Exploring vernacular reorganisation with pre-schoolers’ short front vowels

Joshua Wilson Black, Lynn Clark, Margaret Blackwood, Robert Fromont

Overview

Background and Research Questions
Data
Data wrangling ‘highlights’
Models
Discussion

Kia ora koutou, thanks for having me.

As you’ve seen, the title of this talk is ‘Exploring vernacular reorganisation with preschoolers’ short front vowels’, and that’s exactly what we’re going to do.

I’ll start by 1. giving some background on vernacular reorganisation, the question of the onset of vernacular reorganisation, and set out a couple of research questions. I’ll then 2. tell you a bit about our data, which is at the same time, an incredibly valuable data set for exploring early dialect acquisition and responsible for a reduction in my life expectancy by at least a few years. That’s why, 3. I’ll say a bit about the data wrangling process in the hopes that it might be useful to some of you here. I’ll then run you through our 4. the models we used to explore the data and the results from them. Before turning to 5. a discussion of what they mean for our research questions.

The Team

Background

Vernacular reorganisation

[…] we necessarily begin with the phonetics, phonology, morphology, and syntax that we acquired from our first caretaker, normally female. The general condition for linguistic change can then be stated in a very simple way: children must learn to talk differently from their mothers. Let us refer to this process as vernacular re-organization. (Labov 2001, 415)

We investigate the onset of vernacular reorganisation.

Vernacular reorganisation

Transmission: Acquire caregiver’s vernacular.
Incrementation: Advance linguistic changes already underway in the community.

Trigger: shift from caregiver-dominated norms to peer-dominated norms around 4-5 and typically taken to be beginning of schooling.

This plot is frequently used to illustrate Labov’s model of vernacular reorganisation. This is the expected movement of an individual speaker in F2 for an unnamed vowel which is taken to be fronting in the community as the speaker ages.

The imagined speaker starts at their caregiver’s level. The acquisition of the caregiver’s vernacular is called transmission.

Then, at sometime between 4 and 5, but around at 4 in the plot here, the speaker departs from the caregivers model and begins to use more innovative variants as they age. This process is called incrementation.

At a later stage, we get stabilization, but this is not the topic of our work.

It is also worth noting that this model, as originally stated, assumes that the caregiver is female and that only female children engage in incrementation. Male children pick up their caregivers vernacular, while female children go ahead of it. This is supposed to explain the one-generation lag sometimes seen between male and female speakers with respect to sound changes.

–>

Research Questions

Is there evidence of vernacular reorganisation in the speech of preschool aged children?
Are all changing accent features advanced at the same time and at the same rate during incrementation?

The overall aim of our research is to understand how children’s vowel production changes over the period that spans their pre-school years.

In Labov’s model, the child’s transition to school at around 5 years of age and this is expected to mark the time when parental influence begins to wane and the peer group becomes important. But what if this happens sooner? What if there is already evidence of vernacular reorganisation happening within the preschool stage?

There are a few reasons we might expect to see evidence of vernacular reorganisation in children younger than school age. Labov’s model seems to hinge on a fairly outdated assumption that children undergo a kind of social awakening when they transition from home-based care to school. However, most children these days don’t transition from home-based care directly to school. Moreover, many studies on the acquisition of sociolinguistic variation have found adult-like patterns of linguistic and social variation in the speech of very young children (as young as two).

In NZ, most children spend at least 20 hours per week in an Early Childhood Education setting with other children (pre-school/nursery). This is because the government provides 20 hours of free ECE for children aged 3-5 (soon to lower to 2). Participation in ECE n NZ is higher than the OECD average (87% of NZ 3 year olds are in ECE 20 hours per week, compared with 71% across the OECD). So there is good reason to expect that peer groups are already important before the age of 5.

If we do find evidence of vernacular reorganisation going on in pre-school years, what does that look like? For instance: are all changing accent features advanced at the same time and at the same rate during incrementation?

Spoilers

We focus on the ‘extended’ short front vowel shift (SFVS): dress, trap, kit, nurse, and fleece. using a corpus of preschooler speech derived from a story retell task and a community corpus (QuakeBox).
A mixed bag of changes found, some consistent with change in community, some not.
Preschoolers’ production may be influenced by story telling style.
This can’t be pulled apart from vernacular reorganisation using our data, but should be taken account of in future work on vernacular reorganisation.
Looking at multiple variables is vital.

