You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Relates to MBR decoding outputting CTM. See following exchange on kaldi-help:
Rémi Francis via googlegroups.com
9:01 AM (4 hours ago)
to kaldi-help
Hi,
I'm using lattice-to-ctm-conf with mbr decoding to get the ctm with confidences, but I've noticed that sometimes the timing of the words overlap a bit:
audio-000000 1 92.17 0.85 football 0.66
audio-000000 1 93.03 0.34 scouting 0.99
audio-000000 1 93.36 0.54 agency 1.00
Here "scouting" ends at 93.37 whereas agency starts at 93.36.
I've looked at the code and it seems to come from averaging the times of arcs in the lattice, but I don't know if this is a bug or just something to expect.
An easy fix would be to fiddle the timings so it doesn't overlap, since anyway the timings are precise only up to 0.03s, but I wonder if there is a better fix inside the binaries to do.
to kaldi-help
It's inherent to the algorithm that the timings may overlap a bit... it would be possible to modify the algorithm to do a second pass over the times just to remove any overlap. If you had time to do that it would be great. Search for GetOneBestTimes(). I'll create an issue to keep track of it.
The text was updated successfully, but these errors were encountered:
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
This issue has been automatically closed by a bot strictly because of inactivity. This does not mean that we think that this issue is not important! If you believe it has been closed hastily, add a comment to the issue and mention @kkm000, and I'll gladly reopen it.
It's not exactly what was asked, but I've written a Perl script that does the fiddling of the timings. resolve_ctm_overlaps.py also resolves overlaps, but does so by deleting overlapping segments. My script does the dumb thing and decreases the duration of the segments until they no longer overlap with later segments, plus some handling of empty segments.
This script could easily be added as an optional final stage to the get_ctm_conf{,_fast}.sh scripts. Let me know if you're interested in me making a PR.
Relates to MBR decoding outputting CTM. See following exchange on kaldi-help:
Rémi Francis via googlegroups.com
9:01 AM (4 hours ago)
to kaldi-help
Hi,
I'm using lattice-to-ctm-conf with mbr decoding to get the ctm with confidences, but I've noticed that sometimes the timing of the words overlap a bit:
audio-000000 1 92.17 0.85 football 0.66
audio-000000 1 93.03 0.34 scouting 0.99
audio-000000 1 93.36 0.54 agency 1.00
Here "scouting" ends at 93.37 whereas agency starts at 93.36.
I've looked at the code and it seems to come from averaging the times of arcs in the lattice, but I don't know if this is a bug or just something to expect.
An easy fix would be to fiddle the timings so it doesn't overlap, since anyway the timings are precise only up to 0.03s, but I wonder if there is a better fix inside the binaries to do.
Daniel Povey dpovey@gmail.com
1:12 PM (1 minute ago)
to kaldi-help
It's inherent to the algorithm that the timings may overlap a bit... it would be possible to modify the algorithm to do a second pass over the times just to remove any overlap. If you had time to do that it would be great. Search for GetOneBestTimes(). I'll create an issue to keep track of it.
The text was updated successfully, but these errors were encountered: