Describing the development of intonational categories using a target-oriented parametric approach

Britta Lintfert, Bernd MöbiusInterspeech 2012

TL;DR

This paper introduces a novel parametric approach to analyze the development of intonation in children, moving beyond simply imposing adult linguistic categories. By using a data-driven clustering method on acoustic features, they demonstrate that children's intonation contours gradually evolve towards adult targets, providing a more nuanced understanding of language acquisition.

What Problem Does This Paper Address?

Understanding how children acquire intonation is crucial for developmental linguistics and speech technology. Traditionally, researchers have tried to map children's speech to established adult intonation categories, like those in the ToBI (Tones and Break Indices) framework. However, this approach assumes children can already produce these complex adult categories, which might not be true, especially in early stages of language development. This paper addresses the challenge of describing intonation development without pre-imposing adult categories, aiming to capture the continuous, gradual nature of this process. They want to see if a data-driven approach can reveal how children's intonation patterns evolve towards adult norms.

Methods

The core of this paper is a two-step methodology:

  1. Parametric Modeling (PaIntE): Instead of directly analyzing raw pitch contours, the researchers use the PaIntE (Parametrized Intonation Events) method. PaIntE models stretches of pitch (FO) contours using a combination of rising and falling sigmoid functions. This captures key phonetic features like peak height, temporal alignment, and the amplitude/steepness of pitch movements. This parametric representation is more robust and can capture fine phonetic details relevant to intonation.
  2. Clustering and Evaluation: The parametrized features from PaIntE are then fed into a K-means clustering algorithm. This algorithm groups similar intonation contours together, forming data-driven "categories." The number of clusters is varied (from 2 to 9) to explore different levels of granularity. To evaluate these clusters, they are compared against manually annotated German ToBI (GTOBI(S)) categories. The "classification accuracy" is calculated by assigning each cluster the most frequent GTOBI(S) category within it and then seeing how often the manually labeled category matches the cluster's assigned category. This allows them to assess how well the data-driven clusters correspond to established linguistic categories.

Datasets

The study utilizes longitudinal data from 9 German-speaking children aged between 1 and 8 years. This data is part of the Stuttgart Child Language Corpus. In addition to the children's speech, the recordings also include:

  • Child-Directed Speech (CDS): Speech produced by the mothers of the children, intended for the children.
  • Adult-Directed Speech (ADS): Speech produced by the mothers, directed towards adults (presumably the researchers).

The paper notes that the mothers' speech (both CDS and ADS) was also analyzed to establish adult target categories for comparison.

Key Insights and Ideas

  • Data-Driven Categories: The research demonstrates that a parametric approach combined with clustering can reveal meaningful intonational patterns in children's speech without relying on pre-defined adult categories.
  • Gradual Development: Intonation development isn't an all-or-nothing process. Children's contours gradually evolve towards adult targets, rather than suddenly adopting them.
  • Mothers' Speech Consistency: The study found that mothers' intonation categories in Child-Directed Speech (CDS) are similar to those in Adult-Directed Speech (ADS), suggesting that mothers don't drastically alter their prosodic patterns when speaking to children.
  • Age-Dependent Correlation: The correlation between data-driven clusters and adult GTOBI(S) categories significantly improves with age, particularly from around 5 years old. This suggests that younger children's intonation might not be adequately described by adult categories.
  • PaIntE's Advantage: The PaIntE parametrization is highlighted for its ability to capture fine phonetic details like peak alignment and rise/fall amplitudes, which are crucial for understanding intonation contours.

Key Results

  • Adult Speech Comparison: Classification accuracies for both CDS and ADS data (when mapped to clusters) were significantly better than chance baselines (around 65-70% vs. 44-49%). This indicates that the clustering method successfully identifies patterns corresponding to GTOBI(S) categories in adult speech.
  • Child Speech Development:
    • For children aged 1-2 years, classification accuracy was only slightly better than the baseline, suggesting their intonation categories are not yet well-aligned with adult GTOBI(S) categories.
  • From around 5 years of age onwards, classification accuracies for children's speech showed a marked improvement and consistently exceeded the baseline, indicating a strong correlation with adult GTOBI(S) categories.
  • When mapping child data to adult target categories (Experiment 3), accuracies were initially worse than baseline for children under 4 years. However, from age 4 onwards, accuracies improved, and by age 5, children's categories were comparable to adult targets.

Limitations and Future Work

  • Limitations: The paper acknowledges that for younger children (under 5 years), GTOBI(S) categories may not adequately describe their produced intonation. The current method might not fully capture the nuances of early developmental stages.
  • Future Work: The authors suggest extending the method by including additional temporal parameters. They also envision applying this approach to longitudinal studies of intonation acquisition, as the same method can be used across different developmental stages. The automatic nature of the approach is also highlighted as a benefit for future research, allowing for consistent analysis across various theoretical frameworks and potentially incorporating discourse structure.

Summary updated

BibTeX
@inproceedings{lintfert2012describing,
  author = {Britta Lintfert and Bernd Möbius},
  title = {{Describing the development of intonational categories using a target-oriented parametric approach}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure