Date post: | 17-Dec-2015 |
Category: |
Documents |
Upload: | brandon-aron-small |
View: | 214 times |
Download: | 0 times |
Pitch accent alignment in Egyptian Arabicmore evidence for cross-linguistic variation
Sam Hellmuth [email protected]
PaPI 2005, Barcelona20th June 2005
Egyptian Arabic pitch accent alignment
Aim:• to explore the surface phonetic alignment patterns of Egyptian
Arabic pitch targets – in rising pre-nuclear (= non-final) pitch accents– in different syllable types
to establish their phonological representation to contribute to the growing range of cross-linguistic alignment data
towards pitch-accent typology
Egyptian Arabic (EA):• = Egyptian Arabic: the dialect of Arabic spoken in Cairo
– and also by educated people throughout Egypt– all data reported here collected in Cairo
autosegmental-metrical theoryIn AM theory, intonational contours are analysed as:• a sequence of pitch targets
– H or L or bitonal combinations thereof
• autosegmentally associated with prosodic structure– aka metrical structure: syllables, feet, words, phrases...
Gussenhoven 2002
Ladd 1996, P&B1988 inter alia
‘starred’ tone: associated with the stressed syllable of the main stress foot of the accented word
pitch accent alignment• recent discussion: can phonological association (‘starredness’) be
inferred from surface alignment of individual pitch targets? Ladd 2003, Prieto et al (in press)
• effects of prosodic context on surface alignment– Prieto et al (1995) Mexican Spanish pre-nuclear rising accents
• in open syllables eg número– L very stably aligned at left edge of stressed syllable but see Prieto (in press)
– H alignment is affected systematically by:• proximity to a prosodic boundary• proximity to other pitch accents
– results reproduced for Lebanese Arabic (LA) Chahal 2001
• patterns of surface alignment to segmental landmarks– eg Arvaniti et al (1998) Greek pre-nuclear rising accents – targets independently aligned to specific landmarks in the string– L aligned very stably at the left edge of the stressed syllable
• onset of the initial consonant of the stressed syllable (C0)– H also aligned stably to segmental landmarks ‘segmental anchoring’
pitch accent alignment
two key studies for comparison here:• Atterer & Ladd 2004
– comparison of L/H target alignment – in two dialects of German– open syllables
• Ladd, Mennen & Schepman 2000– comparison of alignment in Dutch long vs short vowels– L alignment very stable (at C0)– H aligned:
• within stressed vowel in CVV (long/tense)• into following consonant in CV (short/lax)
• research questions: – how are pitch targets in EA non-final pitch accents aligned?– does alignment of EA pitch targets vary across syllable types?
AL2004:187
pitch accent alignmentwhat is known about EA pitch accents?• highly populated pitch accent distribution
– “Arabic seems to have a greater tendency to accent all words..” Mitchell 1993:230
– “lexical stress of every content word will be stressed in continuous speech if.. nothing to cause suppression of the stress” Heliel 1977:125
• cf Spanish, Greek (Jun 2004), NEP, Brazilian Portuguese (Vigario & Frota 2003)
• non-final pitch accents are bitonal– “an ‘up-and-down’, ‘see-saw’ effect.. characterises the spoken
language” .... “unaccented syllables in the same word.. remain on the same height.. whereas pitch dips markedly lower to pre-accentual syllables in the following word.. from which a ‘jump’ takes place to the height of the following accented syllable” .... Mitchell 1993
– “pre-final stressed syllables.. are depicted by a late peak situated on the last point of the syllable... [and] are all rising” Rifaat 1991
methodologystudy modelled on Atterer & Ladd 2004 BUT:• three types of target syllable
1 CV light open short tense vowel2 CVC heavy closed short lax vowel Shahin 1996
3 CVV heavy open long tense vowel• target syllables word-initial, target word non-initial in sentence
– to clarify alignment facts in heavy vs light syllables– to facilitate comparison with the results of other studies (some CV, some CVC)
• word-medial CVC closed syllables also tested:4 CVC heavy closed short lax vowel– is alignment of pitch targets to word edge(s) or to the stressed syllable?– to facilitate comparison with the results of other studies (some word-initial,
some word-medial)
methodology• targets placed in frame sentences, as ‘natural’ as possible
– 6 sentences per ‘set’ > 24 target sentences + distractors• read three times by 15 EA speakers
– 6 female & 9 male– all at pre-intermediate level or lower in English
• 24 x 15 x 3 = 1080 (270 per set) > 939 fluent tokens for analysis– digital recordings using ProTools 6.1 on MBox, headset microphone
• 44.1KHz 16 bit, re-sampled to 22.5KHz– F0/spectrogram & measurements extracted using Praat 4.2
• in effort to achieve naturalness > clash context not fully controlled # intervening σ before # intervening σ after set 1 0 or 1 2-4set 2 0 or 1 2-4set 3 0 or 1 1-2set 4 1 or 2 1-3
Error analysis Frequency
Valid disfluent 78
fluent 939
H phrase accent 42
L phrase accent 13
preceding phrase accent
8
Total 1080
methodology
sample target sentences:
• šufna malik il-?urdun lamma ruHna l-?urdunWe saw the king of Jordan when we went to Jordan
• ?akalna manga laziiza giddan min-is-suu? We ate a really delicious mango from the market.
• ir-ruzz da maaliH ?awwi wiTa9muh waaHiš That rice is really salty and tastes horrible.
• il-mudarris mimalmil min iT-Talaba The teacher is nervous of his students.
methodology• pitch event variables: L1 H L2• alignment variables:L1-C0 L1-V0 (L1-X) H-C1 (H-C2) H-V1• peak delay: H-C0
syllable duration#1: treats C1 as end of syllable in set 1 (CV.CV)syllable duration#2: treats V1 as end of syllable in set 1 (CVC.V)> relative peak delay (RPD): peak delay/syllable duration (RPD1/RPD2)
NB L2 observed during transcription always to fall within following word
results
L alignment variables H alignment variables relative peak delay
Report
7.2130040303 -58.2395 628.3664 43.4505 -5.3129 1.3369 .9744
215 215 215 215 215 215 215
26.32698999 27.46485 175.06527 37.36155 36.03676 .26205 .18240
.9431435004 -59.5140 584.9197 44.5366 -1.7864 1.3611 .9894
16.23682462 -44.8495 658.1045 25.8998 -45.0429 -114.4879 .7895 .7895
249 249 249 249 249 249 249 249
29.32466354 31.68055 301.02599 29.07838 40.01888 46.66035 .17662 .17662
7.2104102649 -48.9990 605.3918 23.9491 -43.9744 -114.4021 .7742 .7742
10.40832087 -57.9137 557.2166 3.5714 -48.2291 1.0403 1.0403
245 245 244 245 245 245 245
33.61346130 34.28911 186.86404 46.53571 46.73823 .28189 .28189
2.4434110385 -60.7945 552.9166 -13.7227 -60.6610 .9282 .9282
5.8167213496 -58.9393 115.6099 20.9136 -48.7410 -111.2681 .7808 .7837
230 230 230 230 230 230 230 230
30.03691783 31.09688 70.63411 31.47868 41.04119 40.22659 .18171 .17152
4.2407925799 -56.5006 94.2672 17.1441 -49.5622 -111.6065 .7737 .7737
10.09760354 -54.7752 492.0233 22.8712 -46.8186 -71.4138 .9781 .8958
939 939 938 939 479 939 939 939
30.26705470 31.86077 297.89220 39.30651 40.51273 62.12170 .31961 .23799
3.7422844715 -55.8797 533.6717 20.0253 -46.1332 -79.2413 .8791 .8634
Mean
N
Std. Deviation
Median
Mean
N
Std. Deviation
Median
Mean
N
Std. Deviation
Median
Mean
N
Std. Deviation
Median
Mean
N
Std. Deviation
Median
Set #Set 1
Set 2
Set 3
Set 4
Total
L1-C0 L1-V0 L1-X H-C1 H-C2 H-V1Relative Peak
Delay #1Relative Peak
Delay #2
results: L alignment
L alignment variables, all speakers, by set:• L is aligned closer to C0 than to V0 • ie to the left edge of the syllable
results: L alignmentdetails of average speaker behaviour in L alignment across sets:
– based on speaker means within each set: # who align L before C0 # who align L after C0set 1 8 3F,5M 7 3F,4Mset 2 3 3F,4M 14 6F,8Mset 3 5 1F,4M 10 5F,5Mset 4 3 1F,2M 12 5F, 7M
