Introduction

The question

Are the choice of intonation contour and the form of the contour independent of the words used?​

The Autosegmental-Metric Theory

• The assignment of intonation is autonomous from the segmental level

  • The separation of the lexical and the tonal level

  • Realization of pitch contours is determined by phonetic rules

    • (refering to tonal sequences and metrical organization)

  • At odds with the idea of storing acoustic details with lexical items

The Exemplar Theory

• Detailed episodic storage

• Frequent units lead to many exemplars

• Frequent units display less variation

  • “entrenchment” (Lee et al., 1999; Pierrehumber, 2001)
• Present study: lexical and word+accent frequency

Exemplar models and prosody/intonation

• Current exemplar models do not take account of intonation

• Frequency effects on prosody

  • acquisition of the prosodic word (Vigario et al., 2006)

  • word stress assignment (Daelemans et al., 1994)

  • predictability of syllable duration (Schiweitzer & Mobius, 2004)

Corpus Analysis

Parametrisation of pitch accent shape

(The PaIntE model)

• a1 and a2: steepness of rise and fall

• b: alignment of the peak
• c1 and c2: amplitude of the falling and rising sigmoid
• It is also possible to use one sigmoid only (c)
• d: height of the peak

Experiment 1

absolute frequency of pitch accent + word

Data

• Annotated part of the DIRNDL-Corpus, German radio news broadcasts (5 hours and 16 min)

• 7871 L*H and 6118 H*L tokens 

• For each word type, frequencies of combination with L*H and H*L were calculated
  • outliners for any PaIntE parameters were removed
  • patterns that are not "clear" were removed

Results of Plausibility checks

Methods

• Dependent variable: accent range (c1 and c2)

• Fixed Factors:
  • frequency of word+accent pairs (logged)
  • number of accent to the next IP boundary
  • coda size (number of segments)
  • onset/coda classification: -V, +S, +V-S

• Random intercepts and slopes for word and speaker

• Linear mixed model with likelihood ratio tests
  • add fixed factors separately, only keep the significant ones

Result for H*L

The range is increased by approx. 0.89 Hz for each unit increase in logged frequency of occurrence

Result for L*H

• higher word+PA frequency, bigger range
• sonorant coda, bigger range
• closer to the boundary, bigger range

Experiment 2

relative frequency of pitch accent + word

Methods: Accent variability

• Statistical tests: as in Exp 1
  • New fixed factor: relative frequency: "how often is does the word type bear the accent?"

• Calculating the variability of realization
  1. z-scoring the PaIntE parameters for speakers and accent types
  2. The vector for each token = {a1, a2, b, c1, c2, d}
  3. For each token, the vector's distance with other tokens of the same type (word+PA) was calculated
  4. Average distance = variability

Methods: Accent variability

• For example, for a token "Porsche"+H*L
  • Variability (DV) = it's average distance with all other tokens for  "Porsche"+H*L
  • Absolute frequency (IV) = type frequency of  "Porsche"+H*L
  • Relative frequency (IV) = 

Result for H*L

higher rel.freq, lower variability

higher abs.freq, higher variability

sonorant coda, lower variability

Result for L*H

• Result: higher abs.freq, higher variability

Experiment 3

relative frequency of word in lexical context

Data

• A subset of the Switchboard corpus (6 hours)

  • PA location and boundary location were labeled.

• Two datasets were extracted
  • prosodic pattern dataset
    • trigrams with a frequency > 4
  • pitch accent variability dataset
    • trigrams with a PA on the middle word

• Relative frequency: 

Prosodic pattern variability: Methods

• Prosodic pattern: two variables
  • NoAcc vs. Acc
  • NoBound vs. Bound
  • e.g. NoAcc-NoBound—Acc-NoBound—NoAcc-Bound.

• Variability:

  • Most common pattern for a word in any trigram (e.g., lot)

  • For each trigram with the middle word (lot), determine whether it has that common pattern

    • This is the binary dependent variable

Prosodic pattern variability: Methods

• For example: for the trigram: "a lot of"
  • find the most frequent prosodic pattern for "lot" in all kind of trigrams when it's the middle word
  • " NoAcc-NoBound—Acc-NoBound—NoAcc-Bound"
  • For each token of "a lot of", determine if it has the pattern (DV)

• Independent variables:
  • absolute frequency of "lot"
  • relative frequency:  "a lot of" divided by "lot"

Prosodic pattern variability: Results

Pitch accent variability: Methods

• Calculation: same as Exp. 2
  1. z-scoring the PaIntE parameters for speakers
  2. The vector for each token (middle word) = {a1, a2, b, c1, c2, d}
  3. For each token, the vector's distance with other tokens of the same type was calculated
  4. Average distance = variability

• For example, for a token of "a lot of"
  • Variability (DV) = it's average distance with all other tokens for "lot" in  "a lot of"
  • Absolute frequency (IV) = type frequency of  "lot"
  • Relative frequency (IV) = "a lot of" divided by "lot"

Pitch accent variability: Methods

Pitch accent variability: Results

• higher rel.word.freq, lower variability

Discussion

and Conclusion

PA+word freq and Accent range

• PA's F0 range increases as PA+word's frequency increases

• The relationship between tonal and word levels has no obvious explanation in AM theory

• Theories of episodic storage are able to explain this

  • linguistic units as concrete and highly specified instances, containing properties of the tonal and the lexical

• Why greater range with greater frequency?
  • choosing the exemplars meeting communication goals
  • PA = prominence

PA+word rel.freq and PA variability

• Less variability with increasing relative frequency (H*L)

  • if a word if often realized with a particular accent type (communicative function), then the set of exemplars will be more homogeneous

• More variability with increasing absolute frequency 

  • Greater absolute frequency = greater variability because tokens are coming from all contexts

Trigram frequency, prosodic context, PA realization

• As relative trigram frequency increases...

  • more homogeneous prosodic context (more likely to be the most common pattern)

  • less variability for pitch Accent realization

• Evidence for cohesion between the word and its prosodic realization

• Explained with the exemplar theory:

  • tonal events are stored with lexical sequences

  • words collocate together will be stored together and acquire particular phonetic characteristics

Other Discussions

• Word and PA tokens separately with co-index?

• Pitch contours and words together?
  • This explains the results of entrenched intonation (less variability) better

• Effects too subtle?

  • intonational information is very uneven across the lexicon, so effects do not show well across words

Conclusion

• This study demonstrates effects on intonation that
  should be considered in exemplar-theoretic models.

• Entrenchment should be modelled for tonal parameters.

• Frequency of occurrence effects are
  acknowledged