recent kaldi version add a new decoder class of threaded processing,
with a different API than the previous (SingleUtteranceNnet2DecoderThreaded
instead of SingleUtteranceNnet2Decoder).

this commit adds compile-time support to the threaded decoder, by
ifdef'ing all the relevant API changes. future commit should make it
prettier, but for now, in order to avoid everybody updating kaldi source
code, this seems better.

compile the threaded decoder with make CPPFLAGS=-DTHREADED_DECODER.
Signed-off-by: Amit Beka <amit.beka@gmail.com>

Made chunk length configurable, can be increased to e.g. 0.2 to get some speedup when multithreaded BLAS is used

Inverse scale

add .gitignore with makefile output files

LM scaling is intended to control the ratio between the LM strength and
the acoustic model strength *after* decoding, when the best path in the
lattice is generated.
Setting this to a value higher than 1.0 (default) makes the LM more
prominent in the lattice path weighting, so the best path may be more
aligned with the LM, and not with the acoustic model.
Signed-off-by: Amit Beka <amit.beka@gmail.com>

When decoding the input, the output lattice is scaled using the acoustic
scale parameter. When 'inverse-scale' is set to TRUE, the lattice is
re-scaled back (inverse the acoustic scale) before the hypothesis is
extracted.

This behavior is common when playing with different scales after the
decoding phase.
Signed-off-by: Amit Beka <amit.beka@gmail.com>

Signed-off-by: Amit Beka <amit.beka@gmail.com>

Fix for adaptation state being get_property() is called soon after initialization (https://github.com/alumae/gst-kaldi-nnet2-online/issues/1)