• two speakers align L on average before C0 in 3 out of 4 sets• mrf/mun
– most instances of early alignment are in set 1– BUT no one speaker aligns before C0 consistently across sets
• strong tendency to align L just after C0 (but not universal)
working hypothesis: in EA L is aligned “just after C0”
L alignment
• in EA L aligns to left edge of the syllable, most often just after C0
• however there is a considerable variation in the dataset– across a
range of 200 ms for some speakers
626059556457676559705962636968N =
Speaker
muy
mun
mrf
moa
mns
mma
miz
mgk
meh
fsf
fna
fhx
fhm
fhg
faa
L1
-C0
200
100
0
-100
-200
Set 2Set 2Set 1Set 4
Set 3Set 3
Set 3Set 2Set 2Set 3Set 4
Set 3
Set 4Set 4Set 3Set 2Set 4Set 4Set 4
Set 4Set 4Set 4Set 3Set 3Set 4Set 4
Set 2Set 2Set 3
Set 3Set 4Set 3Set 3Set 4
Set 4
Set 4Set 4Set 4Set 3Set 3Set 4
Set 4
Set 4Set 4
Set 3
Set 2
Set 1Set 4Set 3Set 3
Set 4Set 2Set 2
Set 2
Set 4
Set 4Set 4
Set 4Set 4Set 3Set 4
Set 2Set 2
Set 2
Set 4
Set 2
Set 3
Set 3
Set 1
Set 4Set 3
Set 3
Set 1
Set 3
Set 3
Set 2Set 4
Set 1Set 1Set 2
Set 2Set 2
Set 4Set 1Set 3Set 4Set 1
Set 2Set 3Set 4Set 4Set 4Set 2Set 3
Set 4
Set 4
Set 4
Set 3
Set 2Set 1
clash
results: H alignment
H alignment variables, all speakers, by set:• H is aligned after C1• ie to the right edge of the syllable?
Set #
4.54.03.53.02.52.01.51.0.5
H-C
1
300
200
100
0
-100
-200
Gender
Male
Female
results: H alignment
rise duration (H-L) x syllable duration:• rise duration maps more closely to sylldur#2 than sylldur#1• suggests that alignment of H best described in terms of syllable definition #2
results: H alignment
in terms of segmental landmarks, H alignment patterns differently in light vs heavy syllables
• in CV (set 1) just before/after V1– 8 speakers (1F/7M) align H before V1– 7 speakers (5F/2M) align H after V1
• mean RPD1 > 1 (H aligned outside stressed syllable)
• in CVC (set 2/4) between C1 & C2– all speakers align H between C1 & C2
• mean RPD1 < 1 (H aligned well inside stressed syllable)
• in CVV (set 3) just before/after C1– 8 speakers (1F/7M) align H before C1– 7 speakers (5F/2M) align H after C1
• RPD1: 8 speakers: <1; 7 speakers >1
clash
C V C V
C V C C
C V V C
results: H alignmentdistance from H to syllable end (#2):
• H aligns later in open syllables (CVV & CV) than in closed syllables (CVC)– an effect of vowel
quality? (tense/lax)
• however there is considerable variation in the dataset...
