The roles of reconstruction and lexical storage in the comprehension of regular pronunciation variants

Mirjam ErnestusInterspeech 2009

TL;DR

This study investigates how we understand speech with common pronunciation variations, like shortened or dropped sounds. The research suggests that when we hear a new word with a regular reduction, we both store that specific reduced version and reconstruct its full, unreduced form, indicating a dual mechanism for processing speech.

What Problem Does This Paper Address?

Conversational speech is full of variations, especially "speech reduction," where sounds or syllables are weakened or dropped. This leads to many different pronunciations for the same word. For example, the Dutch word "natuurlijk" can be pronounced in numerous ways, some significantly altered from its full form. This raises a fundamental question in speech comprehension: how do listeners understand these variations? Two main hypotheses exist: 1) Listeners might reconstruct the full, unreduced pronunciation of a word using general rules or constraints, and then match this to a stored unreduced form in their mental lexicon. 2) Alternatively, listeners might store multiple pronunciation variants for each word directly in their mental lexicon and match the incoming sound to any of these stored forms. This paper aims to disentangle these two mechanisms by examining how listeners learn and process new words with regular pronunciation variants.

Methods

The researchers employed a word-learning paradigm using novel Dutch past-participles. The core of their methodology involved two studies, each with three phases:

  • Familiarization Phase: Participants learned new, non-existent past-participles. Crucially, half of these words were presented with an unreduced prefix ("ge-", pronounced [xə]), and the other half with a reduced prefix ("ge-", pronounced [x]). This reduction is common in Dutch and often results in phonotactically "illegal" consonant clusters, which can cue listeners to the presence of the schwa.
  • Control Lexical Decision Experiment: This phase, conducted a week after familiarization, aimed to assess if participants had integrated the new words into their mental lexicon. Participants performed a lexical decision task (deciding if a presented word is real or not) on existing low-frequency words and their phonological neighbors. If the new words were integrated, they should interfere with the recognition of their neighbors.
  • Main Lexical Decision Experiment: This phase directly tested the comprehension of the learned pronunciation variants. Participants performed a lexical decision task on the learned past-participles. The key difference between Study 1 and Study 2 lay in this phase: in Study 1, all learned words were presented in their reduced forms, while in Study 2, they were presented in their unreduced forms.

Response latencies (how quickly participants responded) and error rates were the primary measures. Statistical analysis involved multi-level regression modeling to account for participant and item variability.

Datasets

The study did not use pre-existing datasets for training or evaluation. Instead, the researchers created a novel set of 26 non-existing Dutch past-participles. These were designed to mimic regular Dutch morphology and phonotactics, with stems starting with obstruents. The words were also created to be phonological neighbors of low-frequency existing Dutch words. Stimuli were recorded by six female speakers of Standard Dutch, producing both unreduced and reduced versions.

Key Insights and Ideas

  • Listeners Store Word-Specific Variants: Study 1 demonstrated that when presented with new words in a reduced form, listeners store these specific reduced variants. This was evidenced by faster reaction times to words they had previously heard in their reduced form, even when the stimuli in the lexical decision task were also reduced (and longer than the familiarization stimuli). This suggests listeners are sensitive to word-specific pronunciation variation.
  • Reconstruction of Unreduced Forms is Likely: Study 2, where all learned words were presented in their unreduced form during the main lexical decision task, revealed no difference in reaction times based on whether the word was initially learned in its reduced or unreduced form. This implies that listeners likely reconstruct the unreduced form from the reduced variant and store both.
  • Unreduced Forms May Have a Privileged Status: The findings from Study 2 suggest that upon hearing a reduced form, listeners reconstruct the unreduced form and store both. This aligns with previous research suggesting that the unreduced pronunciation might hold a special status in speech comprehension.
  • Hybrid Models are Supported: The results collectively support hybrid models of speech processing, which propose that both the direct storage of pronunciation variants (exemplars) and the abstract reconstruction of unreduced forms play crucial roles in understanding speech.

Key Results

  • Study 1: Participants in the familiarization group were faster to respond to learned words in the main lexical decision task if they had heard those words in their reduced form during familiarization, compared to those who heard them unreduced. This effect was observed even though the reduced stimuli in the main task were longer than those in familiarization.
  • Study 2: Participants in Study 2 reacted equally quickly to learned words regardless of whether they had heard them in their reduced or unreduced form during familiarization, when all words were presented in their unreduced form in the main lexical decision task.
  • Control Lexical Decision Experiment: This experiment, designed to detect lexical integration, did not show significant interference with phonological neighbors, suggesting that the learned words were not fully integrated into the mental lexicon in a way that would hinder neighbor recognition, at least with the sensitivity of this particular task. However, the authors suggest this task might not have been sensitive enough.

Limitations and Future Work

  • Sensitivity of the Control Lexical Decision Task: The authors acknowledge that the control lexical decision experiment did not show the expected inhibition of phonological neighbors, which would have provided stronger evidence for lexical integration. They suggest this task might not have been sensitive enough, possibly due to the short duration of the learned words or the location of distinguishing segments. Further research is needed to understand these discrepancies with prior work.
  • Task Effects: The study notes that participants in the main lexical decision experiment were very careful with pseudo-words, potentially due to the familiarization phase focusing on pseudo-words. This raises questions about the influence of task effects on the results.
  • Generalizability: While the study focuses on regular reduction processes (schwa shortening), future work could explore how these mechanisms apply to other types of pronunciation variation.
  • Detailed Insights: The authors call for further research to gain more detailed insights into the precise relevance and interplay of storage and reconstruction mechanisms in general speech processing.

Summary updated

BibTeX
@inproceedings{ernestus2009roles,
  author = {Mirjam Ernestus},
  title = {{The roles of reconstruction and lexical storage in the comprehension of regular pronunciation variants}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure