meeting-2009-09-10
Kick-off meeting
- antti keeps the system
- the parts we will have to do for new languages
- letter-to-sound
- how we will model prosody
- phrasing
- how to split training sentences
- letter-to-sound
- a big issue is the handling of numbers and abbreviations
- inflect numbers for case
- tts for prosody
- html marking as clues
- hmm synthesis
- does not glue
- make a reverse speech recognition
- not tobi
- perceptual prominence model
- accent
- stress on the first syllable
The Antti model for now:
- 0 - deaccented
- 1 - weak cont wo
- 2 - strong foc
- 3 - emphatic
As comparision: the Trondheim model:
- 1 - nonfoc
- 2 - foc
- 3 - acc
3Pekkakin sen möi Minä puhun 2Pekastakin
cg-stype rule set:
- prom of -1 0 1 on word level
- prom of -1 0 1 on syll level
word level
funct w deacc pron v weak num adv N adj strong NP cost can be considered having focus and background theme rheme verb .... if theme has two constituents the first is sprominent and others are secontary or olc 3Pekka 2kalan 1osta 3-keskiviikkona 2-kaikki tutkijat 1-læhtivæt kotiin
rheme
S O systemic ordering deviations from syystemic ordering has prosodic contecequence if A B in text sys B A act A B => prominent-B, less-prominent-A hypothesis for finnish SO subj < obj < loc < manner < osurce < goal if any of these is moved to the left in the string it will be part of the theme and therefor less prominent. pekka osti KALAN-obj TORILTA-src Pekka osti torilta-src KALAN-obj
- if word order follows the systemic order we cannot say anything
- if word order deviates from systemic order (as indicated)
rule against data driven approch:
- pro
- insight to phon
- motivated
- not huge c
- rare
- insight to phon
- con