Introduction and Background

Traditional view: Regional variation is limited. "Australia is, generally speaking, linguistically unified" (Mitchell & Delbridge 1965: 13)

AusE has “begun to exhibit more widespread social and regional variation than has previously been acknowledged” (Cox and Fletcher, 2017, p. 20)

This study: Investigation of regional variation of prelateral merger of /e/ and [æ] in a large dataset

• Prelateral merger of /e/ and [æ] (e.g., celery = salary, Ellen = Alan)

• Investigated in small-scale studies in Southern Victoria/Melbourne (Diskin et al. 2017; Diskin-Holdaway et al., 2024; Loakes et al., 2017, 2024)

• What about in large datasets and in other locations?

Methods: Vowel extraction

• Data from CoANZSE Audio (https://coanzse.org) (Coats 2024a,b)
• ASR transcripts and audio from 38,786 videos uploaded to YouTube channels of Australian councils in 404 locations
• Alignment with the Montreal Forced Aligner (McAuliffe et al. 2017)
• F1 and F2 formant values for /e/ and /æ/ extracted at vowel midpoints using Parselmouth-Praat (Jadoul et al., 2018), a Python port of Praat (Boersma & Weenink 2024)
• Vowels were extracted in two contexts: prelateral (/æl/ and /el/, e.g. value, well) and non-prelateral (/æC/ and /eC/, e.g. fact, next)
• Values from CMU pronunciation dictionary (Weide et al. 1998)
• After filtering (only stressed syllables, only locations with at least 20 tokens in all contexts): 4,297,259 vowel tokens from 252 locations
• F1 and F2 formants, values normalized with Nearey's method

Methods: Bhattacharyya Difference, Spatial Autocorrelation

• Many recent studies use Pillai's trace to quantify vowel overlap.
• Pillai's trace is is from MANOVA, used for quantifying additional covariates and generating p-values
• We use the Bhattacharyya Distance (Bhattacharyya 1943) $D_B$, defined as:

$$D_B = -\log{\int \sqrt{P(x) \cdot Q(x)} \, dx}$$

for two multivariate probability distributions $P$ and $Q$ (i.e., $f_1, f_2$ for /e/ and /æ/)

• For each location, we calculate the $D_B$ between /e/ and /æ/, based on all measurements at that location, for
  • 1) Non-prelateral contexts ($vC$), and
  • 2) Prelateral contexts ($vL$)

We then calculate the difference between the two contexts at each location:

$$\text{Diff}_B = D_{\text{vC}} - D_{\text{vL}}$$

Methods: Spatial autocorrelation

• Moran's I and Getis-Ord G^*_i from $\text{Diff}_B$
  • Spatial weights matrix $W$ based on distance band $w_{ij}=1/d_{ij}$
  • Minimum distance: all locations must have at least one neighbor

Moran's I (Moran 1950): Takes into account attribute values at all locations in a dataset and summarizes the overall extent of spatial correlation

• 1: perfect clustering, 0: random distribution, -1: pefect dispersion

Getis-Ord G^*_i (Getis & Ord, 1992; Ord & Getis, 1995): Identifies spatial clusters by evaluating the values at each location in the dataset in comparison with neighboring locations, in relation to the global dataset

• Positive: location is ~> global mean; 0: location ~= global mean; negative: location ~< global mean

Results: Regional distribution

Results: Regional distribution, G^*_i values

Caveats, summary, outlook

• ASR and formant extraction methods can introduce errors
• No demographic data, but some can be semi-automatically annotated (cf. Bredin 2023; Plaquet & Bredin 2023; Ferreira 2024)
• Pipeline approaches can be used to automatically process large data volumes and extract formants for millions of vowels
• Merger confirmed as Southern Victoria/Melbourne phenomenon
• Need to look more closely at Western Australia

References