There is no such thing as confusion networks in Kaldi, although there
is something with a similar structure that you can get from
lattice-mbr-decode. Perhaps you have another toolkit in mind.
Generally speaking the scores in lattices will be likelihoods, and you
would have to do a forward-backward to get the posteriors. In Kaldi
lattices the negative-log of the graph (LM) and acoustic likelihoods
are stored separately. There is documentation on lattices at
kaldi-asr.org/doc/. (it has its own section). If you have any
*specific* questions and that documentation doesn't answer them, ask
again.
Dan
> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to
kaldi-help+...@googlegroups.com.
> For more options, visit
https://groups.google.com/d/optout.