Re: [Kaldi-users] Memory requirement for FSTs

[Josef N. <jos...@gm...>]

>> What might be alternatives if one want to further reduce the size of
HCLG.fst?

It is also possible to simulate a kind of 'poor mans' dynamic composition.

You can build the the component FSTs:  H, Cdet(L), G, use openfst (1.3.2+)
to perform the operations needed to prepare CL and G for lookahead
composition.

Then you modify one of the decoder interfaces - gmm-decode-faster or
whatever to accept the three arguments, build a ComposeFst from that and
feed it to the decoder core.

The cascade build chains in kaldi are not really designed to support this,
but you can make it work with a bit of elbow grease.

Accuracy plateau should be the same; it will definitely be slower, and
operate at different beam points to your vanilla builds, but it may be an
interesting path if you foresee yourself hitting a wall.