Cours 03
45
pages
English
Documents
Le téléchargement nécessite un accès à la bibliothèque YouScribe Tout savoir sur nos offres
45
pages
English
Documents
Le téléchargement nécessite un accès à la bibliothèque YouScribe Tout savoir sur nos offres
Word segmentation…l e p e t i ch i en m o r s e r a en t é r é d e m ainle pe ti chien mor se ra en té ré de mainle petit chien mort sera enterré demain0 2.1Recall• Finnish vowels:front backi y ue Ø oorthography: ö = /Ø/æ aä = /æ/‘vowel harmony’: within words, /y,Ø,æ/ cannot co-occur with /u,o,a/ (/i,e/ can occur with both /y,Ø,æ/ and /u,o,a/)• Finnish stress: all words have initial stressWord segmentation strategies• transitional probabilities• phonotactic cues• allophonic cues• prosodic boundaries Where do fixed stress and vowel harmony fit in?Word stress versus vowel harmony• word stress:– in principle, 100% reliable: every Finnish word starts with a stressed syllable– in practice, much less reliable: in running speech, not every stressed syllable is realized with clear acoustic stress cues (F0 and energy), and these same cues have other functions than signaling stress (intonation, emotion, phrase boundaries,…)Word stress versus vowel harmony• vowel harmony:– in principle, not 100% reliable:• disharmonic sequences signal word boundaries (except in some compounds and loanwords, e.g. ‘analyysi’), but the reverse does not hold (harmonic sequences are found both within and across words)– in running speech, however, very few exceptions to what is expected in principle• front and back vowels are reliably articulated as front and back, respectively!Vroomen et al. 1998• Word spotting: detect disyllabic words embedded at the end ...
Voir
Publié par
Langue
English
l e
le
le
p e
pe
t
ti
petit
i
Word segmentation
ch i en m
chien
chien
o
mor
mort
r
s e r a en t é r é d e
se ra en té ré de
sera
enterré
m
ain
main
demain
•
•
Finnish vowels: front back i y u eØo æa
Recall
orthography: ö = /Ø/ ä = /æ/
‘vowel harmony’: within words, /y,Ø,æ/ cannot co-occur with /u,o,a/ (/i,e/ can occur with both /y,Ø,æ/ and /u,o,a/)
Finnish stress: all words have initial stress
•
•
•
•
Word segmentation strategies
transitional probabilities
phonotactic cues
allophonic cues
prosodic boundaries
Where do fixed stress and vowel harmony fit in?
•
Word stress versus vowelharmony
word stress: in principle, 100% reliable: every Finnish word starts with a stressed syllable in practice, much less reliable: in running speech, not every stressed syllable is realized with clear acoustic stress cues (F0 and energy), and these same cues have other functions than signaling stress (intonation, emotion, phrase boundaries,
)
•
Word stress versus vowelharmony
vowel harmony: in principle, not 100% reliable: • disharmonic sequences signal word boundaries(except in some compounds and loanwords, e.g. ‘analyysi’), but the reverse does not hold (harmonic sequences are found both within and across words) in running speech, however, very few exceptions to what is expected in principle • front and back vowels are reliably articulated as front and back, respectively!
•
•
Vroomen et al. 1998
Word spotting: detect disyllabic words embedded at the end of trisyllabic non-words
Exp. 1 cue: vowel harmony only • disharmonic vs. harmonic context • no stress on embedded word • ex.: detectHYmyinPUhymyvs.PYhymy results: • better performance forPUhymythan forPYhymy
•
•
Vroomen et al. 1998
Word spotting: detect disyllabic words embedded at the end of trisyllabic non-words
Exp. 2 vowel harmony + stress • disharmonic vs. harmonic context • stress on initial syllable of disyllabic word • ex.: detectHYmyinpuHYmyvs.pyHYmy results: • marginally better performance forpuHYmy thanandpyHYmy • overall better performance in this experiment than in the previous one
Exp. 3: artificial language learning • four artificial languages, each containing 6 CVCVCV ‘words’
•
no stress
initial stress
harmonious motamu
'motamu
disharmonious motamy
'motamy
each language consists of random concatenation of the 6 words produced by a speech synthesizer no pauses between words all syllables have equal length in stress languages: raised F0 on initial syllable
• •
•
•
Exp. 3: artificial language learning
Finnish and French participants background assumption: listeners use their native language segmentation strategies Finnish: initial stress, harmony French: final stress, no harmony each participant exposed to one of the four languages length of exposure: 10 minutes (each word appears 150 times) test: forced choice (word versus non-word)
100
80 60
40
20
0
*
*
Finnish
Results
Finnish - all groups better than chance - main effect of stress - marginal stress-harmony interaction - no-stress condition: effect of harmony
French
no stress harmonious no stress disharmonious initial stress harmonious initial stress disharmonious
French - all groups better than chance - no main effects or interaction
•
Extra discussion
(no tin Vrooemnet al.)
How did the French manage to perform better than chance? suppose they incorrectly inferred the presence of a word boundaryaftereach stressed syllable: shouldn’t they be worse than chance? this incorrect strategy raises the number of words from 6 to 30 (can you show this?)impossible to learn within 10 minutes French subjects probably relied on transitional probabilities
