about "lattice-to-post" and "kaldi::LatticeForwardBackward"

[trung...@gmail.com]

I want to "export" state posterior from lattice, and it seem "lattice-to-post" is a recommended choice. But I can't understand the implementation of the tool.

As my understanding, Lattice contain both graph cost and AM cost, so why do we need "forward-backward over lattices and computes the posterior probabilities of the arcs" (kaldi::LatticeForwardBackward) instead of get it from lattice weight directly.

(forward-backward algorithm is complicated for me, so I haven't get the idea of LatticeForwardBackward yet)

Thanks

[Nickolay Shmyrev]

Lattice arc weights and lattice arc posterior are different things. Lattice arc posterior is score of all paths passing through the arc divided to score of all paths in lattice. So you need forward-backward to calculate arc posterior value.

For more information you can read section "3.1 Lattice-based Posterior Estimation" from 

POSTERIOR PROBABILITY DECODING, CONFIDENCE ESTIMATION AND SYSTEM COMBINATION 

G. Evermann & P.C. Woodland
http://mi.eng.cam.ac.uk/reports/svr-ftp/evermann_stw00.pdf

[trung...@gmail.com]

Thank❤️ you for the helpful information.