Set #
4.54.03.53.02.52.01.51.0.5
H t
o s
ylle
nd
#2
200
100
0
-100
-200
Gender
Male
Female
H alignment
626059556457676559705962636968N =
Speaker
muy
mun
mrf
moa
mns
mma
miz
mgk
meh
fsf
fna
fhx
fhm
fhg
faa
H t
o s
ylle
nd
#2
300
200
100
0
-100
-200
-300
Set 3
Set 3
Set 2Set 4
Set 3
Set 3
Set 4
Set 4Set 4Set 3
results: towards explanationsQ: is H aligned a fixed
distance from L?• as already seen,
there is some correlation between rise duration & syllable duration (#2)– suggesting that as
the duration of the syllable increases the position of H also moves
• the correlation is weak but is significant– Kendall’s τ 0.262
• p < 0.01
some support for ‘segmental anchoring’ in EA
results: towards explanationsQ: is there a fixed slope
(rate of F0 change)?• F0 change
(semitones) x rise duration
• there is a correlation– unlike Greek
(Arvaniti et al 1998)
– suggesting that as the duration of the syllable increases the position of H also moves
• again, the correlation is weak but is significant– Kendall’s τ 0.136
• p < 0.01F0 change (sem)
121086420
Ris
e D
ura
tion
500
400
300
200
100
0
Set #
Set 4
Set 3
Set 2
Set 1
Total Population
Rsq = 0.2266
compare findings of Elzarki 1996 (EA pronunciation of Modern Standard Arabic)..
results: towards explanationsQ: how stable is H
scaling?• mean H F0 (semitones)
– ie are speakers aiming at a specific H pitch target level?
• unable yet to normalise for individual speaker pitch range (work in progress)– but visually there
does not seem to be an effect of syllable type on H scaling
Set #
Set 4Set 3Set 2Set 1
Me
an
HF
0S
30
20
10
0
fna
fsf
meh
mgk
miz
mma
mns
moa
mrf
mun
muy
results: summary
• L alignment– at left edge of stressed syllable
• H alignment: – at right edge of stressed syllable
• explanations:– fixed duration?– fixed slope?– segmental anchoring?– all three seem to be relevant
• ? due to enlarged speaker set and resulting variation
• in this context the consistent alignment of L and H to the syllable edges is all the more striking
C V C V
C V C C
C V V C
discussion: cross-linguistic variation in alignment
comparison to other Arabic dialects:• Lebanese Arabic Chahal 2001 (4 speakers)
• LA: L aligns before/after C0 (depending on word position)• EA: L aligns just after C0 (slight variation due to word position)
– but in same direction (L aligned earlier in medial syllable than initial)
• LA: H aligns outside the stressed syllable in CVC syllables• EA: H aligns inside the stressed syllable in CVC syllablesIn LA L aligns earlier than in EA, and H aligns later...
discussion: cross-linguistic variation in alignment
comparison to other Arabic dialects:• Moroccan Arabic Yeou 2004 (5 speakers)
in MA:• L aligns “close to the onset of the syllable”• H aligns after C1 in CV syllables
– “after the end of the stressed vowel” • H aligns after C1/before C2 in CVC syllables
– within the stressed syllable (inferred from RPD value)
In MA both L and H align similarly to their EA counterparts
discussion: cross-linguistic variation in alignment
comparison to other languages:• comparing data in short open syllables
additional evidence in support of a continuum of cross-linguistic variation in phonetic alignment of phonologically parallel pitch accents
is it appropriate however to make a direct comparison of EA with these languages?
are these pitch accents phonologically parallel?
discussion: phonological specification of EA pitch accents
• unable at present to choose from among the three possible explanations – fixed duration vs. fixed slope vs. segmental anchoring– BUT: “association cannot be based on phonetic alignment in any straightforward
way”.. Arvaniti et al (2000) (emphasis mine)
• working hypothesis for EA pitch accents: • in the spirit of Prieto et al (in press)
– bitonal pitch accent L+H– primary association of H to stressed syllable
• perceptual salience of H cf Rifaat 2003
– no secondary association of H?– default alignment of L to onset of stressed syllable
• problem?: association of strong element in pitch accent...• with weak element in foot
• it is only meaningful to compare EA surface pitch accent alignment facts then with languages which also employ L+H* (defined under the same set of assumptions)
– ?Catalan L+H* “rise with delayed peak” (Catalan targets were open syllables)
• additional categories may also be needed:– should influence of fixed slope/duration be phonologically encoded?
L+H*
F
σ σ [ma lik]ω
شكرألف thank you!
With thanks to • the Egyptian Arabic speakers who acted as consultants• audiences at the UCL Phonology Reading Group & Manchester
Phonology Meeting for comments on earlier versions of this paperThis work was funded by AHRB postgraduate award 01/59198.