The segmental anchoring hypothesis' Effects of syllable structure, segmental composition and withinw

1
The segmental anchoring hypothesis. Effects of
syllable structure, segmental composition and
within-word position on H placement in Catalan
and Spanish

• Pilar Prieto, ICREA-UAB
• LPL seminar
• Aix-en-Provence, April 29, 2005

2
Plan of the Talk

• Introduction

• Factors affecting tonal alignment
• Segmental anchoring hypothesis (Ladd et al.,
  1999, Ladd to appear)
• Tonal alignment in Catalan and Spanish

• Segmental Anchoring in Spanish (joint work with
  F. Torreira)
• Experiment 1. To verify if H peaks are anchored
  to a specific segmental site in words contrasting
  in syllable structure and segmental composition.
• Experiment 2. To verify if H peaks are anchored
  to a specific segmental site in words contrasting
  in syllable structure and uttered at different
  speech rates.

3
3. Effects of within-word position in Catalan and
Spanish (joint work with E. Estebas) Experiment
1 To verify if H peak placement is affected by
within-word position and if Hs are anchored at
the right edge of the prosodic word in Catalan
and Spanish. Experiments 2 and 3 To verify if
Catalan and Spanish listeners are able to use
tonal alignment differences due to within-word
position in the identification and discrimination
of words with ambiguous word boundary location.
4
Regularities in tonal alignment

• Increasing evidence that LH points in the tonal
  space are aligned with the text in extremely
  consistent ways
• .

• Crosslinguistic regularity for L alignment in LH
  rises, i.e. L is anchored at onset of accented
  syllable
• Caspers van Heuven 1993 for Dutch Prieto, van
  Santen Hirschberg 1995 for Spanish Arvaniti,
  Ladd Mennen, 1998 for Greek Ladd, Faulkner,
  Faulkner Schepman 1999 and Ladd, Mennen
  Schepman 2000 for English Xu, 1998 for Mandarin
  Chinese Estebas-Vilaplana 2000 and Prieto 2005
  for Catalan)

• H alignment is quite variable peaks can
  variably be positioned by the end of the stressed
  syllable or at the following unstressed syllable,
  
• Silverman Pierrehumbert 1990 for English
  Prieto, van Santen Hirschberg 1995 for Spanish
  Arvaniti, Ladd Mennen, 1998 for Greek Prieto
  2005 for Catalan.

5
Prosodic factors affecting H alignment

• H placement has been shown to be greatly
  affected by the right-hand prosodic context
• Tonal crowding Peaks are retracted before
  upcoming pitch accents and boundary tones (Bruce,
  1977 Silverman Pierrehumbert 1990, Prieto et
  al. 1995, DImperio 2001, Prieto 2005, etc.)
• Word boundaries Peaks are retracted before
  upcoming word boundaries (see Silverman
  Pierrehumbert 1990 for English, Prieto, van
  Santen Hirschberg 1995 for Spanish and
  Arvaniti, Ladd Mennen 1998 for Greek).

6
Segmental anchoring hypothesis

• An unexpected stability effect was found when
  little or no tonal pressure was exerted on Greek
  pitch accents (Arvaniti, Ladd Mennen 1998). The
  H target aligned with the end of the
  postaccentual onset.
• Strict alignment effects are pervasive under
  changes of syllabic/segmental structure and
  speech rate
• Arvaniti Ladd 1995 Greek Ladd et al. 1999
  English Xu 1998 Chinese Schepman et al. subm
  Dutch Atterer Ladd, to appear, German, Ladd to
  appear)

Segmental anchoring hypothesis Ls and Hs are
anchored to specific points in the segmental
structure, regardless of speech rate, segmental
composition or syllable structure (Ladd et al
1999, Ladd in press).
Yet, other studies have shown effects of
segmental composition (van Santen and Hirschberg,
1994 Rietveld and Gussenhoven, 1995) and
syllable structure (DImperio, 2000)
7
Phonetic differences in H alignment across
languages
Segmental and syllabic anchoring of trailing
tones across different languages in contexts
without tonal coarticulation. Segments
Cconsonant, Vvowel, Ncoda, Ssyllable
Structure 0accented syllable, 1postaccentual
syllable (ex C0V0N0.C1V1).
8
Segmental anchoring in Spanish

• GOALS OF EXPERIMENTS 1 and 2 To verify the
  strong version of the segmental anchoring
  hypothesis for Spanish, that is, if prenuclear H
  peaks are anchored to a specific point in the
  segmental string.
• HYPOTHESES Prenuclear peak position should not
  be affected by the following factors
• segmental composition in the syllable?
• syllable structure?
• speech rate?

Experiment 1 focuses on the effects of segmental
and syllabic composition on prenuclear peak
placement.
Experiment 2 focuses on the effects of syllable
structure and speech rate on prenuclear peak
placement.
9
Experiment 1

• GOAL test the effects of segmental and syllabic
  composition on prenuclear peak placement.
• 4 speakers of Castilian Spanish read a corpus of
  96 sentences exhibiting the phonological variety
  needed to test our hypothesis
• A total of 432 tokens (48 utterances x 3 speakers
  x 3 repetitions)

10
Example of prenuclear LH

• Contour typically found in read speech initial
  prenuclear H does not exhibit peak displacement.

11

• Three female speakers of Castilian Spanish read
  the set of 48 test sentences three times, for a
  total of 432 utterances (3 speakers x 48
  sentences x 3 repetitions).

12
Labeling Scheme

• The following measurements in ms were manually
  extracted from the accented region
• o onset of the utterance
• c0 beginning of the initial consonant in the
  accented syllable
• v0 beginning of the vowel in the accented
  syllable
• k0 beginning of the coda in the accented
  syllable
• c1 beginning of the initial consonant in the
  postaccentual syllable
• v1 beginning of the vowel in the
  postaccentual syllable
• c2 end of the postaccentual syllable
• ew end of the word
• phrase-initial and final points (only Exp 2)
• L valley in LH accent
• H peak in LH accent

13
Example of prenuclear LH
Beginning of the sentence Emilio Rodríguez
terminó ayer (02b, Raquel 2rep)
14
Effects of syllable duration

• High positive correlation between syllable
  duration and H delay
• H is retracted in closed syllables

15
Effects of syllable structure

• Closed syllables display a significantly later H
  alignment than open syllables for the 3 speakers
• A two-way ANOVA showed significant effects for
  both SYLSTRUC and SPEAKER (plt0.000), plus an
  interaction between the two (p0.009)

16
Effects of syllable structure

• Tonal H targets are not strictly anchored at the
  vocalic offset it depends on syllable structure
  and it is speaker-dependent.

Tonal target for mándola is normally realized
within the coda consonant.
Tonal H target for nómina is not strictly
anchored at the vocalic offset.
Late Alignment -gt mándola

Early Alignment -gt nómina
Ex.
nómina payroll
mándola mandolin
l
Parallel findings for other languages Ladd et al
(2000) for Dutch, DImperio (2000) and DImperio,
Petrone Nguyen (2005) for Neapolitan Italian.
17
Effects of coda type

• Separate two-way ANOVAs revealed significant
  effects for SPEAKER (plt0.000), CODATYPE (plt0.000)
  both for hdelay and htoev0. For hdelay/codatype
  an interaction between SPEAKER and CODATYPE was
  found (plt0.013). No significant interaction
  (p0.097) was found in the case of htoev0.

18
Conclusions (Exp. 1)

• Experiment 1 reveals a loose alignment towards
  the end of v0 depending on the following factors
• Syllable duration Durations of onset and vowel
  have a linear effect on H placement.
• Syllable structure Closed syllables display a
  significant later alignment than open syllables.
• Segmental composition Coda type has an effect on
  peak alignment, as peaks align some ms later in
  nasal codas than in lateral codas.

19
Experiment 2

• GOAL to test the effects of syllable structure
  and speech rate on prenuclear H placement.
• Same 3 speakers of Castilian Spanish read twice
  16 target syllables in proparoxytone words, at a
  normal, fast and slow rates.
• A total of 288 utterances (16 utterances x 3
  speech rates x 3 speakers x 2 repetitions).

20
Speech materials

• The database consisted of a total of 16 test
  proparoxytonic words, divided in 2 groups open
  vs. closed syllables.

21
Effects of syllable duration

• High correlation (around 0,90) between syllable
  duration and H delay for the 3 speakers (exc 1
  case)
• Less delay in syllables with no coda

22
Effects of syllable structure

• Peaks more retracted into the syllable in closed
  syllables differences are statistically
  significant for the 3 speakers
• No strict anchoring at end of V0 or at end of
  syllable

23
No anchoring at syllable or vowel ends

• Peaks signif. displaced to the right in closed
  syllables (t-test significances at p lt 0,0001)
• No anchoring at end of V

24
Effects of speech rate

• As expected, speech rate (fast, normal, slow) had
  a significant effect on utterance length (in ms).
  ANOVA shows significant effect at 0,0001 for
  each speaker.

25
Effects of speech rate, 2

• Peaks are retracted as speech rate decreases, for
  the 3 speakers.
• ANOVAs show statistically significant effects of
  speech rate for 2 speakers

Patterns for English are the opposite (Ladd et al
1999) and probably Chinese (Xu 1998). .
26
Conclusion (Exp. 2)

• Exp. 2 also shows a loose alignment towards the
  end of v0 depending on the following factors
• Duration of the syllable of the syllable is
  positively correlated with H delay
• Syllable structure Closed syllables display a
  significant later alignment than open syllables.
• Speech rate peaks are retracted as speech rate
  decreases (fast gt normal gt slow).

27
General conclusion

• Strong version of the segmental anchoring
  hypothesis cannot be maintained.
• Prenuclear peaks in Spanish loosely anchored at
  end of V0, depending on the following factors
• Syllable structure
• Speech rate
• Coda type
• The 3 factors (together with syllable duration)
  significantly contribute to linear models of peak
  placement, sylldur and syllstruct being the most
  relevant (R2 80-90).

28
Part 2Effects of within-word position in H
placement in Catalan and Spanish
29
Catalan and Spanish prenuclear accents

• L anchored with onset of accented syllable
• H generally displaced to the postaccentual
  syllable (Prieto 1995 for Spanish and
  Estebas-Vilaplana 2000, 2003 for Catalan).

Broad focus statements

• Volen una nena La Marina vol
  demanar-lhi
• They want a girl Mary wants to ask him

30
Effects of word boundaries on H location in
Catalan and Spanish

• Recent experiments have observed that prenuclear
  peaks in both Catalan and Spanish strictly align
  with the word-boundary location (Arranz Garrido
  ms for Spanish and Estebas-Vilaplana 2000, 2003
  for Catalan).
• Estebas-Vilaplana (2000, 2003) analyzes Catalan
  prenuclear rises as sequences of a low pitch
  accent (L) plus a word edge tone (H) anchored at
  the end of the word.
• Other experiments show that peaks tend to shift
  backwards as their associated syllables approach
  the end of the word
• Silverman Pierrehumbert 1990 for English,
  Prieto et al 1995 and Prieto Torreira 2004 for
  Spanish Arvaniti, Ladd Mennen 1998 for Greek
  Ishihara in press for Japanese.
• Prieto et al. (1995) showed a significant effect
  of word position on peak delay in phrases such as
  número rápido, numero nervioso, numeró regular)
  where potential effects of clash were
  neutralized.

31
Goals of the study

• To clarify the role of word-edge tones in
  prenuclear accents of focus declaratives in
  Catalan and in Spanish
• Experiment 1. To verify the effects of
  within-word position on H peak placement and test
  the hypothesis that H is anchored at the right
  edge of the prosodic word.
• Experiments 2 and 3. To verify whether Catalan
  and Spanish listeners are able to use tonal
  alignment differences due to within-word position
  in the identification and discrimination of words
  with ambiguous word boundary location

32
Experiment 1

• MATERIALS 20 pairs of potentially ambiguous
  utterances which are only distinguished by word
  boundary location.
• CATALAN words with a pitch accent on the final
  and penultimate syllables.

Comprà ventalls (s)he bought fans w-fin Compra
ven talls they bought pieces w-med

• SPANISH Words with a pitch accent on the
  final, penultimate and
• antepenultimate syllables.

Ve bovinos s(he) sees cows fin Bebo
vinos s(he) drinks wines pen pen Sube
Melino Melino goes up pen Súbeme lino
Bring up the linen for me ante

• 3 speakers read the 20 pairs 4 times (40 x 4
  160 sentences per speaker, for a total of 480
  utterances for each language)

33
Comprà ventalls Compraven talls
Comprà ventallets de vim Compraven
tallets de vim
34
Labeling Scheme

• Key points manually placed
• Segmental landmarks (in ms)
• On the target accented syllable beginning of
  onset and vowel
• On the target postaccentual syllable beginning
  of onset, vowel, and coda (whenever present) end
  of the postaccentual syllable, end of the onset
  of the following syllable.
• Phrase-final point
• Pitch landmarks
• L1, valley of the first pitch accent (in Hz)
• H1, peak of the first pitch accent (in Hz)

35
Example of prenuclear LH
Waveform display, F0 contour, and labels
corresponding to the utterance Compraven talls
they bought pieces (speaker AG).
36
Results

• Strong correlation between H delay and syllable
  duration for the 3 speakers (correlation coef.
  0,67-0,82)
• Difference in H delay between two groups Hs in
  word-final position (e.g., comprà ventalls) are
  less delayed than in word-medial position (e.g.,
  compraven talls)

37
Spanish

• Difference in H delay between the three groups
  1) Hs in fin-accented words (e.g. ve bovinos)
  are less delayed than Hs in pen-accented words
  (e.g. bebo vinos) 2) Hs in pen-accented words
  (e.g. sube Melino) are less delayed than Hs in
  ante-accented words (e.g. súbeme lino).

Speaker EV
Speaker RA
Speaker TE
38
H Alignment

• All peaks displaced to the postaccentual syllable
• Clear effects of within-word position on H
  placement in the two languages peaks are less
  displaced in fin-accented words than in
  pen-accented and ante-accented words (longer H
  delay).
• Differences statistically significant for the 3
  speakers for the 2 languages (two-tailed t-tests
  significant at p lt 0,0005), except for pen/fin.

Catalan
Spanish
39

• No strict word anchoring effects found peaks are
  located before the end of the word (-72 ms) if
  the accented syllable is word-medial and after
  (48 ms) if it is word-final.
• Differences statistically significant for the 3
  speakers for the 2 languages separate ANOVAs for
  the 2 peak placement variables revealed a
  significant effect of word boundary in Catalan
  and ante vs pen/fin in Spanish.

Catalan
Spanish
40

• F0 alignment differences (Catalan)
• w-medial w-final
• compraven talls comprà ventalls

100 ms
60 ms
?
?
41
Duration patterns

• Accented syllables in word-final position are
  slighly longer than syllables in internal
  position, except for speaker AG.
• Separate ANOVAs for this duration variable
  revealed no significant effects of word boundary
  for none of the speakers (except for speaker AG).

Spanish
Catalan
42

• Postaccentual syllables in word-final position
  are slighly longer than syllables in word-initial
  position, except for speaker AG.
• Separate ANOVAs for this duration variable
  revealed no significant effects of word boundary
  for all of the speakers (except for speaker PP).
• No consistent effects of within-word position on
  duration patterns.