Since there’s going to be quite a lot of methodological detail here before we get to the results, I thought I’d give you a few spoilers.

some of our results are consistent with change in the direction of the community,
Kids seem to start from a much more ‘conservative’ place that would be expected on the standard model of vernacular reorganisation.
Child directed speech, or perhaps better, ‘storyteller’ speech, may explain these features of the preschooler vowel spaces.
More generally, the consequences of the shift from child-directed speech to adult speech on children’s language production may look like ‘vernacular reorganisation’ even though it is not actually a shift away from the caregivers vernacular. This is important for future research to take account of.
Looking at multiple variables together helps to avoid nice, simple, but likely false stories.

Data

Story Retell

Source: the UC Child Well-Being Research Institute’s Better Start Literacy Approach | Te Ara Reo Matatini.
Children were presented with a story and asked to retell it in their own words.
Two stories were used: Hana and the Tui and Tama and the Playground (Gillon, McNeill, and Scott 2019).
Both stories were written by the CWRI to match the NZ cultural context and contain a wide variety of literacy-relevant linguistic features (Gillon et al. 2023; Scott et al. 2022).

Our child speech data comes from a series of story retell tasks carried out by preschoolers in Christchurch New Zealand. The data were recorded by our collaborators in the Child Wellbeing Research Institute as part of a large government-funded study called the “Better Start Literacy Approach” project.

A large battery of language measures were collected from hundreds of pre-school children across Christchurch, and part of that suite of measures included a story re-tell task. Two stories were used: Hana and the Tui and Tama and the Playground. Both stories were constructed to contain a wide variety of vowel and consonant sounds and to match the New Zealand cultural context (Gillon et al. 2023; Scott et al. 2022).

Children were presented with a recorded voice and a series of still images (like this one).

After hearing the story and viewing the images, the children were shown the images again and asked to tell the story in their own words.

We were given access to anonymised transcripts and audio recordings from these preschoolers from multiple early childhood centres across Christchurch.

Tell:

Retell:

Preschooler corpus

18 centres
141 children
F: 76, M: 65
63 with two recordings
21 with three recordings
2 with four recordings

Preschooler corpus

100 NZ European
19 Māori
5 Pasifika
14 Asian
4 Other
Age at recording: 3;10 - 5;5
Median age: 4;6

Preschooler corpus

dress: 839 (F1), 834 (F2)
fleece: 730 (F1), 723 (F2)
kit: 1129 (F1), 1111 (F2)
nurse 388 (F1), 389 (F2)
trap 771 (F1), 760 (F2)

Wrangling

😢

Story retells were passed through Otter.ai for automatic transcription and then corrected by the CWRI team.
Problems:
- No reliable timestamps(!!!)
- SALT conventions not best for forced alignment
Recordings were often in noisy environments, e.g.:

Utterance correction

Manually corrected utterances in Transcriber

We manually corrected timestamps at the utterance level
Utterances were modified to avoid major sources of noise.
Transcripts and audio were then uploaded to LaBB-CAT (Fromont and Hay 2008).

Standard forced alignment settings failed.

Solution: Manual phonemic correction in Praat.

(Thanks Maggie!)

For this story, see: (Fromont 2023)

Standard formant tracking settings failed.

Formant tracking

Young children have very high frequency formants. This causes problems in addition to noise etc (see, e.g.: Valentine et al. 2023).
We use FastTrack (Barreda 2021)
The FastTrack process:
1. Formants are fit in Praat using a series of upper-limit values,
2. Smooths (really, quadratic models) are fit through each formant track,
3. Smoother analyses are favoured over wiggly ones,
4. Heuristics are applied.
Settings were fine tuned for each vowel type.

Before

After

Preprocessing

We filter:
- Tokens with insufficient data for FastTrack,
- Hesitations,
- Tokens without words,
- Formants with very high bandwidths,
- F1 > 1500hz,
- Tokens preceeding liquids,
- Tokens from ‘Hana’,
- Tokens ± 2 s.d. within 3 month age bins.
We track:
- Stopwords,
- Stresss

Modelling

General Approach

We don’t have all the data we need to directly test core hypotheses
- …so we go exploratory.
We want to discern an onset in sound change
- …so we use GAMMs.
Data is sparse and model convergence difficult
- …so we adopt a Bayesian approach

Bayesian approach increases the complexity of specifying models, but massively increases our flexibility in terms of model fit and examining the resulting models.
We use the brms package (Bürkner 2017).

