kooma-ai/bambara-orthography

Bambara orthography and text normalization

Writing conventions for Bambara (Bamanankan) in the standard Latin alphabet, a normalization table from common ASCII spellings to the standard, and a reference normalizer in Python. Maintained by Kooma.

Why this exists: Bambara is written in many ways in the wild — with or without ɛ/ɔ/ɲ/ŋ, with ny/ng digraphs, with or without tone marks, with French spellings for loanwords. Any comparison between two Bambara texts (a transcription and its reference, two translations, a search query and a document) is meaningless until both sides are written the same way. This repository fixes one way, documents it, and ships the code.

The alphabet

a b c d e ɛ f g h i j k l m n ɲ ŋ o ɔ p r s t u w y z — 27 letters. ɛ (open e), ɔ (open o), ɲ (palatal nasal, as in French gn), ŋ (velar nasal, as in English sing). Tone marks (à, á, ǎ…) and vowel length (aa) are part of the official orthography but optional in everyday writing.

Normalization table (excerpt)

Seen in the wild	Standard	Rule
ny (nyɛ, nyama)	ɲ (ɲɛ, ɲama)	digraph → letter
ng at syllable onset (ngala)	ŋ (ŋala)	digraph → letter
è / e for open e (bè, cè)	ɛ (bɛ, cɛ)	accent or plain e → ɛ when the word is in the lexicon with ɛ
ò / o for open o (dòn)	ɔ (dɔn)	same, with ɔ
ka n'a / k'a	k'a	elision written with apostrophe
tone marks (kà, kǎ)	removed for comparison	tone-insensitive matching by default
Sikasso	Sikaso	Bambara spelling of place names

The full table (normalization.csv, 260 rows) covers digraphs, accents, apostrophes, punctuation, numerals and the most frequent loanwords.

Reference normalizer

from bambara_normalize import normalize
normalize("Ne bè taa Sikasso nyɛ fɛ")   # -> "ne bɛ taa sikaso ɲɛ fɛ"

bambara_normalize.py is ~120 lines, no dependencies, MIT-licensed. It is the normalizer to apply to both hypothesis and reference before computing WER or CER on Bambara; without it, error rates mostly measure spelling conventions rather than recognition.

What can and cannot be recovered

• The accent convention (è, ò) and the old digraphs (ny, ng, ty, dy) are recovered by rule, every time.
• Bare ASCII (e for ɛ, o for ɔ) is ambiguous: be may be bɛ but `ne* (I) is not nɛ. It is recovered only through the lexical table, which covers the most frequent words and place names, and never overwrites a form that is itself a standard word. On the reference sentences, half of the bare-ASCII rewrites are recovered in full; the rest keep one or two ASCII vowels. This is by design: a wrong guess is worse than a missing one.
• Tone marks are removed, so the comparison is tone-insensitive. Pass strip_tones=False to keep them.

Conventions used by Kooma

• Standard alphabet, no tone marks in output.
• French loanwords keep the French spelling when pronounced as French (téléphone), Bambara spelling when established (lɛri, mobili).
• Place and person names in Bambara spelling.
• Sentence punctuation as in French; no space before ? and !.

Files

• README.md — this document
• normalization.csv — the table
• bambara_normalize.py — the reference normalizer
• examples.jsonl — 220 before/after pairs used as tests
• test_normalize.py — the test suite (python test_normalize.py)
• LICENSE — MIT for the code; CC BY 4.0 for documentation and data

License and citation

CC BY 4.0 (documentation and data), MIT (code).

@misc{kooma-bambara-orthography,
  title  = {Bambara orthography and text normalization},
  author = {Kooma},
  year   = {2026},
  url    = {https://huggingface.co/datasets/kooma-ai/bambara-orthography}
}

---

Orthographe et normalisation du bambara (français)

Conventions d'écriture du bambara (bamanankan) en alphabet latin standard (27 lettres, dont ɛ, ɔ, ɲ, ŋ), table de normalisation des graphies courantes vers le standard, et normaliseur Python de référence. À appliquer aux deux côtés avant tout calcul de WER/CER : sans cela, on mesure des conventions d'orthographe, pas la reconnaissance. Maintenu par Kooma, https://kooma.ai