02f5
test case for class LMs, to be completed soon.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7485 94700074-3cef-4d97-a70e-9c8c206c02f5
have to read unigrams into memory anyway).
* Extend the language model APIs to include (non-stateful) backoff
mode calculation.
* Store stuff we'd need to undo unigram weights (but not do that yet)
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7472 94700074-3cef-4d97-a70e-9c8c206c02f5
annoy future generations), and add a plan.
I'm not convinced that the "virtual function" implementation here is
actually going to be necessary, but it will make things easier for the
time being.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7462 94700074-3cef-4d97-a70e-9c8c206c02f5
sphinx3_continuous will not confuse about size
of the structure
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7457 94700074-3cef-4d97-a70e-9c8c206c02f5
stupid this requires some extra caution. And it might have broken
Windows compilation, which I will fix later today.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7431 94700074-3cef-4d97-a70e-9c8c206c02f5
Actually going to base this on the PocketSphinx version rather than
the Sphinx3 version, because the Sphinx3 is overengineered.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7410 94700074-3cef-4d97-a70e-9c8c206c02f5
are never smaller than 256 elements. This seems totally arbitrary but
is necessary to do fast addition of quantized log probabilities.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7368 94700074-3cef-4d97-a70e-9c8c206c02f5
table. This is all still up in the air and requires some serious
experimentation to figure out what is fastest.
git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/sphinxbase@7359 94700074-3cef-4d97-a70e-9c8c206c02f5
dhdfu
nshmyrev
tkharris