Phonetic imitation of Japanese vowel devoicing

Kuniko NielsenInterspeech 2010

TL;DR

This study investigates whether speakers unconsciously imitate subtle phonetic details, specifically Japanese vowel devoicing, even when it's phonologically constrained. The findings show that imitation does occur, suggesting that our brains retain and can reproduce fine-grained phonetic information, influencing speech production beyond just phonemic contrasts.

What Problem Does This Paper Address?

The paper delves into the fascinating phenomenon of phonetic imitation, where speakers unconsciously mimic the speech sounds they hear. Previous research has established that this imitation isn't always automatic; it can be modulated by factors like phonemic contrast (e.g., distinguishing between voiced and voiceless sounds). However, a key question remains: do speakers also imitate phonetic details that are governed by less critical phonological rules, or that don't threaten phonemic distinctions? This paper specifically explores Japanese vowel devoicing, a process where high vowels are often weakened or disappear between voiceless consonants. This phenomenon is interesting because it's considered a gradient phonetic process, not strictly tied to phonemic contrast, making it a good test case for understanding the boundaries of phonetic imitation. The authors want to know if imitation occurs even when phonological factors are at play, which could shed light on how phonetic information is stored and accessed.

Methods

The researchers employed a word-naming imitation paradigm, similar to previous studies. Participants were first recorded speaking a list of Japanese words (baseline). Then, they listened to recordings of specific words (the "listening list") that featured extreme Japanese vowel devoicing. Finally, they were asked to read the original word list again (post-exposure recording).

The core of the experimental manipulation involved the feature [+ spread glottis] on the phoneme /u/. The stimuli were carefully constructed to include:

  • Target words: Words from the production list that were also in the listening list.
  • Novel words: Words from the production list that were not in the listening list, used to test generalization.
  • Filler words: Words designed to avoid devoicing environments.

Acoustic analysis focused on measuring the duration of voiced vowels, consonants, and whole words. Two types of measurements were used to quantify devoicing:

  1. Gradient measure: The duration of vowels that were still voiced but weakened.
  2. Categorical measure: The rate of complete vowel deletion or devoicing.

Statistical analyses, primarily repeated-measures ANOVAs, were used to compare baseline and post-exposure productions across different conditions (e.g., target vs. novel words, different vowels, lexical frequency).

Datasets

The study used a custom-constructed dataset of 110 Japanese words. These words were categorized into:

  • 30 Target words (used in the listening phase).
  • 60 Novel words (not used in the listening phase, for generalization tests).
  • 20 Filler words (to control for baseline production).

Lexical frequency and familiarity were controlled, with words categorized as low or high frequency based on the NTT database. The stimuli were produced by a phonetically trained native speaker of Tokyo Japanese.

Key Insights and Ideas

  • Imitation is Robust: Participants unconsciously imitated the extreme degree of Japanese vowel devoicing they heard, even though it was a subtle phonetic detail. This suggests that fine-grained phonetic information is readily captured and reproduced.
  • Generalization Beyond Specific Words: The imitation wasn't limited to the exact words heard. Participants also showed imitation for novel words containing the same vowel (/u/) and even for a different vowel (/i/), indicating generalization at a subphonemic level.
  • Phonetic vs. Phonological Accounts: The observed patterns (e.g., partial devoicing, longer preceding consonants) align better with a gradient phonetic account of vowel devoicing rather than a strict phonological assimilation rule.
  • Word Specificity Exists, But is Nuanced: Imitation was found to be somewhat word-specific, with greater imitation for target words compared to novel words in the gradient analysis. However, this specificity was not observed in the categorical analysis.
  • Lexical Frequency Doesn't Drive Imitation: Contrary to some previous findings, lexical frequency did not significantly influence the degree of phonetic imitation in this study.

Key Results

  • Gradient Imitation: The mean duration of devoiceable vowels decreased significantly from 53 ms (baseline) to 44 ms (post-exposure).
  • Categorical Imitation: The rate of complete devoicing increased by 5% from baseline to post-exposure production.
  • Phoneme-level Generalization: Imitation occurred for both Target /u/ and Novel /u/ words, though it was stronger for Target words.
  • Subphonemic Generalization: Imitation was observed for Novel /i/ as well as Novel /u/, demonstrating generalization beyond the specific vowel modeled.
  • Word Specificity (Gradient): Imitation was greater for Target words than Novel words in the gradient analysis.
  • No Lexical Frequency Effect: Neither gradient nor categorical analyses showed a significant effect of lexical frequency on imitation.

Limitations and Future Work

  • Individual Variability: The study noted significant individual variability in the degree of imitation, suggesting that not all speakers are equally prone to this type of accommodation.
  • Limited Scope of Vowels: While generalization to /i/ was observed, the primary focus was on /u/, and the study did not deeply explore imitation across a wider range of vowels or other phonetic features.
  • No Perceptual Assessment: The baseline productions were not perceptually assessed, which could have provided additional context for understanding individual differences.

Future work could explore the role of other phonological factors, investigate imitation in different languages, and delve deeper into the cognitive mechanisms underlying this selective phonetic accommodation.

Summary updated

BibTeX
@inproceedings{nielsen2010phonetic,
  author = {Kuniko Nielsen},
  title = {{Phonetic imitation of Japanese vowel devoicing}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure