Tracking the New Zealand English NEAR/SQUARE Merger Using Functional Principal Components Analysis

Michele Gubian, Jonathan Harrington, Mary Stevens, Florian Schiel, Paul WarrenInterspeech 2019

TL;DR

This paper introduces Functional Principal Components Analysis (FPCA) as a novel method to track subtle sound changes in speech, specifically the NEAR/SQUARE diphthong merger in New Zealand English. FPCA allows for a more holistic analysis of formant trajectories, revealing that the merger is characterized by increased centralization and monophthongization, and is more advanced in older speakers.

What Problem Does This Paper Address?

The paper tackles the challenge of accurately quantifying ongoing sound changes in spoken language, particularly those involving gradual shifts in pronunciation. Traditional acoustic analyses often focus on single time points within a sound, which can miss crucial dynamic information. The specific phenomenon under investigation is the merger of the New Zealand English (NZE) diphthongs /eə/ (SQUARE) and /ɪə/ (NEAR), where /eə/ is shifting towards /ɪə/. Previous studies have relied heavily on auditory analysis and perception experiments, with acoustic analyses being less common and often limited to single time points. This approach risks overlooking the full trajectory of formant changes and the influence of factors like tongue movement and diphthongal shape.

Methods

The core methodological innovation is the application of Functional Principal Components Analysis (FPCA) to analyze the dynamic shapes of formant trajectories. FPCA treats formant frequencies (F1 and F2) as continuous functions over time, allowing for a data-driven parametrization of these trajectories. Instead of analyzing discrete points, FPCA captures the overall shape and its variations.

The process involved:

  1. Acoustic Data Collection: Recording read speech from speakers across different age groups and regions.
  2. Forced Alignment: Using the Munich Automatic Segmentation System (MAUS) to align speech segments.
  3. Formant Tracking: Extracting F1 and F2 frequencies using the Praat formant tracker.
  4. Manual Correction: Correcting segment boundaries and formant frequencies.
  5. Speaker Normalization: Normalizing formant frequencies using z-scores relative to a speaker's own vowels.
  6. Time Normalization: Linearly normalizing all formant contours to 11 equally spaced time points.
  7. Smoothing: Applying a 5-point median filter to remove short-term fluctuations.
  8. Interpolation: Using B-splines to interpolate the formant contours, preparing them for FPCA.
  9. FPCA Application: Applying FPCA to the normalized and interpolated formant tracks to derive principal components (PCs) that capture the main modes of variation in the diphthong shapes.
  10. Statistical Modeling: Using Linear Mixed-Effects (LME) models to test the influence of factors like age, diphthong type, lexical frequency, and preceding consonant on the FPCA-derived scores.

Datasets

The study utilized the New Zealand Spoken English Database. This database contains:

  • Isolated /hVd/ words.
  • 200 read sentences drawn from a mixture of the TIMIT and ANDOSL corpora.

The data included recordings from speakers across three age groups (younger: 18-30, mid-age: 31-45, older: 46-60) and two regions (Hamilton and Wellington). A total of 2328 /eə/ and 1894 /ɪə/ diphthongs were analyzed.

Key Insights and Ideas

  • FPCA as a powerful tool for sound change: The study demonstrates FPCA's effectiveness in quantifying subtle, multi-dimensional changes in speech dynamics, moving beyond single-point analyses.
  • The merger is dynamic: The NEAR/SQUARE merger isn't just about a static shift in vowel quality but involves changes in the shape of the diphthongal trajectory, including increased centralization and monophthongization.
  • Age is a key factor: Older speakers show more advanced stages of the merger, supporting the idea of an ongoing sound change progressing through generations.
  • Diphthongal differentiation: The merger involves /eə/ becoming more monophthongal, while simultaneously becoming more differentiated internally, leading to a shift towards /ɪə/.
  • Lexical frequency's limited role: Contrary to some hypotheses, lexical frequency did not show a significant effect in driving this particular sound change, though a non-significant trend suggested younger speakers' low-frequency words might be more affected.
  • Coarticulation's limited role: The place of articulation of the preceding consonant did not significantly influence the merger, suggesting it's not primarily driven by coarticulatory effects.

Key Results

  • PC1: Captures overall centralization/peripheralization of the diphthong.
  • PC2: Relates to F2-raising in the initial part of the trajectory, suggesting fronting.
  • PC3: Crucially, captures the degree of monophthongization and internal differentiation of the diphthong. A negative s3 score corresponds to a shift from /eə/ towards /ə/.
  • Age × Diphthong Interaction: A significant interaction for PC3 indicated that acoustic differences between /eə/ and /ɪə/ were more pronounced in older speakers, supporting the hypothesis of an age-graded sound change.
  • Diphthong Effect: Both /eə/ and /ɪə/ showed significant differences in PC1 and PC3, with /eə/ being more centralized and less monophthongal than /ɪə/ when canonical.
  • Lexical Frequency: A non-significant interaction (p=0.07) between Age and Log Lexical Frequency suggested a slight tendency for younger speakers' low-frequency /eə/-words to shift towards /ɪə/, but no strong evidence for high-frequency words leading the change.
  • Place of Articulation: No significant interaction between Age and Place of Articulation was found, indicating that the merger is not more advanced in coronal contexts.

Limitations and Future Work

  • Lexical Frequency Trend: The interaction between Age and Lexical Frequency was not statistically significant (p=0.07), meaning the hypothesized role of high-frequency words leading the change was not strongly supported by the data.
  • Coarticulation Trend: Similarly, the interaction between Age and Place of Articulation was not significant, suggesting that the sound change is not primarily driven by coarticulatory factors in coronal contexts.
  • Future Work: The authors suggest that FPCA offers a valuable new approach for analyzing multidimensional dynamic phonetic contrasts, which could be applied to other sound changes, even those involving more complex vowel shifts like those in American English. The study highlights the potential of FPCA for understanding the intricate dynamics of speech evolution.

Summary updated

BibTeX
@inproceedings{gubian2019tracking,
  author = {Michele Gubian and Jonathan Harrington and Mary Stevens and Florian Schiel and Paul Warren},
  title = {{Tracking the New Zealand English NEAR/SQUARE Merger Using Functional Principal Components Analysis}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure