About lattice-to-post

[Anis Chihi]

Hello Everybody,

I have just a question concerning lattice-to-post:

I run lattice-to-post over a lattice generated by decoding some data. This is the command I used:

lattice-to-post --acoustic-scale=0.1 "ark:gunzip -c exp/tri2/decode_eval2000_nosp_sw1_tg/lat.1.gz |" ark,t:- | less > RADIOACTIVE/latticepost.txt

which worked very fine. I just wanted to know the meaning of the output I got: so here is the first line I got over an utterance with 81 frames:

en_4156-A_030185-030248 [ 4 1 ] [ 16 1 ] [ 15 1 ] [ 18 1 ] [ 17 1 ] [ 17 1 ] [ 17 1 ] [ 17 0.9916295 18616 0.00837055 ] [ 17 0.9880814 15924 0.003548065 18615 0.00837055 ] [ 7478 0.9880814 15952 0.003548065 18646 0.00837055 ] [ 7526 0.9880814 15992 0.003548065 18645 0.00837055 ] [ 7525 0.9880814 15991 0.003548065 18678 0.00837055 ] [ 1276 0.00837055 7172 0.003548065 7525 0.9880814 ] [ 1328 0.00837055 7204 0.003548065 7525 0.9880814 ] [ 1420 0.00837055 7232 0.003548065 7525 0.9880814 ] [ 7231 0.003548065 7562 0.9880814 12432 0.00837055 ] [ 7231 0.003548065 7561 0.9880814 12516 0.00837055 ] [ 7231 0.003548065 7561 0.9880814 12515 0.00837055 ] [ 7231 0.003548065 7561 0.9880814 12515 0.00837055 ] [ 7231 0.003548065 12515 0.00837055 19192 0.9880814 ] [ 7231 0.003548065 12554 0.00837055 19191 0.9880814 ] [ 19186 0.01191861 19191 0.9880814 ] [ 19185 0.01191861 19191 0.9880814 ] [ 19185 0.01191861 19191 0.9880814 ] [ 19185 0.01191861 19191 0.9880814 ] [ 19185 0.01191861 19226 0.9880814 ] [ 19225 0.9880814 19226 0.01191861 ] [ 19225 1 ] [ 19225 1 ] [ 19225 1 ] [ 19225 0.9916295 19262 0.00837055 ] [ 19260 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 19259 0.9916295 19261 0.00837055 ] [ 20114 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20113 1 ] [ 20174 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20173 1 ] [ 20232 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 20231 1 ] [ 4 1 ] [ 1 1 ] [ 16 1 ] [ 15 1 ] [ 15 1 ] [ 15 1 ] [ 15 1 ] [ 15 1 ] [ 18 1 ] 

Just want to be sure what this means, as far as I unsderstand, this is a per frame posterior transition probability. for example  [ 17 0.9916295 18616 0.00837055 ]  which is the 8th Entry in my case, would mean that:
Prob(transition-id=17|frame8)=0.991..  and Prob(transition-id=18616|frame8)=0.008.. . Of course all other post-Probs over the left Transition-ids would be zero.

Is what I am saying here right?

Thank you in advance.

[Daniel Povey]

It's a posterior over transition-ids-- transition-ids are an index
that identifies both the pdf-id and which arc in the FST you took
(equivalent to which HMM-state transition you took). For the most
part you can view it as a distribution over pdf-ids, at least that's
normally the most useful information in it.
It's kind of a way of summarizing the distribution over paths in the
lattice, and their probabilities.
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.

[Anis Chihi]

Thanks Dan,

Now if I want for example to use the posteriors as labels for the data for the case of unsupervised adaptation, I could basically use post-to-pdf-post to get the pdf-level posteriors, since my triphone states correspond to pdf-ids. Am I right?

Thank you.

[Daniel Povey]

Yes.

[Anis Chihi]

Thanks a lot for your help