David Schlangen : Home Page > minutes051107
- present: Michaela, Timo, David
- done:
- parser features are already useful on "is this the last word?"
task. Parser still has bug in computing probabilities, though.
[ Michaela ]
- evaluation of f0 trackers via average agreement. Snack
methods seem to agree well, InPrF0 is getting there. [ Timo ]
- Suggestion: compare higher level tendencies, do they in
general *move* in same direction?
- Octave mistakes may not lead to problems, if all we need are
the directions / tendencies of the f0 contour.
- next steps
- for pentomino dialogue system, we probably should go back to the
lattice parsing idea, with a hand-crafted grammar (+ learned
rule probabilities to get most likely parse). But for now, for
TRP-projection project, we should try to stay domain independent
& continue with Amit's parser.
- immediate next step: join parsing features and acoustic
features, offline. (A.k.a. "offline-Verdengelung".)
- parser features are word-based, ac feats are time-based. Hence
parser features need to be spread over time / aligned to time
"ticks". Here we can experiment with different versions,
e.g. putting parser feature at frame of word end, or using a
fixed delay...
- would be best to have version of parsing results that allows
computing (on the fly) of different combinations of this and the
acoustic information in one file, which then can be basis of
learning & testing.
- this file (or the scripts producing it) should also be flexible
enough to allow definition of different learning tasks:
- is this word the turn final one?
- does turn come in the invertval (bs_n, be_n) milliseconds from
now? (I.e., (bs_n, be_n) forms bins; here we could experiment
with equal-size bins and bins that get bigger.)
das, 11/05/07 02:58 (GMT)
Keyword: inpro,
meetings,
minutes,
TRP-predictorAdd a new page under this one
