Extracting accent features in spoken Brazilian Portuguese without sociolinguistic labels

Overview

Research Questions & Answers

Limitation: these three markers do not fully disambiguate all Brazilian accents; broader coverage requires further markers (e.g., /l/ palatalization, vowel openness) and prosodic cues.

The Three Accent Markers

Speakers are classified according to their realization of three phonological variables, which map to distinct regional varieties of Brazilian Portuguese.

Listen: Marker Realizations

Each clip is a 160 ms window centered on a ZIPA-detected phoneme spike, the exact unit used for feature extraction in the paper. Use the 0.5× button to slow playback for easier perception.

/s/ Coda

What to listen for: chiado produces a “sh”-like sound with energy spread lower and broader in frequency; sibilant concentrates energy higher (>4 kHz), giving a sharper “s”. The spectrograms show this clearly: chiado has a diffuse frication band, sibilant a narrow bright ridge at the top.

Chiado

chiado [ʃ] · #1

chiado [ʃ] · #2

chiado [ʃ] · #3

Sibilant

sibilant [s] · #1

sibilant [s] · #2

sibilant [s] · #3

/r/ Coda

What to listen for: carioca sounds like a breathy velar or glottal fricative (similar to German “ch”); tap is a quick flap similar to the British English “r” in “very”; caipira is the retroflexed “r” of interior São Paulo, colored and retroflex.

Carioca

carioca [x]/[h] · #1

carioca [x]/[h] · #2

carioca [x]/[h] · #3

Tap

tap [ɾ] · #1

tap [ɾ] · #2

tap [ɾ] · #3

Caipira

caipira [ɻ] · #1

caipira [ɻ] · #2

caipira [ɻ] · #3

/d/–/t/ Palatalization

What to listen for: palatalized speakers produce “djia” / “tchia” for words like dia / tia (affricate followed by vowel); non-palatalized speakers keep a plain “d” / “t” before /i/. The spectrogram shows an extra frication burst after the stop in palatalized examples.

Palatalized

palatalized [dʒ]/[tʃ] · #1

palatalized [dʒ]/[tʃ] · #2

palatalized [dʒ]/[tʃ] · #3

Non-palatalized

non-palatalized [d]/[t] · #1

non-palatalized [d]/[t] · #2

non-palatalized [d]/[t] · #3

Full Ablation Study

Best speaker accuracy per feature set and task (best classifier per cell selected among XGBoost, logistic regression, and linear SVM), under balanced stratified cross-validation with ~5 test speakers per fold and speaker isolation between folds. The mean ± std is computed across folds. This is the full study summarized in the paper (referenced there as companion page footnote 5).

Best speaker accuracy per feature set (best classifier selected per cell). Values show mean ± std across balanced stratified cross-validation folds (∼5 test speakers/fold). Heatmap colors match the paper.

Reproducing This Work

Corpora

Every corpus used in this paper has a bundled download script — CETUC/Alcaim, CORAA (including its C-ORAL Brasil and TED subsets), Common Voice, CoLingPB, CML-TTS, Certas Palavras, GneutralSpeech, BRSpeech-DF, TAGARELA and Sotaque Brasileiro — with the sole exception of the Ynoguti corpus, which is distributed only by its authors. Each script fetches the corpus and organizes it into the per-speaker layout the pipeline expects:

git clone https://github.com/pedrohlopes/pt-br-accent-toolbox
cd pt-br-accent-toolbox && pip install -e ".[datasets]"

python tools/download_datasets.py list          # every corpus, with its source and caveats
python tools/download_alcaim.py --sentences balanced50
python tools/download_coraa.py --split train --subsets nurc,coral,ted
python tools/download_common_voice.py

Alcaim / CETUC is streamed and filtered on the fly rather than downloaded whole, and --sentences balanced50 reproduces the phonetically balanced 50-sentence subset used throughout the paper. Corpora that are only available on request from their authors are listed with contact pointers instead of a script.

Speaker annotations

The 115 hand-labelled speakers — nine corpora, three markers (/s/ coda, /r/ coda, /d/–/t/ palatalization) — ship with the package as CSVs under pt_br_accent_toolbox/data/annotations/: the raw per-annotator rows, the per-speaker majority label with agreement counts, and a summary table. Speaker IDs are the directory names the download scripts produce, so the labels join directly against a local copy of each corpus.

Read that folder's README before using the labels. In particular: 22 of the Alcaim speakers carry only an /s/-coda label and were selected because of how they realize it, so that marker is not a random sample of the corpus; and only three speakers were labelled by two annotators, which is too few to report as an inter-annotator agreement figure.

Citation

@misc{leite2026accentfeatures,
  title         = {Extracting accent features in spoken Brazilian Portuguese
                   without sociolinguistic labels},
  author        = {Leite, Pedro H. L. and Valadares, Pedro Benevenuto and
                   Biscainho, Luiz W. P.},
  year          = {2026},
  eprint        = {2605.30457},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS}
}