The processes underlying two frequent casual speech phenomena in Dutch: a production experiment

Iris Hanique, Mirjam ErnestusInterspeech 2012

TL;DR

This study investigates how Dutch speakers reduce common speech sounds like schwa and /t/ in fast, casual speech. Using a shadowing task, they found that these reductions are primarily driven by gradient co-articulation, suggesting that the absence of sounds is an extreme form of shortening rather than a distinct categorical process, contrary to some previous corpus-based findings.

What Problem Does This Paper Address?

Casual speech is characterized by significant reductions in pronunciation, making words sound different from their formal, dictionary pronunciations. Understanding the mechanisms behind these reductions is crucial for fields like speech recognition, synthesis, and linguistics. Specifically, this paper addresses how two common reduction phenomena in Dutch – the shortening and absence of schwa (a neutral vowel sound) and the consonant /t/ in past participles – are produced. Previous research using corpus studies suggested that some reductions might be categorical (all-or-nothing processes), while others might be gradient (gradual shortening). This paper aims to clarify whether these reductions are primarily gradient or categorical by using a controlled production experiment, specifically a shadowing task, and comparing its results to those from corpus analyses.

Methods

The researchers employed a controlled production experiment using a shadowing task. In this task, participants listened to pre-recorded Dutch sentences and were instructed to repeat them as quickly and accurately as possible. This method aims to elicit spontaneous, fast speech.

The core of the methodology involved:

  • Stimuli Creation: 180 target Dutch past participles (e.g., "gedanst" - danced) were created, often starting with "ge-" and ending in a consonant, with a schwa in the initial syllable followed by a consonant. These were embedded in sentences produced in both "casual" and "careful" speech styles.
  • Data Collection: 35 native Dutch speakers participated. Their repeated utterances were recorded.
  • Automatic Phonetic Transcription: Forced alignment software was used to generate broad phonetic transcriptions of the recorded speech. This system selected the most likely pronunciation variant from a lexicon.
  • Data Processing and Exclusion: A rigorous process was used to exclude trials that were not produced fluently, contained silences, or had potentially incorrect transcriptions. This resulted in the exclusion of a significant portion (35%) of the trials, indicating the difficulty of the task.
  • Statistical Modeling: Mixed-effects regression models were used to analyze the data. These models account for individual differences in participants and stimuli. Predictors included experimental factors (register, block), acoustic features (durations of surrounding segments), linguistic factors (word frequency, predictability), and phonetic factors (manner, voicing of surrounding segments).

Datasets

The study utilized the Spoken Dutch Corpus [6] to determine the frequency of occurrence for the target past participles. This corpus data was used to inform the selection of stimuli and the calculation of predictability measures. No new datasets were introduced.

Key Insights and Ideas

  • Shadowing Elicits Non-Casual Fast Speech: The study found that the shadowing task, while eliciting fast speech, did not perfectly replicate the reduction patterns observed in spontaneous conversational corpora. This suggests that the shadowing task elicits a specific type of fast speech that might differ from truly spontaneous, informal speech.
  • Absence as Extreme Shortening: The results strongly suggest that the absence of schwa and /t/ is not a distinct categorical process but rather the extreme end of gradient shortening due to co-articulation. The same variables that influenced shortening also influenced absence.
  • Co-articulation is Key: Reductions in schwa and /t/ are heavily influenced by the acoustic properties of surrounding sounds (e.g., their duration, manner, and voicing). This indicates that articulatory gestures of adjacent sounds overlap and influence each other.
  • Frequency Matters (but differently): Higher word frequency was associated with more reduction, consistent with the idea that frequent words are more automatized and produced with more overlapping gestures.
  • Limitations of Shadowing for Corpus-like Reductions: The participants did not reproduce the highly reduced forms sometimes found in spontaneous speech corpora, suggesting that the shadowing task might not fully capture the most extreme forms of reduction found in naturalistic, unconstrained speech.

Key Results

  • Schwa:
    • Schwa was more often absent and shorter in the "casual" condition (29.8% absent) compared to the "careful" condition (22.9% absent).
  • Schwa was shorter when surrounded by longer consonants.
  • The manner of articulation of the following consonant influenced schwa duration (longer with plosives, shorter with fricatives).
  • /t/:
    • /t/ was shorter at higher speech rates.
  • /t/ was more likely to be absent and shorter when followed by a consonant (28.3% absent) than a vowel (10.5% absent).
  • The duration of the following segment influenced /t/ reduction, particularly if it was a consonant.
  • Word-final /t/ was more likely to be absent and shorter in more predictable word combinations.

Limitations and Future Work

  • Limitations:
    • The shadowing task did not elicit the most extreme reductions observed in spontaneous speech corpora. Participants seemed to aim for producing the full form of the word rather than perfectly mimicking the reduced variant they heard.
  • A significant portion of trials (35%) had to be excluded due to task difficulty, highlighting challenges in controlling for fluency and transcription accuracy in production experiments.
  • Future Work:
    • The authors recommend using shadowing tasks in conjunction with other research methods (like corpus analysis) to gain a more comprehensive understanding of natural speech.
  • Further investigation into why participants did not fully replicate the most reduced forms heard is warranted.
  • Exploring other experimental tasks that might elicit more naturalistic casual speech reductions could be beneficial.

Summary updated

BibTeX
@inproceedings{hanique2012processes,
  author = {Iris Hanique and Mirjam Ernestus},
  title = {{The processes underlying two frequent casual speech phenomena in Dutch: a production experiment}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure