Maloba

Data

The corpora, and who they belong to

The Maloba corpora are built from identified sources (rural radio stations, Voice-a-thons, the SIL Congo collection), transcribed by trained annotators, then harmonised by a deterministic program: no sentence is rewritten by a machine. The original text is always kept beside the harmonised one.

Licence and governance

Four reference corpora

The same template for each: content, construction, figures, access, licence, limits, citation.

Spelling: what harmonisation does

Lingala is written in several ways. In the corpora, the same words appeared under 600 competing spellings. Harmonisation, automatic and deterministic, brings them down to 290, without ever rewriting a sentence:

  1. Unicode look-alikes are corrected (ε → ɛ, ↄ and ᴐ → ɔ, Greek ο → o), then the text is normalised.
  2. Tone marks are removed.
  3. The rules of the writing guide are applied.
  4. Each word takes the majority spelling of ɔ and ɛ across the Maloba corpora, from a frozen lexicon.
  5. French words in a protected vocabulary and non-lexical tags are left untouched.

Spelling variation on its own accounts for about 5 % of word error rate on the test references. The module is published with the corpus and reproduces the harmonised column exactly. See the guide under Resources.

Versions of the speech corpus

  1. V2Rebuild from the annotation service: audio re-extracted losslessly, fingerprint per segment. The version Whisper was trained on.
  2. V3Deduplication on the audio fingerprint; splits drawn so that no audio is shared between train, validation and test.
  3. V3.1Transcriptions harmonised under the Maloba writing convention, alongside the original text.
  4. V3.2Reference versionUnicode look-alike ͻ (U+037B) mapped back to ɔ (115 occurrences). Reference version.

Corpus V3.2 speech figures

  • 18 767 segments, 58,13 h, 210 source recordings.
  • Training 15 175 · validation 1 795 · test 1 797.
  • Source collection: the Radio Rurale digital archives. About 150 h of magnetic tape are not digitised.

The document-reading chain, in figures

1 747 text regions read across 117 annotation tasks, of which 1 323 unique validated transcriptions and 1 282 harmonised lines. The published training set keeps 417 of them, and 50 for the test.

Limits at closure

The corpus is frozen as of 30 September 2026. What follows is observed, quantified, and calls for no follow-up.

Documenté

What is measured, and what is not

  • 58,1 h annotated and verified, below the 80–100 h range of the terms of reference.
  • The number of distinct speakers in corpus V3.2 is not established.
  • Agreement between human validators stays below the 0,60 threshold set before the campaigns: the rates are published as indicative.
  • The Impfondo and Talangaï collections were not carried out.
Documenté

Resources gathered, outside the annotated corpus

  • Health Voice-a-thon: 11 recordings from 10 people. No duration of speech is published.
  • About 150 h of Radio Rurale magnetic tape, not digitised.
  • SIL Congo written corpus: 22 documents, 1 008 pages. They were used for document reading; they are not released as running text.

Release: three datasets are submitted, pending validation, on Mozilla Data Collective.

Earlier versions

Kept so the first training runs remain reproducible. Prefer the reference versions.