Creating Vowel Charts in R

Author

Martin Schweinberger

Published

January 1, 2026

Introduction

This tutorial demonstrates how to create personalised vowel charts using Praat and R. Vowel charts are among the most informative visualisations in phonetics and language learning: they map the acoustic properties of vowel sounds onto a two-dimensional space that closely mirrors the articulatory position of the tongue during vowel production.

The tutorial is aimed at beginners and intermediate users of R with an interest in phonetics, second language acquisition, or corpus phonology. It walks through the full workflow from recording to visualisation: recording words in Praat, extracting formant values, loading and processing the data in R, and producing publication-quality vowel charts with ggplot2. We also introduce vowel normalisation — an essential step when comparing speakers of different ages, sexes, or vocal tract sizes.

The aim is not to provide a fully-fledged acoustic analysis but to give a practical, step-by-step guide to vowel chart production that can be adapted for teaching and research purposes.

Prerequisite Knowledge

Before working through this tutorial, you should be familiar with:

Citation

Martin Schweinberger. 2026. Creating Vowel Charts in R. The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia. url: https://ladal.edu.au/tutorials/vowelchart/vowelchart.html (Version 2026.03.27), doi: 10.5281/zenodo.19242479.

Setup

Installing Packages

Code 
# Run once — comment out after installation
install.packages("tidyverse")
install.packages("flextable")
install.packages("ggforce")
install.packages("scales")

Loading Packages

Code 
library(tidyverse)
library(flextable)
library(ggforce)
library(scales)
klippy::klippy()

Vowel Sounds and Acoustic Phonetics

Section Overview

What you will learn: How vowels differ from consonants; what formants are and why F1 and F2 map onto the vowel space; and how a vowel chart relates to articulatory and acoustic phonetics

Vowels and How They Differ from Consonants

When learning or studying a language, one of the first distinctions encountered is that between consonants and vowels (Rogers 2014). Consonants are produced with an obstruction of the airstream; they cannot form the nucleus of a syllable on their own. Vowels, by contrast, are produced with an open, unobstructed vocal tract and form the carrying core of syllables (Zsiga 2012).

Because vowels are produced without obstruction, they are classified by different criteria than consonants:

  • Number of tongue positions during production: monophthongs (one stable position), diphthongs (glide between two positions), triphthongs (three positions)
  • Tongue height: high (close), mid, low (open)
  • Tongue backness: front, central, back
  • Lip rounding: rounded or unrounded

The classical IPA vowel quadrilateral maps these articulatory dimensions onto a two-dimensional chart — the standard representation of the vowel space. Below is the chart for monophthongal vowels of Received Pronunciation (RP).

(a) Vowel Chart (RP)

Tongue positions for vowels
Figure 1: Vowel Chart of monophthongal vowel sounds in Received Pronunciation (RP) (left); Tongue positions for the vowels /i, e, ɛ, a/ (right)

From Articulatory to Acoustic: What Are Formants?

Vowels are periodic sounds — rhythmic compressions and decompressions of air produced by the vibration of the vocal folds. Each vowel is a complex periodic wave that can be decomposed into multiple simple periodic waves (formants) through acoustic analysis (Johnson 2003; Ladefoged 1996).

Formants are resonant frequency bands produced by the shape of the vocal tract. Different vowels are associated with different vocal tract shapes and therefore different formant patterns. The first two formants are linguistically the most important:

Formant

Articulatory correlate

Typical range (adult)

F1 (first formant)

Tongue height — low vowels have high F1; high vowels have low F1

250–900 Hz

F2 (second formant)

Tongue backness + lip rounding — front vowels have high F2; back/rounded vowels have low F2

700–2,500 Hz

F3 (third formant)

Lip rounding and individual voice quality — less used in vowel classification

1,700–3,500 Hz

The key insight that makes acoustic vowel charts possible is that plotting F1 (y-axis, reversed) against F2 (x-axis, reversed) produces a display that closely mirrors the traditional articulatory vowel quadrilateral — high front vowels appear in the top-left, low back vowels in the bottom-right (Peterson and Barney 1952). This means that acoustic measurements from audio recordings can be used to produce a personalised vowel chart without needing to image the speaker’s tongue directly.

Why Are the Axes Reversed?

In a traditional vowel quadrilateral, high vowels are at the top and front vowels are on the left. Because F1 increases as the tongue lowers (low vowels = high F1), the y-axis is reversed so that low F1 (= high vowel) appears at the top. Similarly, F2 increases as the tongue moves forward (front vowels = high F2), so the x-axis is reversed so that high F2 (= front vowel) appears on the left. Reversing both axes aligns the acoustic plot with the traditional articulatory chart.

Recording and Measuring Formants in Praat

Section Overview

What you will learn: How to record words in Praat; how to set formant parameters; how to use TextGrids for reproducible annotation; how to extract F1 and F2 values; and an introduction to automated extraction via Praat scripting

Word Selection

To produce a vowel chart, we need recordings of words containing each monophthongal vowel sound. The standard approach uses minimal pairs in a controlled phonetic environment — typically the /h_d/ frame, which places vowels between two consonants with minimal coarticulation effects.

Word

IPA

Phonemic context

Vowel class

had

æ

h _ d

low front

hard

ɑ

h _ d

low back

head

e

h _ d

mid front

heed

i

h _ d

high front

herd

ɜ

h _ d

mid central

hid

ɪ

h _ d

high front lax

hoard

ɔ

h _ d

mid back rounded

hod

ɒ

h _ d

low back rounded

hood

ʊ

h _ d

high back lax

hud

ʌ

h _ d

mid central low

who'd

u

h _ d

high back

Each word should be repeated at least three times to allow for averaging across tokens, which reduces measurement error. Aim for a natural speaking rate — neither too fast nor too slow.

Recording in Praat

Download Praat from www.praat.org and install it. After installation:

  1. Open Praat — two windows appear: the Objects window (left) and the Picture window (right). Close the Picture window.
  2. Go to New → Record mono sound in the Objects window.
  3. In the SoundRecorder window, click Record and read the word list aloud. Repeat each word three times with a short pause between repetitions.
  4. When finished, click Stop, enter a name (e.g. vowels) in the Name field, and click Save to list & close.
  5. Back in the Objects window, select the sound object and click Save → Save as WAV file to save a backup.
Recording Quality

Microphone quality substantially affects formant accuracy. For research purposes, use a headset or close-talking microphone in a quiet room. For pedagogical purposes — e.g. giving a learner feedback on their vowel production — a laptop microphone in a reasonably quiet room is sufficient. Avoid recording near fans, air conditioning, or open windows.

Figure 2

Formant Settings

Before extracting formants, the formant parameters must be set appropriately for the speaker:

  1. Select the sound object and click View & Edit
  2. In the edit window, go to Formant → Formant settings
  3. Enable Show formants
  4. Set Maximum formant (Hz) according to the speaker:
    • Adult male: 5,000 Hz
    • Adult female: 5,500 Hz
    • Child: up to 8,000 Hz
  5. Set Number of formants (default 5 is usually appropriate; try 4–6 if formant tracks appear unstable)
Diagnosing Poor Formant Tracking

If the formant dots in the edit window are scattered or erratic rather than forming smooth horizontal lines, try adjusting the Maximum formant setting up or down by 500 Hz, or change the Number of formants between 4 and 6. Poor tracking is most common in low back vowels where F1 and F2 are close together.

Measuring Formants Manually

For each vowel token:

  1. Listen to the recording and identify the steady state of the vowel — the portion where the formant lines are stable and horizontal, corresponding to the target vowel quality
  2. Click and drag to highlight the steady state
  3. Go to Formant → Get first formant — note the Hz value
  4. Go to Formant → Get second formant — note the Hz value
  5. Also note the start and end time of the selection from the time axis

Record all values in a table. A completed table for one speaker is shown in the Data section below.

Using TextGrids for Reproducible Annotation

For research purposes, manual note-taking is not recommended — it is slow, error-prone, and not reproducible. TextGrids provide a structured, reusable annotation layer attached to the audio file.

To create a TextGrid:

  1. Select the sound object in the Objects window
  2. Go to Annotate → To TextGrid and create an interval tier named vowel and a point tier named label
  3. In the edit window, mark the steady-state boundaries of each vowel token as interval boundaries on the vowel tier and label each interval with the IPA symbol or word

Once annotated, formants can be extracted for all labelled intervals using a Praat script, which eliminates the need to measure each token manually.

Automated Extraction with a Praat Script

The following Praat script extracts the mean F1 and F2 from the midpoint of each labelled TextGrid interval and writes the results to a tab-delimited text file. Save this as extract_formants.praat and run it from Praat → Open Praat script.

# extract_formants.praat
# Extracts mean F1 and F2 from TextGrid interval midpoints
# Output: tab-delimited text file

form Extract formant values
  sentence Sound_file vowels.wav
  sentence TextGrid_file vowels.TextGrid
  real Maximum_formant_(Hz) 5500
  integer Number_of_formants 5
  sentence Output_file formants_output.txt
endform

Read from file: sound_file$
sound = selected("Sound")
Read from file: textGrid_file$
tg = selected("TextGrid")

# Create formant object
selectObject: sound
To Formant (burg): 0, number_of_formants, maximum_formant, 0.025, 50
formant = selected("Formant")

# Open output file
writeFileLine: output_file$, "label", tab$, "time", tab$, "F1", tab$, "F2"

# Loop over all intervals in tier 1
selectObject: tg
n = Get number of intervals: 1

for i from 1 to n
  label$ = Get label of interval: 1, i
  if label$ <> ""
    t_start = Get start time of interval: 1, i
    t_end   = Get end time of interval:   1, i
    t_mid   = (t_start + t_end) / 2

    selectObject: formant
    f1 = Get value at time: 1, t_mid, "Hertz", "Linear"
    f2 = Get value at time: 2, t_mid, "Hertz", "Linear"

    appendFileLine: output_file$,
      ... label$, tab$, fixed$(t_mid, 3), tab$,
      ... fixed$(f1, 2), tab$, fixed$(f2, 2)

    selectObject: tg
  endif
endfor

writeInfoLine: "Done! Output written to: ", output_file$
Why Use Praat Scripting?

A Praat script makes the measurement process reproducible (anyone can re-run it on the same audio and get the same values), scalable (it works equally well for 10 or 10,000 tokens), and consistent (all measurements are taken at the same relative position within each interval). For any analysis beyond a personal demonstration, scripted extraction is strongly preferred over manual measurement.

Data and Processing in R

Section Overview

What you will learn: How to load and inspect formant data; how to combine a learner’s data with a native-speaker reference; how to compute mean formant values by vowel and speaker; and how to prepare the data for plotting

The Data

For this showcase we use two datasets:

  • nns — formant measurements from one L1-German learner of English, recorded using the /h_d/ word list and measured manually in Praat
  • ns — reference formant values for Received Pronunciation from Hawkins and Midgley (2005), representing five L1-RP speakers aged 20–25
Code 
ns  <- read.table("tutorials/vowelchart/data/rpvowels.txt", header = TRUE, sep = "\t")
nns <- read.table("tutorials/vowelchart/data/vowels.txt",   header = TRUE, sep = "\t") |>
  dplyr::select(-file)

The learner data contains three repetitions of each word. Let us inspect the first few rows:

subject

item

trial

F1

F2

ms

had

1

717.3

1,868

ms

had

2

743.5

1,904

ms

had

3

721.0

1,939

ms

hard

1

734.5

1,493

ms

hard

2

832.9

1,408

ms

hard

3

797.3

1,498

ms

head

1

610.9

2,063

ms

head

2

722.3

2,131

ms

head

3

625.1

2,010

ms

heed

1

263.4

2,833

ms

heed

2

301.4

2,746

ms

heed

3

287.0

2,823

The native-speaker reference data contains one row per vowel — the mean F1 and F2 across all speakers and tokens, plus standard deviations:

subject

item

trial

F1

F2

S4-1

heed

1

261

2,282

S4-1

heed

2

288

2,304

S4-1

heed

3

307

2,373

S4-1

heed

4

328

2,416

S4-2

heed

1

290

2,215

S4-2

heed

2

257

2,263

S4-2

heed

3

290

2,255

S4-2

heed

4

251

2,223

S4-3

heed

1

235

2,552

S4-3

heed

2

270

2,479

S4-3

heed

3

238

2,434

S4-3

heed

4

295

2,459

S4-4

heed

1

295

1,962

S4-4

heed

2

291

1,996

S4-4

heed

3

309

1,983

S4-4

heed

4

269

2,045

S4-5

heed

1

244

2,632

S4-5

heed

2

268

2,640

S4-5

heed

3

284

2,627

S4-5

heed

4

253

2,612

S4-1

hid

1

454

2,375

S4-1

hid

2

452

2,032

S4-1

hid

3

424

2,003

S4-1

hid

4

441

2,098

S4-2

hid

1

368

2,105

S4-2

hid

2

345

2,089

S4-2

hid

3

338

2,161

S4-2

hid

4

390

2,079

S4-3

hid

1

378

2,438

S4-3

hid

2

411

2,369

S4-3

hid

3

421

2,360

S4-3

hid

4

366

2,442

S4-4

hid

1

346

2,000

S4-4

hid

2

342

1,974

S4-4

hid

3

374

1,930

S4-4

hid

4

335

1,996

S4-5

hid

1

414

2,311

S4-5

hid

2

454

2,256

S4-5

hid

3

389

2,278

S4-5

hid

4

415

2,191

S4-1

head

1

652

1,811

S4-1

head

2

638

1,838

S4-1

head

3

656

2,006

S4-1

head

4

547

1,960

S4-2

head

1

509

1,795

S4-2

head

2

409

1,845

S4-2

head

3

563

2,023

S4-2

head

4

473

1,979

S4-3

head

1

609

2,097

S4-3

head

2

721

2,130

S4-3

head

3

720

2,191

S4-3

head

4

706

1,950

S4-4

head

1

518

1,677

S4-4

head

2

444

1,985

S4-4

head

3

523

1,694

S4-4

head

4

515

1,726

S4-5

head

1

682

1,862

S4-5

head

2

737

1,912

S4-5

head

3

704

1,963

S4-5

head

4

673

2,070

S4-1

had

1

821

1,493

S4-1

had

2

845

1,458

S4-1

had

3

713

1,431

S4-1

had

4

875

1,470

S4-2

had

1

889

1,472

S4-2

had

2

805

1,379

S4-2

had

3

755

1,413

S4-2

had

4

839

1,345

S4-3

had

1

1,121

1,712

S4-3

had

2

1,021

1,578

S4-3

had

3

1,113

1,724

S4-3

had

4

1,127

1,695

S4-4

had

1

963

1,370

S4-4

had

2

1,049

1,459

S4-4

had

3

956

1,292

S4-4

had

4

1,009

1,396

S4-5

had

1

786

1,375

S4-5

had

2

839

1,476

S4-5

had

3

931

1,468

S4-5

had

4

870

1,457

S4-1

hard

1

627

1,052

S4-1

hard

2

593

1,043

S4-1

hard

3

595

1,109

S4-1

hard

4

635

1,078

S4-2

hard

1

621

1,033

S4-2

hard

2

570

1,040

S4-2

hard

3

570

1,007

S4-2

hard

4

503

970

S4-3

hard

1

574

994

S4-3

hard

2

587

996

S4-3

hard

3

455

986

S4-3

hard

4

570

989

S4-4

hard

1

531

1,044

S4-4

hard

2

620

1,092

S4-4

hard

3

637

1,066

S4-4

hard

4

562

1,081

S4-5

hard

1

668

1,038

S4-5

hard

2

699

1,027

S4-5

hard

3

762

1,068

S4-5

hard

4

704

1,090

S4-1

hod

1

517

900

S4-1

hod

2

518

994

S4-1

hod

3

526

886

S4-1

hod

4

526

850

S4-2

hod

1

486

870

S4-2

hod

2

486

873

S4-2

hod

3

470

872

S4-2

hod

4

443

907

S4-3

hod

1

457

797

S4-3

hod

2

490

816

S4-3

hod

3

508

890

S4-3

hod

4

496

811

S4-4

hod

1

452

830

S4-4

hod

2

451

782

S4-4

hod

3

454

826

S4-4

hod

4

433

891

S4-5

hod

1

520

878

S4-5

hod

2

410

851

S4-5

hod

3

486

855

S4-5

hod

4

533

919

S4-1

hoard

1

405

629

S4-1

hoard

2

389

644

S4-1

hoard

3

400

583

S4-1

hoard

4

396

561

S4-2

hoard

1

368

619

S4-2

hoard

2

370

714

S4-2

hoard

3

360

632

S4-2

hoard

4

353

637

S4-3

hoard

1

324

524

S4-3

hoard

2

367

573

S4-3

hoard

3

381

614

S4-3

hoard

4

391

571

S4-4

hoard

1

386

654

S4-4

hoard

2

335

570

S4-4

hoard

3

452

535

S4-4

hoard

4

469

654

S4-5

hoard

1

369

755

S4-5

hoard

2

419

540

S4-5

hoard

3

445

753

S4-5

hoard

4

454

830

S4-1

hood

1

454

1,550

S4-1

hood

2

420

1,741

S4-1

hood

3

409

1,657

S4-1

hood

4

426

1,610

S4-2

hood

1

400

1,200

S4-2

hood

2

452

1,332

S4-2

hood

3

423

1,198

S4-2

hood

4

381

1,158

S4-3

hood

1

421

1,222

S4-3

hood

2

424

1,150

S4-3

hood

3

446

1,252

S4-3

hood

4

436

1,174

S4-4

hood

1

353

1,087

S4-4

hood

2

370

1,100

S4-4

hood

3

369

1,153

S4-4

hood

4

357

1,281

S4-5

hood

1

453

1,275

S4-5

hood

2

393

1,223

S4-5

hood

3

417

1,125

S4-5

hood

4

453

1,245

S4-1

whod

1

331

1,983

S4-1

whod

2

314

1,956

S4-1

whod

3

311

1,984

S4-1

whod

4

311

1,796

S4-2

whod

1

307

1,416

S4-2

whod

2

330

1,545

S4-2

whod

3

251

1,462

S4-2

whod

4

278

1,314

S4-3

whod

1

294

1,315

S4-3

whod

2

272

1,527

S4-3

whod

3

250

1,589

S4-3

whod

4

281

1,576

S4-4

whod

1

297

1,446

S4-4

whod

2

229

1,241

S4-4

whod

3

262

1,605

S4-4

whod

4

303

1,589

S4-5

whod

1

332

1,917

S4-5

whod

2

251

1,660

S4-5

whod

3

268

1,778

S4-5

whod

4

302

1,627

S4-1

hud

1

592

1,242

S4-1

hud

2

668

1,163

S4-1

hud

3

690

1,202

S4-1

hud

4

654

1,195

S4-2

hud

1

539

1,301

S4-2

hud

2

471

1,239

S4-2

hud

3

537

1,390

S4-2

hud

4

587

1,347

S4-3

hud

1

922

1,224

S4-3

hud

2

805

1,191

S4-3

hud

3

805

1,107

S4-3

hud

4

738

1,124

S4-4

hud

1

500

1,224

S4-4

hud

2

620

1,107

S4-4

hud

3

587

1,207

S4-4

hud

4

572

1,218

S4-5

hud

1

670

1,177

S4-5

hud

2

679

1,155

S4-5

hud

3

784

1,178

S4-5

hud

4

744

1,170

S4-1

herd

1

590

1,498

S4-1

herd

2

571

1,537

S4-1

herd

3

520

1,532

S4-1

herd

4

514

1,545

S4-2

herd

1

479

1,419

S4-2

herd

2

500

1,399

S4-2

herd

3

420

1,391

S4-2

herd

4

466

1,417

S4-3

herd

1

462

1,252

S4-3

herd

2

428

1,224

S4-3

herd

3

464

1,282

S4-3

herd

4

515

1,278

S4-4

herd

1

461

1,334

S4-4

herd

2

449

1,273

S4-4

herd

3

431

1,347

S4-4

herd

4

486

1,344

S4-5

herd

1

537

1,366

S4-5

herd

2

499

1,364

S4-5

herd

3

533

1,314

S4-5

herd

4

546

1,332

Combining and Processing the Data

We first combine the two datasets and rename the key columns:

Code 
voweldata <- rbind(nns, ns) |>
  dplyr::rename(Speaker = subject, Word = item) |>
  dplyr::mutate(
    Speaker = dplyr::case_when(
      Speaker == "ms"    ~ "L1-German learner",
      Speaker == "rpspk" ~ "RP native speakers",
      TRUE               ~ Speaker
    )
  )

Next we add IPA symbols for each word:

Code 
voweldata <- voweldata |>
  dplyr::mutate(
    ipa = dplyr::case_when(
      Word == "had"   ~ "\u00E6",
      Word == "hard"  ~ "\u0251",
      Word == "head"  ~ "e",
      Word == "heed"  ~ "i",
      Word == "herd"  ~ "\u025C",
      Word == "hid"   ~ "\u026A",
      Word == "hoard" ~ "\u0254",
      Word == "hod"   ~ "\u0252",
      Word == "hood"  ~ "\u028A",
      Word == "hud"   ~ "\u028C",
      Word == "whod"  ~ "u",
      TRUE            ~ "?"
    )
  )

Finally, we compute per-vowel mean F1 and F2 within each speaker group and drop the trial column:

Code 
voweldata <- voweldata |>
  dplyr::group_by(Speaker, Word) |>
  dplyr::mutate(
    F1_mean = mean(F1),
    F2_mean = mean(F2)
  ) |>
  dplyr::ungroup() |>
  dplyr::select(-dplyr::any_of("trial"))

The resulting table of per-vowel means looks like this:

Speaker

Word

ipa

F1_mean

F2_mean

L1-German learner

had

æ

727.3

1,903.5

L1-German learner

hard

ɑ

788.2

1,466.5

L1-German learner

head

e

652.8

2,067.7

L1-German learner

heed

i

283.9

2,800.5

L1-German learner

herd

ɜ

531.8

1,743.0

L1-German learner

hid

ɪ

426.5

2,411.4

L1-German learner

hoard

ɔ

579.4

990.5

L1-German learner

hod

ɒ

688.2

1,227.8

L1-German learner

hood

ʊ

435.2

1,382.5

L1-German learner

hud

ʌ

672.6

1,597.4

L1-German learner

whod

u

355.8

1,105.3

S4-1

had

æ

813.5

1,463.0

S4-1

hard

ɑ

612.5

1,070.5

S4-1

head

e

623.2

1,903.8

S4-1

heed

i

296.0

2,343.8

S4-1

herd

ɜ

548.8

1,528.0

S4-1

hid

ɪ

442.8

2,127.0

S4-1

hoard

ɔ

397.5

604.2

S4-1

hod

ɒ

521.8

907.5

S4-1

hood

ʊ

427.2

1,639.5

S4-1

hud

ʌ

651.0

1,200.5

S4-1

whod

u

316.8

1,929.8

S4-2

had

æ

822.0

1,402.2

S4-2

hard

ɑ

566.0

1,012.5

S4-2

head

e

488.5

1,910.5

S4-2

heed

i

272.0

2,239.0

S4-2

herd

ɜ

466.2

1,406.5

S4-2

hid

ɪ

360.2

2,108.5

S4-2

hoard

ɔ

362.8

650.5

S4-2

hod

ɒ

471.2

880.5

S4-2

hood

ʊ

414.0

1,222.0

S4-2

hud

ʌ

533.5

1,319.2

S4-2

whod

u

291.5

1,434.2

S4-3

had

æ

1,095.5

1,677.2

S4-3

hard

ɑ

546.5

991.2

S4-3

head

e

689.0

2,092.0

S4-3

heed

i

259.5

2,481.0

S4-3

herd

ɜ

467.2

1,259.0

S4-3

hid

ɪ

394.0

2,402.2

S4-3

hoard

ɔ

365.8

570.5

S4-3

hod

ɒ

487.8

828.5

S4-3

hood

ʊ

431.8

1,199.5

S4-3

hud

ʌ

817.5

1,161.5

S4-3

whod

u

274.2

1,501.8

S4-4

had

æ

994.2

1,379.2

S4-4

hard

ɑ

587.5

1,070.8

S4-4

head

e

500.0

1,770.5

S4-4

heed

i

291.0

1,996.5

S4-4

herd

ɜ

456.8

1,324.5

S4-4

hid

ɪ

349.2

1,975.0

S4-4

hoard

ɔ

410.5

603.2

S4-4

hod

ɒ

447.5

832.2

S4-4

hood

ʊ

362.2

1,155.2

S4-4

hud

ʌ

569.8

1,189.0

S4-4

whod

u

272.8

1,470.2

S4-5

had

æ

856.5

1,444.0

S4-5

hard

ɑ

708.2

1,055.8

S4-5

head

e

699.0

1,951.8

S4-5

heed

i

262.2

2,627.8

S4-5

herd

ɜ

528.8

1,344.0

S4-5

hid

ɪ

418.0

2,259.0

S4-5

hoard

ɔ

421.8

719.5

S4-5

hod

ɒ

487.2

875.8

S4-5

hood

ʊ

429.0

1,217.0

S4-5

hud

ʌ

719.2

1,170.0

S4-5

whod

u

288.2

1,745.5

Vowel Normalisation

Section Overview

What you will learn: Why raw formant values cannot always be compared across speakers; what Lobanov and Nearey normalisation are; how to apply both methods in R; and when each is appropriate

Why Normalise?

Raw formant values (in Hz) reflect both the vowel category and the speaker’s individual vocal tract size. Adult males have longer vocal tracts than females, who in turn have longer vocal tracts than children — this means that the same vowel produced by a male and a female speaker will have different Hz values even if both are native speakers of the same variety.

When comparing a single speaker to themselves over time or two speakers from the same demographic group (as in this showcase), the effect is modest and raw Hz values are often acceptable. However, when comparing speakers of different sexes, ages, or sizes — or when pooling across many speakers — normalisation is essential to separate linguistic (vowel category) variation from anatomical (vocal tract size) variation.

Method

Description

Best for

Lobanov (z-score)

z-score each formant within each speaker: (F − mean) / SD

Cross-sex, cross-age comparison; large multi-speaker corpora

Nearey (log-mean)

Subtract speaker's log-mean from log-transformed formants

Corpora where speaker means differ but variances are similar

Bark scaling

Convert Hz to Bark scale using a perceptual transform

Perceptually motivated analyses

Watt & Fabricius

Reference vowel triangle method — preserves relative distances

Dialect surveys; comparison of vowel system geometry

Raw Hz (no normalisation)

No transformation — absolute Hz values retained

Single speaker; speakers matched for age and sex

Lobanov Normalisation

The Lobanov method z-scores each formant separately within each speaker — it is the most widely used method for cross-speaker comparison (Lobanov 1971):

\[F_{\text{Lobanov}} = \frac{F - \bar{F}_{\text{speaker}}}{SD_{F,\text{speaker}}}\]

We first compute the z-scored formant values:

Code 
voweldata_lob <- voweldata |>
  dplyr::group_by(Speaker) |>
  dplyr::mutate(
    F1_lob = (F1 - mean(F1, na.rm = TRUE)) / sd(F1, na.rm = TRUE),
    F2_lob = (F2 - mean(F2, na.rm = TRUE)) / sd(F2, na.rm = TRUE)
  ) |>
  dplyr::ungroup()

Then we add per-vowel means on the normalised scale:

Code 
voweldata_lob <- voweldata_lob |>
  dplyr::group_by(Speaker, Word) |>
  dplyr::mutate(
    F1_lob_mean = mean(F1_lob),
    F2_lob_mean = mean(F2_lob)
  ) |>
  dplyr::ungroup()

Nearey Normalisation

The Nearey method applies a log transformation and subtracts the speaker’s log-mean across all formants (Nearey 1989). It is less aggressive than Lobanov and preserves more of the relative distances between vowels:

\[F_{\text{Nearey}} = \log(F) - \frac{1}{n}\sum_{k=1}^{n}\overline{\log(F_k)}\]

We compute the log grand mean per speaker, then subtract it from each log-transformed formant:

Code 
voweldata_nea <- voweldata |>
  dplyr::group_by(Speaker) |>
  dplyr::mutate(
    log_grand_mean = mean(c(log(F1), log(F2)), na.rm = TRUE),
    F1_nea = log(F1) - log_grand_mean,
    F2_nea = log(F2) - log_grand_mean
  ) |>
  dplyr::ungroup()

Then we add per-vowel means on the Nearey scale:

Code 
voweldata_nea <- voweldata_nea |>
  dplyr::group_by(Speaker, Word) |>
  dplyr::mutate(
    F1_nea_mean = mean(F1_nea),
    F2_nea_mean = mean(F2_nea)
  ) |>
  dplyr::ungroup()

Visualising Vowel Spaces

Section Overview

What you will learn: How to produce a basic vowel chart with ggplot2; how to add confidence ellipses; how to create a publication-quality comparison chart with native-speaker reference ellipses; and how to plot normalised vowel spaces

Basic Vowel Chart

We first separate the two speaker groups so that we can apply different ellipse levels to each:

Code 
ns_data  <- voweldata |> dplyr::filter(Speaker == "RP native speakers")
nns_data <- voweldata |> dplyr::filter(Speaker == "L1-German learner")

We also prepare a deduplicated means table for the IPA label layer:

Code 
vowel_means <- voweldata |>
  dplyr::distinct(Speaker, Word, ipa, F1_mean, F2_mean)

Now we build the chart in layers — points, IPA labels, and ellipses — then apply the reversed axes and styling:

Code 
ggplot(voweldata,
       aes(x = F2, y = F1, color = Speaker, fill = Speaker)) +
  geom_point(alpha = 0.15, size = 1.5) +
  geom_text(
    data = vowel_means,
    aes(x = F2_mean, y = F1_mean, label = ipa),
    fontface = "bold", size = 4.5, show.legend = FALSE
  ) +
  stat_ellipse(data = ns_data,  level = 0.50, geom = "polygon",
               alpha = 0.08, linetype = "dashed") +
  stat_ellipse(data = nns_data, level = 0.95, geom = "polygon",
               alpha = 0.08) +
  scale_x_reverse(breaks = seq(500, 3000, 500)) +
  scale_y_reverse(breaks = seq(200, 1000, 100)) +
  scale_color_manual(values = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  scale_fill_manual(values  = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  theme_bw() +
  theme(
    legend.position  = "top",
    legend.title     = element_blank(),
    panel.grid.major = element_blank(),
    panel.grid.minor = element_blank(),
    axis.title       = element_text(size = 11)
  ) +
  labs(
    title   = "Vowel space: L1-German learner vs. RP native speakers",
    x       = "F2 (Hz) →",
    y       = "F1 (Hz) →",
    caption = "Axes reversed to mirror articulatory vowel quadrilateral.\nDashed ellipse = NS 50% region; solid ellipse = learner 95% region."
  )

Publication-Quality Chart with SD Ellipses

For the native-speaker reference data, we have standard deviations provided directly by Hawkins and Midgley (2005). We can use these to draw standard-deviation ellipses that represent the spread of the native-speaker vowel space more accurately than stat_ellipse().

We first prepare the native-speaker means data, adding IPA labels:

Code 
# Reference means and SDs from Hawkins & Midgley (2005)
# Five L1-RP speakers aged 20-25
ns_means <- data.frame(
  Word = c("had","hard","head","heed","herd","hid","hoard","hod","hood","hud","whod"),
  F1   = c(916.35, 604.15, 599.95, 276.15, 493.55, 392.85, 391.65, 483.10, 412.85, 658.20, 288.70),
  F2   = c(1473.15, 1040.15, 1925.70, 2337.60, 1372.40, 2174.35, 629.60, 864.90, 1286.65, 1208.05, 1616.30),
  F1sd = c(124.30, 70.92, 102.23, 25.48, 47.41, 40.84, 39.71, 35.48, 32.98, 116.15, 30.19),
  F2sd = c(119.44, 40.06, 143.60, 223.42, 95.95, 166.86, 81.19, 48.50, 193.70, 72.52, 225.74),
  ipa  = c("\u00E6", "\u0251", "e", "i", "\u025C", "\u026A",
           "\u0254", "\u0252", "\u028A", "\u028C", "u")
)

Next we extract the learner means and join them with the NS means to compute displacement arrows:

Code 
nns_means <- voweldata |>
  dplyr::filter(Speaker == "L1-German learner") |>
  dplyr::distinct(Word, ipa, F1_mean, F2_mean)

arrow_data <- dplyr::left_join(nns_means, ns_means, by = c("Word", "ipa"))

Now we build the chart with ggforce::geom_ellipse() for the SD ellipses and geom_segment() for the displacement arrows:

Code 
ggplot() +
  ggforce::geom_ellipse(
    data = ns_means,
    aes(x0 = F2, y0 = F1, a = F2sd, b = F1sd, angle = 0),
    fill = "#2C3E50", alpha = 0.07, color = "#2C3E50",
    linetype = "dashed", linewidth = 0.5
  ) +
  geom_text(
    data = ns_means,
    aes(x = F2, y = F1, label = ipa),
    color = "#2C3E50", fontface = "bold", size = 5
  ) +
  geom_text(
    data = nns_means,
    aes(x = F2_mean, y = F1_mean, label = ipa),
    color = "#E67E22", fontface = "bold", size = 5
  ) +
  geom_segment(
    data = arrow_data,
    aes(x = F2_mean, y = F1_mean, xend = F2, yend = F1),
    arrow = arrow(length = unit(0.15, "cm"), type = "closed"),
    color = "grey50", linewidth = 0.4, alpha = 0.7
  ) +
  scale_x_reverse(breaks = seq(500, 3000, 500),
                  labels = scales::label_number()) +
  scale_y_reverse(breaks = seq(200, 1000, 100)) +
  theme_bw() +
  theme(
    panel.grid.major = element_blank(),
    panel.grid.minor = element_blank(),
    axis.title       = element_text(size = 11),
    plot.caption     = element_text(size = 8, color = "grey40")
  ) +
  labs(
    title   = "Vowel space comparison: L1-German learner (orange) vs. RP native speakers (dark)",
    x       = "F2 (Hz) →",
    y       = "F1 (Hz) →",
    caption = "Dashed ellipses = ±1 SD of native RP speaker vowel spaces (Hawkins & Midgley 2005).\nArrows indicate direction and magnitude of learner deviation from native norms."
  )

The arrows make displacement patterns immediately interpretable: vowels where the arrow is long indicate larger deviations from the native-speaker target. The chart shows that the L1-German learner’s /iː/ (heed) and /ɔː/ (hoard) are displaced most substantially from RP norms.

Faceted Chart by Vowel

A faceted view allows close inspection of each vowel’s token distribution. We first filter out any rows with missing formant values, then pass the data to ggplot():

Code 
voweldata_clean <- voweldata |>
  dplyr::filter(!is.na(F1), !is.na(F2)) |>
  dplyr::mutate(facet_label = paste0(ipa, " (", Word, ")"))
Code 
ggplot(voweldata_clean,
       aes(x = F2, y = F1, color = Speaker, fill = Speaker)) +
  geom_point(alpha = 0.4, size = 2) +
  stat_ellipse(level = 0.80, geom = "polygon", alpha = 0.12) +
  facet_wrap(~ facet_label, ncol = 4, scales = "free") +
  scale_x_reverse() +
  scale_y_reverse() +
  scale_color_manual(values = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  scale_fill_manual(values  = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  theme_bw(base_size = 9) +
  theme(
    legend.position  = "top",
    legend.title     = element_blank(),
    panel.grid       = element_blank(),
    strip.background = element_rect(fill = "grey92"),
    strip.text       = element_text(face = "bold")
  ) +
  labs(
    title = "Per-vowel token distributions: learner vs. native speakers",
    x = "F2 (Hz) →", y = "F1 (Hz) →"
  )

Normalised Vowel Charts

Lobanov

We first split the Lobanov-normalised data by speaker group:

Code 
ns_lob  <- voweldata_lob |> dplyr::filter(Speaker == "RP native speakers")
nns_lob <- voweldata_lob |> dplyr::filter(Speaker == "L1-German learner")

lob_means <- voweldata_lob |>
  dplyr::distinct(Speaker, Word, ipa, F1_lob_mean, F2_lob_mean)
Code 
ggplot(voweldata_lob,
       aes(x = F2_lob, y = F1_lob, color = Speaker, fill = Speaker)) +
  geom_point(alpha = 0.15, size = 1.5) +
  geom_text(
    data = lob_means,
    aes(x = F2_lob_mean, y = F1_lob_mean, label = ipa),
    fontface = "bold", size = 4.5, show.legend = FALSE
  ) +
  stat_ellipse(data = ns_lob,  level = 0.50, geom = "polygon",
               alpha = 0.08, linetype = "dashed") +
  stat_ellipse(data = nns_lob, level = 0.95, geom = "polygon",
               alpha = 0.08) +
  scale_x_reverse() +
  scale_y_reverse() +
  scale_color_manual(values = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  scale_fill_manual(values  = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  theme_bw() +
  theme(
    legend.position = "top",
    legend.title    = element_blank(),
    panel.grid      = element_blank()
  ) +
  labs(
    title   = "Lobanov-normalised vowel space",
    x       = "F2 (Lobanov z-score) →",
    y       = "F1 (Lobanov z-score) →",
    caption = "Lobanov normalisation removes vocal tract size differences between speakers."
  )

Nearey

We do the same for the Nearey-normalised data:

Code 
ns_nea  <- voweldata_nea |> dplyr::filter(Speaker == "RP native speakers")
nns_nea <- voweldata_nea |> dplyr::filter(Speaker == "L1-German learner")

nea_means <- voweldata_nea |>
  dplyr::distinct(Speaker, Word, ipa, F1_nea_mean, F2_nea_mean)
Code 
ggplot(voweldata_nea,
       aes(x = F2_nea, y = F1_nea, color = Speaker, fill = Speaker)) +
  geom_point(alpha = 0.15, size = 1.5) +
  geom_text(
    data = nea_means,
    aes(x = F2_nea_mean, y = F1_nea_mean, label = ipa),
    fontface = "bold", size = 4.5, show.legend = FALSE
  ) +
  stat_ellipse(data = ns_nea,  level = 0.50, geom = "polygon",
               alpha = 0.08, linetype = "dashed") +
  stat_ellipse(data = nns_nea, level = 0.95, geom = "polygon",
               alpha = 0.08) +
  scale_x_reverse() +
  scale_y_reverse() +
  scale_color_manual(values = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  scale_fill_manual(values  = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  theme_bw() +
  theme(
    legend.position = "top",
    legend.title    = element_blank(),
    panel.grid      = element_blank()
  ) +
  labs(
    title   = "Nearey-normalised vowel space",
    x       = "F2 (Nearey log-normalised) →",
    y       = "F1 (Nearey log-normalised) →",
    caption = "Nearey normalisation uses log-mean subtraction — less aggressive than Lobanov."
  )

Which Normalisation Should I Use?

For a single-speaker showcase like this one, raw Hz values are perfectly adequate and most interpretable. For a study comparing speakers across sexes or ages, Lobanov is the standard choice in sociophonetics. Nearey is preferred when the absolute distances between vowels are theoretically important, as it is less aggressive. If in doubt, report both and note whether the substantive conclusions differ.

Comparing the Three Schemes Side by Side

We assemble one row of means per speaker per vowel under each scheme, tag each with a normalisation label, and combine into a single data frame:

Code 
means_raw <- voweldata |>
  dplyr::distinct(Speaker, Word, ipa, F1_mean, F2_mean) |>
  dplyr::mutate(normalisation = "Raw Hz")
Code 
means_lob <- voweldata_lob |>
  dplyr::distinct(Speaker, Word, ipa, F1_lob_mean, F2_lob_mean) |>
  dplyr::rename(F1_mean = F1_lob_mean, F2_mean = F2_lob_mean) |>
  dplyr::mutate(normalisation = "Lobanov")
Code 
means_nea <- voweldata_nea |>
  dplyr::distinct(Speaker, Word, ipa, F1_nea_mean, F2_nea_mean) |>
  dplyr::rename(F1_mean = F1_nea_mean, F2_mean = F2_nea_mean) |>
  dplyr::mutate(normalisation = "Nearey")
Code 
means_all <- dplyr::bind_rows(means_raw, means_lob, means_nea) |>
  dplyr::mutate(normalisation = factor(normalisation,
                                       levels = c("Raw Hz", "Lobanov", "Nearey")))
Code 
ggplot(means_all,
       aes(x = F2_mean, y = F1_mean, color = Speaker, label = ipa)) +
  geom_text(fontface = "bold", size = 3.5) +
  facet_wrap(~ normalisation, scales = "free") +
  scale_x_reverse() +
  scale_y_reverse() +
  scale_color_manual(values = c("L1-German learner"  = "#E67E22",
                                "RP native speakers" = "#2C3E50")) +
  theme_bw(base_size = 10) +
  theme(
    legend.position  = "top",
    legend.title     = element_blank(),
    panel.grid       = element_blank(),
    strip.background = element_rect(fill = "grey92"),
    strip.text       = element_text(face = "bold")
  ) +
  labs(
    title   = "Vowel means under three normalisation schemes",
    x = "F2 →", y = "F1 →",
    caption = "Raw Hz values, Lobanov z-scores, and Nearey log-normalised values plotted at mean positions."
  )

Interpretation

The vowel charts reveal several patterns in the L1-German learner’s English vowel production relative to the RP native-speaker norms:

  • /iː/ (heed): The learner’s vowel is produced more centrally — lower F2 — than the RP target, suggesting insufficient fronting of the tongue.
  • /ɔː/ (hoard): The learner’s vowel is substantially higher (lower F1) than the RP target, consistent with German-influenced vowel transfer where the back rounded vowels are produced in a higher tongue position.
  • /æ/ (had): The vowel spaces differ considerably, which may reflect the absence of this vowel category in German — German speakers learning English often produce /æ/ too high.
  • /ʊ/ (hood) and /ʌ/ (hud): These vowels are produced relatively close to the RP norms, suggesting that for this learner these categories are already well-established.
  • Overall vowel space size: The learner’s ellipses tend to be larger than the native-speaker 50% ellipses, indicating more within-category variability — a common feature of L2 production.

