PS lattice file format

[Vassil Panayotov]

Hi,

What is the exact meaning first-endframe and last-endframe fields in .lat
files (for example see
http://www.pasteall.org/18753)? As a related
question: for which frames are the acoustic scores (in the edge table section
of the file) calculated? Is it for example for the frames between start frame
and first-lastframe, between start and last-endframe or something in between?

Thank you!

[Vassil Panayotov]

Thinking again about my second question(about the acoustic scores) it seems
logical that the scores are calculated between the start frame of the source
node and the start frame of the destination node. If this is the case for the
above link one(and only in this case) hypothesis for the word "go" spans the
frames between 46 and frame 64, which is the first frame of the word
"forward". For this hypothesis the acoustic score (log_1.0001) is -1653760

21 19 -1653760

. Can someone (e.g. Nickolay) confirm this?

[Vassil Panayotov]

Here is the GraphViz picture corresponding to the lattice in my first post:
http://www.pasteall.org/pic/8711 .
This picture was produced using the lattice.py module from SphinxTrain (not
sure this is the proper way to calculate posteriors, but it is irrelevant for
now):

import cmusphinx.lattice

dag = cmusphinx.lattice.Dag('goforward.lat')
dag.posterior()
dag.dag2dot('goforward.dot')

What I find puzzling is that for some of the nodes there are no arcs. For
example I think there should be an arc between "s/0" and "s/46" node and
between for example ten/117 and leaders/152.
Is it a bug?

[Nickolay V. Shmyrev]

Hello

This is not really a bug, algorithm just works this way. It builds lattice
from the end trying to find predecessor candidates and link to them. Sometimes
it can happen that no good predecessor can be found and the node says
unlinked. For details see the function

ps_lattice_t * ngram_search_lattice(ps_search_t *search)

Of course there could be a step to clean such nodes, but it's just not
implemented. As far as I understand it doesn't affect any further decoding
steps