A comparative account of the suprasegmental and rhythmic features of British English dialects

1

• A comparative account of the suprasegmental and
  rhythmic features of British English dialects
• Emmanuel Ferragne, François Pellegrino
• Laboratoire Dynamique Du Langage
• UMR CNRS 5596
• LYON, FRANCE

2
Goals

• Characterize the 14 dialects in the Accents of
  the British Isles database in terms of rhythm.
  Correlates of rhythm studied here
• Metrics based on the duration of vocalic and
  consonantal intervals
• Spectral vowel reduction

3
The Accents of the British Isles database

• Recorded 2003
• Dialects 14
• Speakers 20 (10 ?, 10 ?) per dialect
• Read passage appr. 290 words
• Uncontrolled factors
• age, social class, education
• speech rate, dysfluencies

4
Rhythm studies in multilingual contexts some
landmarks

• 1940-70s
• Notion of rhythm classes (stress-timed
  /syllable-timed) popularized by Pike (1945) and
  Abercrombie (60s).
• Syllabic structure as a correlate of "phonic
  impression" (rhythm?) (Delattre et al. 1969).
• 1980s
• Objective isochrony rejected for good (Roach
  1982, Dauer 1983).
• 1990s
• Introduction of duration-based metrics that
  mainly capture syllable properties 2 commonly
  used metrics
• 1) Method introduced by Ramus et al. (1999)
• 2) Method introduced by Low and Nolan (late
  90s), then Grabe and Low (2002)

5
Durational correlates of rhythm in the world's
language

• Grabe, E. and Low, E.L. (2002). Durational
  Variability in Speech and the Rhythm Class
  Hypothesis. Papers in Laboratory Phonology 7,
  Mouton.

6
The effect of speech rate

• D

• Dellwo, V. et al. (2004). The BonnTempo-Corpus
  BonnTempo-Tools Proceedings of the 8th ICSLP,
  Jeju Island, Korea.
• Tempo has more effect on deltaC than on V,
  except for Italian.
• Intended speech rate corresponds to various
  actual speech rates (syllables/second) depending
  of the language

7
Speech and music
Patel, A.D., et al. (2004). Comparing rhythm and
melody in speech and music The case of English
and French. JASA, 1162645
8
Why apply these methods to British English
dialects?

• The literature says rhythmic differences exist
  but no empirical evidence yet
• Vowel
• little (if any) contrastive vowel length in
  Ireland and Scotland
• "boats are stronger and more stable"
• Southern Standard 185 ms
• Scottish Highlands 100 ms
• Some dialects (North Celtic countries) resist
  vowel reduction in unstressed syllables or have
  relatively peripheral realizations of unstressed
  vowels

9
Method

• Following passage manually segmented into vowels
  and consonants
• "boats are stronger and more stable, protecting
  against undue exposure. Tools and instruments are
  more accurate and more reliable, helping in all
  weather and conditions."
• Successive segments of the same type (vowel or
  consonant) merged into one single interval
• Caveat
• velarized L (syllabic or not) often segmented as
  a vowel (it is even vocalized in some dialects),
• final R ranges from nothing ? r-colouring ? full
  apical approximant (or flap) often merged with
  preceding vowel in segmentation
• devoiced, or fricated syllable nuclei count as
  vowels inSTRUments

10
Computation

• meanV mean duration of vocalic intervals
• meanC mean duration of consonantal intervals
• V percentage of vocalic duration over the whole
  passage
• ?C standard deviation of consonantal interval
  duration
• ?V standard deviation of vocalic interval
  duration
• varcoC ?C expressed as a fraction of mean
  consonantal interval duration
• varcoV ?V expressed as a fraction of mean
  vocalic interval duration
• mean_rpviv mean (over a whole passage) of the
  difference in duration between two consecutive
  vocalic intervals
• mean_npviv same as above except that the
  duration difference for each pair is divided by
  the sum of the duration of the two vocalic
  intervals divided by two
• med_rpviv and med_npviv same as previous two,
  except that the median, instead of the mean, is
  used
• the consonantal counterparts of the PVI measures
  just mentioned were also computed
• SR speech rate in syllables per second.

11
Results

• ?V maximizes between-/within-dialect variance
  ratio
• However, mean speech rate differs across
  dialects no satisfactory explanation most
  likely due to unbalance in literacy,
  dysfluencies)
• Therefore tempo effect factored out using
  varcoV probably the best gauge

ULSTER
ULSTER
BIRMINGHAM
NEWCASTLE
BIRMINGHAM
EAST YORKSHIRE
12
Vowel spectral reduction (centralization)

• Extreme cases
• Singapore English
• Taiwan English

E-L. Low, E. Grabe, F. Nolan. (2000).
Quantitative characterizations of speech rhythm
syllable-timing in Singapore English. Language
Speech 43, 377-401.
Jian, Hua-Li. (2004). An Acoustic Study of Speech
Rhythm in Taiwan English. Interspeech-ICSLP,
Jeju, Corée.
13
Method and computation

• F1 and F2 values in Bark at temporal midpoint for
  all the vowels
• Normalization for gender was achieved by
  subtracting 1 Bark from F1 and F2 in women
• For each speaker position of the F1/F2 centroid
  and the unweighted Euclidean distance between
  each vowel and the centroid.
• Whenever the distance between a token and the
  centroid was above the 95th percentile, the token
  was removed in order to discard outliers and new
  values for the centroid and the individual
  distances were calculated.
• Each distance in a given speaker was divided by
  the greatest distance observed for this speaker
  as a means to normalize for differences in vowel
  space size across speakers.
• The mean distance, the mean normalized distance,
  the standard deviation of normalized distances,
  and the skewness of normalized distances for each
  speaker were computed.

14
Results

• One-way ANOVA for differences in mean skewness
  across dialects F 2.93, df 13, p lt 0.001,
  post hoc tests reveal differences (p lt 0.05)
  between pairs of dialects. However not
  supportive of our hypotheses.
• Differences in gender more salient.

15
Discussion

• The duration measurements examined here capture
  part of the rhythm differences predicted by
  traditional dialect studies. The best indicator
  seems to be varcoV.
• Focus in rhythm studies must shift to other
  parameters Fo, intensity, vowel
  centralization and voice quality?
• Here, we decided speech rate variation given
  the presumably "levelling" effect of a reading
  task - was due to chance this may not hold for
  spontaneous speech (Birmingham).
• On women using more centralized reduced vowels
  total agreement with the fact that women are more
  prone to using prestige/standard form of a
  language (Labov, Trudgill, etc.)
• The portion of the corpus may not have been long
  enough to guarantee enough occurrences of
  dialect-specific features

16
Take-home message

• Time to move on to more correlates of rhythm
  duration is only half the story