Acoustic Properties of Canonical and Non-Canonical Stress in French, Turkish, Armenian and Brazilian Portuguese

Angeliki Athanasopoulou, Irene Vogel, Hossep DolatianInterspeech 2017

TL;DR

This paper investigates how predictable stress in languages affects the acoustic realization of stress. They found that stress is less clearly signaled when it's predictable (even in its canonical position) and more clearly signaled when it's unpredictable, challenging the idea that canonical positions inherently have stronger cues.

What Problem Does This Paper Address?

Languages are often categorized by how predictable their word stress is. Some languages have "fixed" stress, where stress reliably falls on a particular syllable (e.g., the last one). Others have "non-predictable" stress, where stress can fall on different syllables, and this variation can even distinguish words (minimal pairs). The "Smooth Signal Redundancy Hypothesis" suggests that when linguistic information is less predictable, speakers add extra acoustic cues to make it clearer. This paper explores whether this hypothesis applies to word stress. Specifically, they ask: does the predictability of stress position influence how clearly stress is acoustically marked? They focus on languages with "fixed" stress that also have exceptions (French, Turkish, Armenian) and compare them to a language with non-predictable stress (Brazilian Portuguese).

Methods

The researchers analyzed acoustic properties of stressed and unstressed vowels in real words produced by native speakers of French, Turkish, Armenian, and Brazilian Portuguese. They focused on four key acoustic features:

  • Duration: How long the vowel lasts.
  • F0 (Fundamental Frequency): The perceived pitch of the voice. They looked at the overall mean F0 and the change in F0 across the vowel (ΔF0).
  • Intensity: The loudness of the vowel.
  • Vowel Centralization: How close the vowel's sound is to the center of the vowel space.

They used Binary Logistic Regression Analyses (BLRAs) to determine how well these acoustic features could distinguish between stressed and unstressed vowels, and also between vowels in canonical (expected) and non-canonical (exceptional) stress positions. The "clarity" or "robustness" of stress manifestation was measured by the percentage of correctly classified vowels.

Datasets

The study used a dataset of recorded speech from 10 native speakers of each of the four languages (Turkish, French, Armenian, and Brazilian Portuguese). The stimuli consisted of real trisyllabic words, with target vowels in open syllables. For each language, two sets of words were created: one with stress in the canonical position and another with stress in a non-canonical position. The specific non-canonical positions were chosen based on known exceptions or common patterns in each language (e.g., using specific suffixes in Turkish and Armenian, or exploiting contexts that favor schwa production in French).

Key Insights and Ideas

  • Canonical position doesn't guarantee stronger cues: Contrary to what one might expect, the canonical stress position in "fixed" stress languages did not consistently show clearer acoustic cues than non-canonical positions. This suggests that simply being in the "expected" spot doesn't automatically make stress more prominent.
  • Unpredictability is key to clarity: The most significant finding is that stress is acoustically more robustly manifested when its position is unpredictable. This aligns with the Smooth Signal Redundancy Hypothesis.
  • French as an interesting case: French, with its less predictable non-canonical stress, showed an increase in stress distinction in these non-canonical positions compared to its canonical final stress. This is similar to Portuguese, where stress is generally unpredictable.
  • "Stress Deafness" linked to predictability, not just fixed stress: The results suggest that the phenomenon of "stress deafness" (difficulty perceiving stress) might be more closely tied to the overall predictability of stress in a language, rather than simply having a fixed stress system with exceptions. Turkish speakers, who have predictable exceptions, showed similar stress perception patterns to French speakers, who have less predictable exceptions.

Key Results

  • Canonical vs. Non-Canonical Distinction: In the "fixed" stress languages (Turkish, Armenian, French), the overall classification rates distinguishing canonical from non-canonical stressed vowels were relatively weak (around 62-68%). This indicates that the acoustic manifestation of stress is not drastically different between these positions.
  • Stressed vs. Unstressed Distinction:
    • In canonical stress positions, the distinction between stressed and unstressed vowels was moderately clear across languages (71-81%).
  • In non-canonical stress positions:
  • Turkish and Armenian (predictable non-canonical stress) showed a slight decrease in classification rates compared to their canonical positions.
  • French (less predictable non-canonical stress) and Portuguese (unpredictable stress) showed an increase in classification rates in their non-canonical positions.
  • Main Acoustic Correlates: F0 (mean and contour) and intensity were consistently important acoustic cues for stress distinction across languages. Duration also played a role in some cases.

Limitations and Future Work

  • Stimuli Limitations: The study used specific types of trisyllabic words and focused on particular phonological contexts for non-canonical stress. The authors acknowledge that French targets with intended penultimate stress were excluded if the final schwa was not produced, which might affect the French results.
  • Cross-linguistic Comparison Nuances: While patterns were observed, the authors note that direct comparison of specific percentage rates across languages is not straightforward due to differences in stimuli and the nature of stress in each language.
  • Future Directions: The paper suggests that future research could further explore the interplay of predictability, stress, and other prosodic phenomena like focus. The findings also have implications for understanding cross-linguistic differences in stress perception and acquisition.

Summary updated

BibTeX
@inproceedings{athanasopoulou2017acoustic,
  author = {Angeliki Athanasopoulou and Irene Vogel and Hossep Dolatian},
  title = {{Acoustic Properties of Canonical and Non-Canonical Stress in French, Turkish, Armenian and Brazilian Portuguese}},
  booktitle = {Interspeech},
  year = {2017},
}

Loading related content...

Figure