- kurzfristige Projekte:
- bababa2, SIGdial Poster
- TO DOs, unprioritisiert: a) Silbengrenzen, von
Aussprachewörterbuch kommend; b) echtes Audio verwenden,
Kielkorpus; c) ASR verwenden, Wörter, ngramme; d) bessere
speech states, phrasengrenzen (f. BCs); e) besser
TT-Strategien; f) simulation, constant time < (or >)
real-time; g) bessere Evaluation; h) interruption management;
i) BC management; j) Parametrisierung (chattiness,
interruption propability, etc.); k) adaptivity
- mögliche Ansätze f. Paper:
- in Richtung David T., `believable, non-scripted content-free
background chatter'
Nicht sehr überzeugend; um online erzeugt zu werden, doch
ein wenig resourcenhungrig. Nur für Hintergrundgerede würde
das wohl niemand ernsthaft einsetzen.
- `simple rules create realistic turn-taking patterns'
SSJ rules as *generative* rules, not just descriptive. Shows
that such a set of rules, together w/ some audio magic, are
enough to produce patterns that are `natural' (in a way that
needs to be defined properly). Again sort of upper-bound; to
get something like this working properly within a real
system, here's what we would need in terms of components.
- to do first: b), d), e), g).
- needed: more principled metric for `naturalness' of
resulting corpus. Multi-dimensional: distribution of gaps
& overlaps, balance btw speakers, turn length (in time,
but also # of utterances).
- `syntactic and prosodic language modelling for incremental
utterance segmentation', für Coling
utterance end pointing, but in an incremental set up. Needed to
know where to clear the chart of the parser. Connected to a
well-researched task (i.e., easy to motivate & compare), but
different in that we don't allow (as much?) right context.
- method:
- select only multi-utterance turns; EOUs to find are the
turn-internal ones.
- use original data & variants w/ various WER.
Those need plausible time information. How much does
this degrade performance?
- what's a good way to evaluate this? follow-on effects of wrong
decisions: an insert for example makes us restart the parser,
and hence get other things wrong?
das, 03/03/08 10:37 (GMT)
Add a new page under this one
David Schlangen : Home Page > minutes030208b > 030208cont