Multidimensional acoustic variation in vowels across English dialects

James Tanner, Morgan Sonderegger, Jane Stuart-SmithNAACL 2022

TL;DR

This paper investigates how acoustic features of vowels, beyond their static position, vary across 21 English dialects. They found that dynamic features like formant trajectories and duration are crucial for understanding dialectal differences, revealing a structured variation that aligns with the traditional "monophthong-diphthong" distinction.

What Problem Does This Paper Address?

For decades, phoneticians have characterized vowels by their static acoustic properties (like pitch and resonance). However, vowels are dynamic sounds, changing over time. While this dynamic aspect is known to be informative, it's unclear how well static versus dynamic features capture the main dimensions of vowel variation, especially across a large number of dialects. This study aims to bridge this gap by exploring how static, dynamic, and duration information contribute to distinguishing vowels within and across English dialects, and whether these features align with established linguistic categories like "monophthong" and "diphthong."

Methods

The researchers employed a computational and exploratory approach. They extracted acoustic features from vowels in a large corpus of speech. Key measures included:

  • Static Formant Position: F1 and F2 values at the vowel onset and offset.
  • Dynamic Trajectory Shape: This was quantified using "vowel section length" (VSL) to measure spectral change between points in the vowel. Specific measures derived from VSL include:
  • Vector Length: Overall spectral change from onset to offset.
  • Trajectory Length: Sum of VSLs from onset to midpoint and midpoint to offset.
  • Vector Angle: Direction of spectral change.
  • Vowel Duration: The time length of the vowel, z-score normalized.

To assess the discriminative power of these features, they used Support Vector Machines (SVMs) for dialect classification. Different combinations of features (formants only, trajectory only, duration only, and various combinations) were used to train SVMs to predict the dialect of a given vowel token.

Datasets

The study utilized data from the Speech Across Dialects of English (SPADE) project. This project collated multi-dialect corpora from the United Kingdom and North America. Specifically, they analyzed 323,060 tokens from 1245 speakers across 21 dialects of British and North American English. The paper provides a detailed table (Table 1) listing the dialects, corpora, speaker counts, and token counts.

Key Insights and Ideas

  • Structured Dialectal Variation: Despite individual speaker variability, the acoustic realization of the "same" vowel across different dialects shows a consistent, structured pattern.
  • Monophthong-Diphthong Spectrum: The dynamic variation in vowels, particularly formant trajectories and duration, broadly aligns with the traditional linguistic distinction between monophthongs (less change) and diphthongs (more change). Vowels like FLEECE showed the least variation, while CHOICE, FACE, and PRICE showed progressively more.
  • Dynamic Features are Crucial: Static formant position alone is insufficient for fully characterizing dialectal variation. Dynamic features, especially trajectory shape and duration, add significant resolution.
  • Duration as a Complementary Cue: While duration alone is a weak classifier, it significantly improves dialect classification when combined with static formant information, suggesting it acts as a meaningful cue alongside formants rather than an independent one.
  • Trajectory Shape's Variable Role: The contribution of trajectory shape to dialect classification varies significantly across vowels, being particularly important for distinguishing MOUTH vowels, which exhibit notable dialectal shifts in their dynamic realization.

Key Results

  • Improved Classification with Dynamics: SVM models trained with combinations of features consistently outperformed baseline models and models using only static formants. The best-performing models always included dynamic information (trajectory or duration).
  • Duration Boosts Accuracy: Combining duration with formants significantly increased classification accuracy across all vowels (Table 3).
  • Trajectory's Impact Varies: Trajectory information provided a notable boost when combined with formants, especially for MOUTH vowels.
  • Overall Best Performance: The model using all three types of information (formants, trajectory, and duration) generally achieved the highest balanced accuracy across vowels, with MOUTH and PRICE showing the highest accuracies (around 70%).

Limitations and Future Work

  • Limited Vowel Set: The study focused on five specific vowels (FLEECE, FACE, PRICE, MOUTH, CHOICE), which tend to be "tense" and longer. Future work should incorporate a wider range of vowels, including "lax" vowels, to get a more comprehensive understanding of duration's role across the entire vowel system.
  • Automatic Measurement Limitations: While automated tools enabled large-scale analysis, manual validation of all measurements was not feasible. Potential biases or inaccuracies in automatic formant tracking and forced alignment segment durations (minimum 30ms) might exist.
  • Trajectory Representation: The study used a specific vector-based representation for trajectory shape. Testing other representations, such as Discrete Cosine Transforms (DCTs), could potentially improve dialect classification accuracy.
  • Further Exploration of Duration: The role of duration in dialectal variation needs further investigation, especially for vowels beyond the "tense" set.
  • Investigating Sources of Variation: Future research should aim to map out the structure of variability more fully and explore its potential sources.

Summary updated

BibTeX
@inproceedings{tanner2022multidimensional,
  author = {James Tanner and Morgan Sonderegger and Jane Stuart-Smith},
  title = {{Multidimensional acoustic variation in vowels across English dialects}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure