Prosodic Effects on Plosive Duration in German and Austrian German

Barbara Schuppler, Margaret ZellersInterspeech 2019

TL;DR

This study investigates how prosodic boundaries and word stress affect the duration of plosive sounds in German and Austrian German. The researchers found that while both varieties use duration to signal these prosodic features, they employ different acoustic cues (closure vs. burst duration) and strategies, highlighting the importance of variety-specific analysis for speech technology.

What Problem Does This Paper Address?

Understanding how speakers signal linguistic structure through subtle phonetic variations is crucial for both linguistic theory and the development of speech technologies like automatic speech recognition (ASR) and text-to-speech (TTS). Prosodic boundaries (like phrase endings) and word stress are known to influence speech sounds, often leading to changes in duration, intensity, and articulation. However, the specific acoustic cues used to mark these prosodic features can vary significantly across different languages and even dialects. This paper addresses the question of whether German and Austrian German, two distinct varieties of the same language, use the same acoustic strategies to mark prosodic boundaries and word stress, particularly focusing on the production of plosive consonants (like /p/, /t/, /k/). Previous research has shown that plosives themselves have a lenis-fortis distinction (e.g., /b/ vs. /p/) that is marked by acoustic cues like closure duration and burst characteristics, and this study aims to see how prosodic factors interact with this phonological distinction and how these interactions differ between the two German varieties.

Methods

The researchers employed a quantitative acoustic analysis using linear mixed-effects regression models. They extracted acoustic measurements from plosive consonants and their surrounding segments in read speech. The dependent variables were various duration measurements (total plosive duration, closure duration, burst duration). Independent variables included:

  • Prosodic Context: Presence or absence of a prosodic boundary, word stress.
  • Phonological Contrast: The plosive consonant itself (/p/, /b/, /d/, /t/, /g/, /k/) and its fortis/lenis status.
  • Segmental Context: Manner of articulation and voicing of preceding and following segments.
  • Speaker and Text Factors: Speaker identity, word class, speech rate (global and local), sex, and language variety (German vs. Austrian German).

Models were built using stepwise backward selection, with non-significant factors and interactions removed based on AIC values and model comparisons.

Datasets

The study utilized two existing corpora of read speech:

  • Kiel Corpus of Spoken German: This corpus provided data for Northern German. It was fully manually annotated phonetically and prosodically.
  • GRASS (Graz Corpus of Read and Spontaneous Speech): This corpus provided data for Austrian German. It was automatically segmented using MAUS, with manual correction for plosive segmentation and prosodic boundary annotation.

The study analyzed data from 37 speakers from GRASS and 9 speakers from the Kiel Corpus, totaling 1497 word-initial plosive tokens after outlier removal.

Key Insights and Ideas

  • Variety-Specific Strategies: The most significant finding is that German and Austrian German employ different acoustic strategies to mark prosodic boundaries and word stress, even though both varieties show effects of these prosodic factors.
  • Closure vs. Burst Duration: While both varieties lengthen plosives at prosodic boundaries and in stressed syllables, the specific cues used differ. German speakers seem to rely more on closure duration for marking stress and boundaries, whereas Austrian German speakers show more variation in burst duration.
  • Burst Duration as a "Free" Cue: Burst duration appears to be less directly controlled by speakers for prosodic marking compared to closure duration. It's suggested that burst duration might be more of an "epiphenomenon" of other articulatory movements, but its differential behavior in the two varieties suggests speakers can manipulate it for prosodic signaling.
  • Speech Rate Influence: Both global (utterance-wide) and local (adjacent segment) speech rates significantly affect plosive durations, but the interaction with local speech rate is also variety-specific.
  • Fortis-Lenis Distinction: The fortis-lenis distinction in plosives is primarily signaled by burst duration in this study's data, rather than closure duration, which contradicts some previous findings but aligns with others.

Key Results

  • Total Plosive Duration: Both varieties show longer plosives in stressed syllables. Plosive duration is also affected by prosodic boundaries, but this effect interacts with variety.
  • Closure Duration: Overall closures are longer for stressed tokens. A significant interaction between boundary and stress indicates that the effect of stress is weaker for plosives in phrase-medial positions compared to those at phrase edges. This effect is more pronounced in Austrian German.
  • Burst Duration: Burst duration is not significantly conditioned by prosodic boundary context in the combined data. However, it is longer in stressed than unstressed tokens. Crucially, the way stress affects burst duration differs: in German, bursts are significantly longer for stressed tokens, while in Austrian German, this effect is not significant.
  • Variety-Specific Stress Marking: German speakers increase burst duration for stressed plosives, while Austrian German speakers increase closure duration.
  • Speech Rate: An increase in global speech rate leads to a significant increase in plosive duration across all contexts and varieties.

Limitations and Future Work

  • Dataset Size: The Kiel Corpus (Northern German) dataset was significantly smaller (n=406) than the GRASS Corpus (Austrian German) dataset (n=1091), limiting direct statistical comparisons of effect sizes and significance.
  • Boundary Token Count: The number of plosive tokens at prosodic boundaries was relatively small (n=93), partly due to silent pauses often occurring at boundaries, making it impossible to measure closure duration.
  • Future Work Implications: The findings underscore the need for variety-specific models in speech technology. Tools trained on one German variety may not perform optimally on another due to these distinct prosodic marking strategies. Future work could involve developing adaptive models or further exploring the phonetic mechanisms underlying these differences.

Summary updated

BibTeX
@inproceedings{schuppler2019prosodic,
  author = {Barbara Schuppler and Margaret Zellers},
  title = {{Prosodic Effects on Plosive Duration in German and Austrian German}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure