Acoustic correlates of phrase-internal lexical boundaries in dutch

Taehong Cho, Elizabeth K. JohnsonInterspeech 2004

TL;DR

This study investigates how Dutch speakers acoustically mark boundaries between words within a phrase, a subtle but crucial aspect of speech. By analyzing recordings of phonemically identical word pairs, the researchers found that speakers use a combination of durational and amplitude cues to signal these internal boundaries, with specific acoustic features like RIME1, RIME2, and C3 duration being particularly predictive.

What Problem Does This Paper Address?

The ability of listeners to segment a continuous speech stream into meaningful units (words) is fundamental to understanding spoken language. While it's well-established that speakers use acoustic cues to mark larger phrase boundaries (like pauses or specific intonation patterns), it's less clear how they signal the subtler boundaries within a phrase, especially when those boundaries don't align with major prosodic breaks. This is particularly challenging for listeners when faced with phonemically ambiguous sequences. Previous research has yielded mixed results, with some suggesting these internal boundaries are acoustically attenuated or inconsistent. This paper aims to systematically investigate whether Dutch speakers reliably produce acoustic cues for phrase-internal lexical boundaries and, if so, what those cues are.

Methods

The researchers designed an experiment using phonemically identical two-word Dutch strings that had the same syllable structure and stress pattern (strong-weak-strong, SWS) but differed in their intended word boundary. For example, "reis # pastei" (trip # pastry) versus "reispas # tij" (passport # tide). Six native Dutch speakers read sentences containing these target strings, with the sentential position of the targets varied to prevent repetitive strategies.

Acoustic analysis focused on several parameters:

  • Durational measures: Rime duration for each syllable (RIME1, RIME2, RIME3) and closure duration for stop consonants in the second and third syllables (C2DUR, C3DUR).
  • Voicing measures: Voice Onset Time (VOT) for stop consonants (C2VOT, C3VOT).
  • Amplitude measures: Peak amplitude during vowels (V1AMP, V2AMP, V3AMP) and peak burst amplitude for stop consonants (C2BURST, C3BURST).
  • Spectral analysis: Center of gravity (COG) for the fricative [s] in a specific string ("reis-pas-tij") to examine coarticulation effects.

Statistical analyses included repeated measures ANOVAs to identify acoustic parameters significantly affected by boundary location (S#WS vs. SW#S) and binary logistic regression to determine which parameters were most predictive of the boundary type.

Datasets

The study used a custom dataset of recorded Dutch sentences. Four phonemically identical target strings were created, each with two possible word boundary interpretations. These targets were embedded in six different sentence contexts. A total of 1153 tokens were initially collected, with 1065 tokens remaining after exclusions due to unclear accentuation or connectedness. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Subtle but present marking: Despite the absence of major prosodic cues (like pauses or boundary tones), Dutch speakers do consistently signal phrase-internal lexical boundaries using fine-grained acoustic details.
  • Acoustic "toolkit": Speakers employ a suite of acoustic parameters, not just one, to mark these boundaries. The study identified eight such parameters.
  • Durational cues are key: While multiple cues are used, durational measures (specifically RIME1, RIME2, and C3 duration) emerged as the most reliable predictors for distinguishing between the two boundary types.
  • Coarticulation matters: The spectral analysis of the [s] sound revealed that its articulation is influenced by the presence or absence of a lexical boundary, suggesting that coarticulation with the following consonant is more constrained across a boundary. This implies that the acoustic realization of sounds can be modulated by the linguistic structure they are part of.
  • Prosodic principles at play: Many of the observed acoustic differences can be explained by established prosodic phenomena like domain-initial strengthening, domain-final lengthening, and accentual lengthening, even at this fine-grained, phrase-internal level.

Key Results

  • ANOVA findings: Significant differences were found in eight acoustic parameters between S#WS (boundary after first syllable) and SW#S (boundary after second syllable) conditions. For example, C2 DURATION and RIME1 DURATION were longer in S#WS, while C2 VOT was shorter. RIME2, RIME3, and C3 DURATION were shorter in S#WS compared to SW#S. V2 AMP and C3 BURST were also smaller in S#WS.
  • Logistic Regression: A binary logistic regression model using the eight significant parameters achieved approximately 70% accuracy in classifying the boundary type.
  • Predictive power: RIME1 DURATION, RIME2 DURATION, and C3 DURATION were identified as the most reliable predictors, contributing most significantly to the classification.
  • [s] coarticulation: The Center of Gravity (COG) of [s] in "reis-pas-tij" was significantly higher when it was word-internal (followed by [t] without a boundary) compared to when it was word-final (followed by a boundary).

Limitations and Future Work

  • Listener perception: While the study demonstrates that speakers produce these acoustic cues, it doesn't directly prove that listeners are sensitive to all of them, or which cues are weighted most heavily in perception. Future perceptual studies are needed to address this.
  • Generalizability: The study focused on Dutch and specific SWS strings. Further research could explore whether these findings generalize to other languages and different phonetic contexts.
  • Complexity of cues: The authors acknowledge that pinpointing the exact source of variation for each acoustic parameter is complex, as they can be influenced by multiple phonetic events. Future work could aim to disentangle these effects more precisely.
  • Subphonemic detail: The paper highlights the importance of "non-contrastive fine-grained phonetic details" in marking boundaries. Future work should investigate how salient these subphonemic cues are and their role in achieving communicative goals.

Summary updated

BibTeX
@inproceedings{cho2004acoustic,
  author = {Taehong Cho and Elizabeth K. Johnson},
  title = {{Acoustic correlates of phrase-internal lexical boundaries in dutch}},
  booktitle = {Interspeech},
  year = {2004},
}
Figure