These observations are consistent with the broader SLA literature on L1-German speakers’ English phonology (see Paganus et al. 2006 for similar analyses in a language-learning feedback context).

Citation and Session Info

Citation

Martin Schweinberger. 2026. Creating Vowel Charts in R. The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia. url: https://ladal.edu.au/tutorials/vowelchart/vowelchart.html (Version 2026.03.27), doi: 10.5281/zenodo.19242479.

@manual{martinschweinberger2026creating,
  author       = {Martin Schweinberger},
  title        = {Creating Vowel Charts in R},
  year         = {2026},
  note         = {https://ladal.edu.au/tutorials/vowelchart/vowelchart.html},
  organization = {The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia},
  edition      = {2026.03.27}
  doi      = {10.5281/zenodo.19242479}
}
Code 
sessionInfo()
R version 4.4.2 (2024-10-31 ucrt)
Platform: x86_64-w64-mingw32/x64
Running under: Windows 11 x64 (build 26200)

Matrix products: default


locale:
[1] LC_COLLATE=English_United States.utf8 
[2] LC_CTYPE=English_United States.utf8   
[3] LC_MONETARY=English_United States.utf8
[4] LC_NUMERIC=C                          
[5] LC_TIME=English_United States.utf8    

time zone: Australia/Brisbane
tzcode source: internal

attached base packages:
[1] stats     graphics  grDevices datasets  utils     methods   base     

other attached packages:
 [1] scales_1.4.0     ggforce_0.4.2    flextable_0.9.11 lubridate_1.9.4 
 [5] forcats_1.0.0    stringr_1.5.1    dplyr_1.2.0      purrr_1.0.4     
 [9] readr_2.1.5      tidyr_1.3.2      tibble_3.2.1     ggplot2_4.0.2   
[13] tidyverse_2.0.0 

loaded via a namespace (and not attached):
 [1] gtable_0.3.6            xfun_0.56               htmlwidgets_1.6.4      
 [4] tzdb_0.4.0              vctrs_0.7.1             tools_4.4.2            
 [7] generics_0.1.3          klippy_0.0.0.9500       pkgconfig_2.0.3        
[10] data.table_1.17.0       RColorBrewer_1.1-3      S7_0.2.1               
[13] assertthat_0.2.1        uuid_1.2-1              lifecycle_1.0.5        
[16] compiler_4.4.2          farver_2.1.2            textshaping_1.0.0      
[19] codetools_0.2-20        fontquiver_0.2.1        fontLiberation_0.1.0   
[22] htmltools_0.5.9         yaml_2.3.10             pillar_1.10.1          
[25] MASS_7.3-61             openssl_2.3.2           fontBitstreamVera_0.1.1
[28] tidyselect_1.2.1        zip_2.3.2               digest_0.6.39          
[31] stringi_1.8.4           labeling_0.4.3          polyclip_1.10-7        
[34] fastmap_1.2.0           grid_4.4.2              cli_3.6.4              
[37] magrittr_2.0.3          patchwork_1.3.0         withr_3.0.2            
[40] gdtools_0.5.0           timechange_0.3.0        rmarkdown_2.30         
[43] officer_0.7.3           askpass_1.2.1           ragg_1.3.3             
[46] hms_1.1.3               evaluate_1.0.3          knitr_1.51             
[49] rlang_1.1.7             Rcpp_1.1.1              glue_1.8.0             
[52] tweenr_2.0.3            xml2_1.3.6              renv_1.1.7             
[55] rstudioapi_0.17.1       jsonlite_1.9.0          R6_2.6.1               
[58] systemfonts_1.3.1
AI Transparency Statement

This tutorial was written with the assistance of Claude (claude.ai), a large language model created by Anthropic. Claude was used to expand and restructure a shorter existing LADAL showcase on vowel chart production. All content was reviewed and approved by Martin Schweinberger, who takes full responsibility for its accuracy.

Back to top

Back to LADAL home

References

Hawkins, Sarah, and Jonathan Midgley. 2005. “Formant Frequencies of RP Monophthongs in Four Age Groups of Speakers.” Journal of the International Phonetic Association 35 (2): 183–99. https://doi.org/https://doi.org/10.1017/s0025100305002124.
Johnson, Keith. 2003. Acoustic and Auditory Phonetics. Vol. 61. Malden, MA: Blackwell. https://doi.org/https://doi.org/10.1159/000078663.
Ladefoged, Peter. 1996. Elements of Acoustic Phonetics. Chigago: University of Chicago Press. https://doi.org/https://doi.org/10.7208/chicago/9780226191010.001.0001.
Lobanov, Boris M. 1971. “Classification of Russian Vowels Spoken by Different Speakers.” The Journal of the Acoustical Society of America 49 (2B): 606–8.
Nearey, Terrance M. 1989. “Static, Dynamic, and Relational Properties in Vowel Perception.” The Journal of the Acoustical Society of America 85 (5): 2088–2113.
Paganus, Annu, Vesa-Petteri Mikkonen, Tomi Mäntylä, Sami Nuuttila, Jouni Isoaho, Olli Aaltonen, and Tapio Salakoski. 2006. “The Vowel Game: Continuous Real-Time Visualization for Pronunciation Learning with Vowel Charts.” In International Conference on Natural Language Processing (in Finland), 696–703. Springer. https://doi.org/https://doi.org/10.1007/11816508_69.
Peterson, Gordon E, and Harold L Barney. 1952. “Control Methods Used in a Study of the Vowels.” The Journal of the Acoustical Society of America 24 (2): 175–84.
Rogers, Henry. 2014. The Sounds of Language: An Introduction to Phonetics. Routledge.
Zsiga, Elizabeth C. 2012. The Sounds of Language: An Introduction to Phonetics and Phonology. John Wiley & Sons.