V-to-V Coarticulation Induced Acoustic and Articulatory Variability of Vowels: The Effect of Pitch-Accent

Andrea Deme, Márton Bartók, Tekla Etelka Gráczi, Tamás Gábor Csapó, Alexandra MarkóInterspeech 2019

TL;DR

This study investigates how pitch accent and vowel quality influence coarticulation between vowels in Hungarian. By simultaneously analyzing acoustic and articulatory data, the researchers found that pitch accent affects vowel realization differently for high front (/i/) and high back (/u/) vowels, challenging previous assumptions about "coarticulatory resistance."

What Problem Does This Paper Address?

Speech sounds are not produced in isolation; they influence each other, a phenomenon known as coarticulation. Vowel-to-vowel (V-to-V) coarticulation, where a vowel's realization is affected by surrounding vowels, is a significant source of variability in speech. Researchers have proposed the concept of "coarticulatory resistance" (CR) to describe how much a vowel resists this influence (higher resistance means less variability). Previous studies suggested that pitch accent might increase CR, making vowels more resistant to coarticulation. Additionally, there's been debate about whether certain vowel qualities (like high vowels /i/ and /u/) are inherently more resistant than others. However, much of this research has relied on either acoustic or articulatory data alone, and the relationship between these two domains is complex. This paper aims to bridge this gap by simultaneously examining acoustic and articulatory data to understand how pitch accent and vowel quality interact to influence V-to-V coarticulation and vowel variability in real Hungarian words.

Methods

The researchers employed a novel approach by collecting parallel acoustic and articulatory data from nine Hungarian adult female speakers. They designed specific VCV (Vowel-Consonant-Vowel) contexts using minimally constrained alveolar consonants and the glottal fricative /h/ to maximize V-to-V effects. The target vowels were embedded in real Hungarian words within sentences, and two accent conditions were tested: one with sentence-level pitch accent on the target vowel and one without.

  • Acoustic Analysis: They measured the first two formants (F1 and F2) of the target vowels at their onset and midpoint using the Praat software. The "locus equation" approach was used to quantify the degree of coarticulation by fitting linear models to F2 onset and midpoint data. Vowel variability was measured by the standard deviation of F2 midpoint values.
  • Articulatory Analysis: An Electromagnetic Articulography (EMA) system was used to track the movement of the tongue body (using sensors TBO1 and TBO2), lips, and jaw. Horizontal displacement of the tongue dorsum was analyzed. Similar to the acoustic analysis, locus equations were used to model the relationship between dorsum position at onset and midpoint. The degree of coarticulation was assessed by comparing dorsum positions in coarticulating vs. non-coarticulating contexts. Vowel variability was quantified by the standard deviation of the horizontal dorsum displacement.
  • Statistical Analysis: Linear mixed-effects models were used to analyze the data, accounting for speaker variability.

Datasets

The study used a custom-designed dataset of real Hungarian words embedded in sentences, specifically constructed to elicit V-to-V coarticulation effects. The dataset included 212 analyzed tokens after excluding mispronounced ones. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Acoustic vs. Articulatory Domains: The study strongly highlights that acoustic and articulatory data do not always align, suggesting that conclusions drawn from one domain alone can be misleading. For instance, while acoustic data might show more variability in one vowel, articulatory data might reveal a different pattern.
  • Pitch Accent's Nuanced Effect: Pitch accent does influence vowel realization, but its effect is not uniform. It impacts vowels differently based on their quality and the specific domain (acoustic vs. articulatory).
  • Challenging "Coarticulatory Resistance": The findings suggest that the concept of "coarticulatory resistance" might need re-evaluation. The study observed patterns that don't fit a simple model of increased resistance under accent for all vowels.
  • Vowel Quality Matters: High front (/i/) and high back (/u/) vowels exhibit distinct behaviors under coarticulation and pitch accent, indicating that vowel quality is a crucial factor in understanding these phenomena.

Key Results

  • Locus Equations: Locus equations showed steep slopes for /i/ and flatter slopes for /u/ in both acoustic and articulatory domains, indicating that /i/ was produced more statically over time than /u/.
  • Acoustic Coarticulation: Pitch accent influenced F2 onset differences. For /u/, coarticulation was suppressed in accented syllables but present in unaccented ones. For /i/, accented syllables showed a "fronting" effect in coarticulating contexts acoustically, while unaccented syllables showed a "backing" effect.
  • Articulatory Coarticulation: In accented syllables, the articulatory difference between coarticulated and non-coarticulated vowels was similar for both /i/ and /u/. However, in unaccented syllables, /u/ showed a greater coarticulatory effect than /i/.
  • Acoustic Variability: /i/ tokens were generally more acoustically variable than /u/ tokens, especially in accented conditions. However, in unaccented, non-coarticulating contexts, /u/s varied more than /i/s.
  • Articulatory Variability: In contrast to acoustic findings, articulatory variability (dorsum displacement) was generally higher in /u/ than in /i/, particularly in accented syllables. This suggests that the acoustic stability of /u/ might be achieved through larger articulatory movements.

Limitations and Future Work

  • Defining Coarticulatory Resistance: The authors raise two critical questions for future research:
  • How should coarticulatory resistance be defined? Should it be viewed as decreased acoustic/articulatory variability, or as an increased capacity for dissimilation/adaptation?
  • What is the primary domain of CR? Should it be measured acoustically or articulately, or is it a combination of both?
  • Domain of Measurement: The paper emphasizes the difficulty in comparing acoustic and articulatory data directly due to their non-linear relationship. Future work could focus on developing more integrated models that account for this relationship.
  • Cross-linguistic Comparison: While this study focused on Hungarian, future research could explore whether these findings generalize to other languages with different prosodic systems and vowel inventories.

Summary updated

BibTeX
@inproceedings{deme2019v,
  author = {Andrea Deme and Márton Bartók and Tekla Etelka Gráczi and Tamás Gábor Csapó and Alexandra Markó},
  title = {{V-to-V Coarticulation Induced Acoustic and Articulatory Variability of Vowels: The Effect of Pitch-Accent}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure