Initial version of kaldi rescorer.

senarvi / theanolm

TheanoLM is a recurrent neural network language modeling tool implemented using Theano

Apache License 2.0

81 stars 29 forks source link

Initial version of kaldi rescorer. #29

Closed psmit closed 6 years ago

psmit commented 7 years ago

In principle this can be merged, but it would need some checking from you.

In a couple place there are now logging statements instead of print statements. This might not work always
At this moment there is a heuristic for tightening the pruning parameters if the lattice is large. Probably that should be made configurable
There might be too much logging in some places

You can merge now, or wait until I'm back from holiday (3 weeks), after which I should have some time to clean it up a bit.

senarvi commented 7 years ago

We can wait 3 weeks. Maybe you could also add short instructions how to use this with Kaldi lattices in this file: https://github.com/senarvi/theanolm/blob/master/docs/applying.rst

The reason that you changed the print statements, was it because you want to output the lattice to standard output? The only problem I see is that the user might not always see error messages. Maybe user messages should be written to standard error?

Anyway, let's continue from this after your vacation.

psmit commented 6 years ago

I have now moved all heuristics to command line options and wrote some documentation. Only some testing is now left.

senarvi commented 6 years ago

Thank's a lot Peter for this!