Catalan
Spanish
Similar findings for English which acknowledge
small duration effects in word-initial or
word-final position, mostly in accented syllables
(Turk White 1999, Turk Shattuck-Hufnagel
2000) .
43
Experiment 2 (Identification)

• GOAL Are differences in range and in peak
  location used by listeners to perceive
  differences in word-boundary location?
• MATERIALS 10 (for Catalan) and 12 (for Spanish)
  ambiguous utterances from the production test.
• SUBJECTS 12 listeners for Catalan and 20 for
  Spanish heard the ambiguous utterances a maximum
  of three times and had to identify the sentence.
•   Comprà ventallets de vim Compraven
  tallets de vim
• Nomenaves comtes al matí Nomenà vescomtes
  al matí
• Mirà batalles Mirava talles

44
Results Catalan

• Low rate of identification (between 35 and 66 of
  correct responses, depending on sentence).
• Type A utterances tend to be less easily
  identified than Type B utterances (60 vs 40).

Type A comprà ventalls Type B compraven
talls
45

• Why?
• The hearer identifies more easily Compraven talls
  (with an H aligned towards the end of the first
  word) than Comprà ventalls (with an H less
  displaced to the right).
• The hearer is probably using a more clear H
  alignment towards the edge of the word as a cue.
  In the second case, H placement is more
  ambiguous.
• A controlled perception experiment is needed.

• Less ambiguous More ambiguous
  
• ? ?

46
Results Spanish

• Variable rate of identification for Spanish
  sentences (25 to 95).

47
Experiment 3 (Discrimination)

• MATERIALS 10 (for Catalan) and 24 (for Spanish)
  ambiguous utterances in pairs from the production
  test.
• SUBJECTS the same listeners heard the utterances
  in pairs and had to chose the order in which the
  stimuli were heard.
•  

48
Results Catalan

• Variable rate of identification for Catalan (30
  to 90 of correct responses).

Catalan
49
Results Spanish

• Variable rate of identification for Spanish (40
  to 95).
• Better identification of Spanish sentences when
  heard in pairs (except for three cases).

Spanish
50
Results perception

• Results slightly support the hypothesis that fine
  allophonic details of H tonal alignment due to
  within-word position are employed by Catalan
  listeners in word identification tasks, even
  though the results are not clear-cut.
• The results of a recent perception study
  (DImperio, Petrone and Nguyen 2004) suggest that
  fine details of tonal alignment might help
  listeners in the identification of closed versus
  open syllables (and are helpful in the
  disambiguation between nono ninth and nonno
  grandfather.
• Results are more consistent in perceptual
  experiments with L points (Welby 2002, 2003 for
  French Ladd Schepman 2003 for English).
• Whether H alignment is crucially used in word
  disambiguation is a substantial empirical
  question that we leave as a matter for future
  research. A more controlled perceptual study is
  needed in order to better evaluate the perceptual
  import of this acoustic cue.

51
Conclusions

• No strict anchoring to word-edges.
• Clear effects of within-word position on H
  location H peaks are more retracted in
  word-final accents than in word-medial accents.
• No effects of within-word position on duration.
• Regression models of peak location reveal that
  (a) the fit of the model is better by using the
  measure peak delay than the measure distance to
  the end of the syllable and (b) that syllable
  duration is a significant factor in the model of
  peak delay, but not in the other model.
• H alignment towards word-edges might act as a
  helpful perceptual cue in disambiguating tasks.

52
General Conclusions

• The segmental anchoring hypothesis cannot be
  maintained for Catalan and Spanish
• Clear effects of syllable structure on H location
• Clear effects of coda type
• Clear effects of speech rate
• Clear effects of within-word position.
• TENTATIVE CONCLUSION 1 H placement is
  conditioned not only by tonal pressure, but also
  by prosodic domain adjustments (ie, syllables and
  prosodic words).
• TENTATIVE CONCLUSION 2 H placement is
  constrained by perceptual mechanisms (ie, the
  coda effect found in a variety of languages).

53

• MERCI BIEN.
• GRÀCIES!