lattice-align-words-lexicon failing

[Rémi Francis]

Some of my lattices fail to align with:

WARNING (lattice-align-words-lexicon:ProcessFinalWrapper():word-align-lattice-lexicon.cc:252) No final-probs to process.

WARNING (lattice-align-words-lexicon:main():lattice-align-words-lexicon.cc:108) Algorithm produced no output (due to --max-expand?), so passing input through as output, for key 20140920_wesat_workers_hand_out_soap_and_advice_as_sierra_leone_locks_down_00002

LOG (lattice-align-words-lexicon:main():lattice-align-words-lexicon.cc:138) Successfully aligned 0 lattices; 1 had errors.

If I use lattice-align-words, I get:

WARNING (lattice-align-words:OutputArcForce():word-align-lattice.cc:578) Invalid word at end of lattice [partial lattice, forced out?]

WARNING (lattice-align-words:AlignLattice():word-align-lattice.cc:320) Number of states in lattice exceeded max-states of 485200, original lattice had 4843 states.  Returning what we have.

Is there any way to fix this?

Also, does lattice-align-words-lexicon return the phone aligned lattice or the original one?

[Daniel Povey]

Is it possible that you have symbols on your graph's output that don't
correspond to actual words?

> Also, does lattice-align-words-lexicon return the phone aligned lattice or
> the original one?

If it succeeds it outputs the word-aligned lattice, but if it fails I
think it outputs the original lattice. You'd have to check the code
to be sure.

Dan

[Rémi Francis]

On Tuesday, 7 June 2016 21:15:05 UTC+1, Dan Povey wrote:

Is it possible that you have symbols on your graph's output that don't
correspond to actual words?

All the words in the graph are in align_lexicon.int, however I have a word that has the same pronunciation than <eps> (both have SIL, no position dependent markers), could that be an issue?

It was working in the past, however I am experimenting a bit with the graphs, so there could be a bad interaction somehwere, and I don't follow completely the alignment algorithm yet.

 

If it succeeds it outputs the word-aligned lattice, but if it fails I
think it outputs the original lattice.  You'd have to check the code
to be sure.

Yeah I just had to look at the output. I does output the original lattice if it fails, not the phone aligned one.

 

[Rémi Francis]

On Tuesday, 7 June 2016 21:15:05 UTC+1, Dan Povey wrote:

Is it possible that you have symbols on your graph's output that don't
correspond to actual words?

Actually you were almost right, all the words are here but I was using an align_lexicon.int that had more words, so the integer ids didn't match. 

When using the right one it works.