Model structure

For each vowel’s F1 and F2 we fit the following model:

formant_value ~ gender + stopword + s(age_s, by=gender, k = 4) + (1|word/unstressed) + (1|participant/collect)

Response: skew-normal.

Similar models fit to QuakeBox.

Results

Here is our main finding – so we’ll spend a bit of time on this slide getting familiar with all that it is showing.

This figure shows model predictions of the 5 monophthongs, for three populations, plotted on a single vowel space for ease of interpretation:

the children’s vowels – the youngest children are represented as circles and the oldest children are represented as arrow heads; change over real time in the children is represented with a line connecting the circles and the arrow heads.
the youngest females in our QB data (represented by diamonds) who should provide us with a sample of speech that is relatively similar to the female caregivers in this community (reminder: we were unable to sample the actual caregivers of these children because of covid restrictions during data collection)
the oldest generation in the QB data. Comparing the Caregivers generation to the Oldest female generation gives us a picture of ongoing community change for females in this community using apparent time data. And comparing the change over time in the children with this gives us a picture of whether or not we’re seeing evidence of vernacular reorganisation at this age and stage.

I’ll take us through each of these vowels, but first we need to look at the sound changes present in the QuakeBox corpus.

e.g.:

Let’s go through each vowel. Note that vowel predictions are from 50 months to 60 months (4;2–5), to avoid plotting large movement at the edges of our data for each vowel.

The preschoolers seem to start with a much higher and fronter realisation of FLEECE, and this continues to raise over the next year and a half, in the opposite direction from the community change in progress.

The children seem to be starting out with relatively conservative realisations of dress, lower than either the caregiver population or the older speakers. We see them raise the vowel over their pre-school years and so they are more or less at line with caregivers’ realisations by 60 months.

Despite generations of lowering and backing, KIT (green) is now fairly stable in the adult population. However, the children’s realisations at 50 months are no-where near the community norm – the children have much higher and fronter realisations of KIT than the adults. This changes over time, so that by 67 months both the boys and girls show lower and more centralised realisations of KIT, although still some way away from the central realisations produced by the caregivers or older speakers.

There is no clear movement over time in the childrens’ speech for nurse (teal) or trap (pink). BUT They are both significantly lower and backer (NURSE) in the vowel space than the community reference points.

Now we look to see if there is any discernible onset time for these changes.

Remember that the overall aim of our research is to understand how children’s vowel production changes over the period that spans their pre-school years.

The original vernacular reorganisation model lays out the expectation that the child’s transition to school at around age 5 is expected to be the catalyst for the onset of vernacular reorganization, in which sound change away from the caregivers begins.

We were interested to know what is happening in the lead up to this time, and whether there is perhaps evidence of vernacular reorganisation from a much younger age, because the transition to school is no longer such a massive social change for children.

Here, I’ve picked out the four most obvious changes in the vowel space plot we’ve just been peering at: FLEECE F1, DRESS F1, KIT F1, and KIT F2. Females are in red and males in blue. Horizontal lines are the same ‘caregiver reference’ levels used in the previous plot. The rug below the plot indicates the presence of actual data points. We don’t want to over-interpret ‘wiggles’ in areas of low data density!

The same story is visible in all of these plots: there is no obvious place at which the changes ‘take off’ over the period we are looking.

DRESS F1 perhaps comes closest to canonical descriptions of vernacular reorganisation – there is some overlap between the male children and the caregiver distributions at age 3;5. By age 5 both male and female children are more or less in the same place as the caregivers but not yet advancing the change.

In both F1 and F2 space, KIT is fairly stable for the adult population. In both cases, the children start off somewhere far more conservative (i.e. the children have a higher fronter KIT realisation, quite unlike the centralised KIT that is typical of NZE). Over time they work towards this more canonical NZE realisation by lowering (females) and backing (males) their KIT realisations, but they still don’t quite reach the adult norms by the age of 5, especially not in the F2 dimension.

And FLEECE seems to be different again. FLEECE has been lowering in Christchurch for some years. But the children start off with a higher FLEECE vowel than the community and instead of converging on the community over time, they continue to raise it, moving in the opposite direction from the community and away from the caregiver norm.

Results summary

Is there evidence of vernacular reorganisation in the speech of pre-school aged children?
- Both stability (nurse, trap) and change (fleece, dress, kit).
- No ‘elbow’ in the smooths indicate rapid acceleration of change. fleece goes in the wrong direction.
- Difficult to interpret these straightforwardly as vernacular reorganisation.
Are all changing accent features advanced at the same time and at the same rate during incrementation?
- Not all features are changing.
- Some are changing faster than others.
- Some are changing differently for males and females.

Interpretation

Developmental?

Greater variation in kids than community.
Donegan (2012): this is unsurprising.
- But: children achieve ‘quite acceptable vowel quality’ before 3.
This is consistent with wider literature (e.g. Vorperian and Kent 2007).
No clear developmental reason for these shifts.

Donegan (2012:29) explains that “variability of vowels is not surprising when we consider the difficulty of controlling the shape of the tongue, which is a complex three-dimensional network of intrinsic longitudinal, vertical, and transverse fibers”. However, he goes on to say that there is enough evidence from a range of different languages to suggest that “many children achieve quite acceptable vowel quality before age three in all but the rhotic vowels” (2012:29).

Vorperian & Kent (2007) review data from 21 studies of F1 & F2 measurements of the four corner vowels in children from 8 months to 11 years. From age 3, the position of the vowels in the corners of the vowels space is shown to be relatively stable and remains this way across childhood (2007:1517).

Donegan (2012) cites work from the 1980s (de Boysson-Bardies, Halle, Sagart, & Durand, 1989) that shows infants in an English-speaking environment begin to produce recognisably different vowel qualities from as young as 10 months old, and as young as 7 months in a Mandarin language environment (Chen & Kent 2010).

Templin’s (1957) seminar study of 480 children aged 3-8 found that already the three year olds were producing vowels with 93.3% accuracy (with no significant increase in accuracy with age).

The implication, then, is that most vowel quality development happens before the age of 3 and so there seems to be no developmental reason to expect that children from 3 years onwards can’t accurately produce the vowels under investigation in this study.In other words, it’s not the case that they can’t produce a higher fronter DRESS vowel at age 4;2 and this is something that takes time to acquire. This means that the vowel changes that we see across the real-time span of our data must be real changes in progress for these children rather than an artefact of developmental delay.

Older caregivers?

The kids look conservative!
Are they recapitulating the SFVS?
But:
- What about fleece?
- Otherwise, they look archaic!
What about hyperarticulation?

CDS/Story-book speech

Triangles connect fleece, trap, and nurse.
Red indicates the preschoolers (4;2).
Blue indicates the recorded storyteller.
Black indicates QuakeBox ‘caregivers’.
Priming or style shifting?

Another possibility is that the audio that the children are hearing from the storybook also has features of hyperarticulation and the children are perhaps being primed by this or are actively copying/mimicking that style.

This is the image that we just saw with the children’s and the young female adults on the same vowel space.

Now let’s look at what happens when we overlay the vowels from the story (in blue). The storyteller also seems to have features of hyperarticulation – notice the much lower TRAP, and much higher FLEECE than is typical for adult speech in NZE. NURSE is also quite far back (even further back than it is for the kids).

There’s no possible way for us to know (with the data that we have) whether this is unconscious priming or whether the children are mimicking the storyteller’s style of speech. But it does seem to help explain why the vowels on the periphery of this space are so extreme.

Upshot

We’ve found sound change before kids start school.
No ‘elbow’, or point at which incrementation begins.
Some changes, e.g. dress raising look like vernacular reorganisation.
But fleece looks entirely different.
Possible influence of CDS/Story-book speech.
Focus on multiple variables at once is vital for this kind of project.

We’ve found some interesting and unexpected stuff. For instance, we found some change over time during a period of time where the model suggest there should be stable variation. And we found no evidence of an ‘elbow’ like this one i.e. no rapid acceleration of change for those vowel that were changing. We found some changes that look like the beginning of vernacular reorganisation (e.g. DRESS raising) and others that look entirely different (but can likely be explained by other phenomena such as priming or stylistic devices).

We think that our study highlights the importance of stepping away from a single-variable approach, which is a methodology that has dominated work in studies of language variation and change for many years. If we had selected any one of these vowels to investigate in isolation, we would have missed the complexity and the systematicity in these data.

Summary

Vernacular reorganisation
Our data
- Preschooler corpus + QuakeBox
Data wrangling challenges
- Transcription, forced alignment, formant tracking, normalisation.
Modelling strategy
- Bayesian GAMMs
Results and Interpretation

References

Barreda, Santiago. 2021. “Fast Track: Fast (Nearly) Automatic Formant-Tracking Using Praat.” Linguistics Vanguard 7 (1). https://doi.org/10.1515/lingvan-2020-0051.

Bürkner, Paul-Christian. 2017. “brms: An R Package for Bayesian Multilevel Models Using Stan.” Journal of Statistical Software 80 (1). https://doi.org/10.18637/jss.v080.i01.

Donegan, Patricia. 2012. “Normal Vowel Development.” In Handbook of Vowels and Vowel Disorders. Taylor & Francis.

Fromont, Robert. 2023. “Maximizing Accuracy of Forced Alignment for Spontaneous Child Speech.” Language Development Research, September. https://doi.org/10.34842/shrr-sv10.

Fromont, Robert, and Jennifer Hay. 2008. “ONZE Miner: The Development of a Browser-Based Research Tool.” Corpora 3 (2): 173–93. https://doi.org/10.3366/e1749503208000142.

Gillon, Gail, Brigid McNeill, and Amy Scott. 2019. Tama and the Playground. University of Canterbury Child Wellbeing Research Institute.

Gillon, Gail, Brigid McNeill, Amy Scott, Megan Gath, and Marleen Westerveld. 2023. “Retelling Stories: The Validity of an Online Oral Narrative Task.” Child Language Teaching and Therapy, 02656590231155861.

Labov, William. 2001. Principles of Linguistic Change: Social Factors. Wiley-Blackwell.

Scott, Amy, Gail Gillon, Brigid McNeill, and Alex Kopach. 2022. “The Evolution of an Innovative Online Task to Monitor Childrens Oral Narrative Development.” Frontiers in Psychology 13 (July). https://doi.org/10.3389/fpsyg.2022.903124.

Valentine, Hannah, Joel MacAuslan, Maria Grigos, and Marisha Speights. 2023. “Comparing Tools for Automation of Formant Estimation in Diverse Pediatric Populations.” The Journal of the Acoustical Society of America 153 (3_supplement): A212–12. https://doi.org/10.1121/10.0018687.

Vorperian, Houri K., and Ray D. Kent. 2007. “Vowel Acoustic Space Development in Children: A Synthesis of Acoustic and Anatomic Data.” Journal of Speech, Language, and Hearing Research. https://doi.org/10.1044/1092-4388(2007/104).

Extra: Formant tracking settings

Upper limits:

Front vowels get 6000-9000 Hz.
Back vowels get 5000-8000 Hz.
Other vowels get 5500-8500 Hz
‘Front vowels’ (NZE) = fleece, dress, nurse, goose, trap, kit
‘Back vowels’ = lot, thought, foot
‘Others’ = strut, start

By-vowel limits:

FLEECE: f2 > 1500,
DRESS: f2 > 1500,
GOOSE: f2 > 1000,
NURSE: f2 > 1200,
THOUGHT: f2 < 2250,
LOT: f2 < 2500
FOOT: f2 > 900
KIT: f2 > 1250

Formant bounds:

label f1lower f1upper f2lower f2upper f3lower f3upper START 350 1500 1200 3500 0 5000 THOUGHT 350 1500 1200 2250 0 5000 TRAP 350 1500 1200 3500 0 5000 NURSE 350 1500 1200 3500 0 5000 DRESS 350 1500 1500 4000 0 5000 FLEECE 350 1500 1500 4000 0 5000 KIT 350 1500 1250 3500 0 5000 LOT 350 1500 1200 2500 0 5000 GOOSE 350 1500 1000 3500 0 5000 FOOT 350 1500 900 3500 0 5000 STRUT 350 1500 1200 3500 0 5000

Extra: I want \(p\) values.

No. But perhaps you’ll like these:

	Model	Gender	Model	Longitudinal
DRESS	F1	F	97.6%	60.5%
DRESS	F1	M	80.1%	52.9%
FLEECE	F1	F	99.6%	57.1%
FLEECE	F1	M	78.7%	60.0%
KIT	F1	F	98.5%	55.6%
KIT	F1	M	83.6%	57.8%
KIT	F2	M	99.1%	60.0%