This dataset contains recordings of human voices and is licensed under NOODL-1.0. Access is reviewed by Congo Digital Services (CDS SARL). The information below is used only to apply the licence tiers and to contact you about benefit sharing where the licence requires it.

audios-lingala-annotatees-v3.2

MALOBA Lingala speech corpus: 18,767 annotated audio segments (58.1 hours), with the original transcriptions and a harmonised version in the MALOBA spelling convention.

MALOBA (digitisation and acceleration of Lingala) is led by Congo Digital Services (CDS SARL) in partnership with UNDP Republic of Congo and the Ministry of Posts, Telecommunications and the Digital Economy.

Licence: Nwulite Obodo Open Data License 1.0 (NOODL-1.0). Access is gated. See Licence and Speakers and privacy.


Language	Lingala (`ln`), with frequent French code-switching
Task	Automatic speech recognition
Audio	FLAC, 16 kHz, mono; segments of 0.3 to 25 s
Size	18,767 segments, 58.1 h
Splits	train / validation / test, grouped by audio fingerprint
Version	v3.2, supersedes v3.1
Licence	NOODL-1.0

Counts

Split	Segments	Hours	Source recordings
train	15,175	47.93	172
validation	1,795	5.37	19
test	1,797	4.83	19
Total	18,767	58.13

Training segment durations: minimum 0.30 s, median 9.09 s, 95th percentile 24.96 s, maximum 25.00 s (the segmentation caps segments at 25 s).

How the corpus was built

Collection and annotation. The recordings are programmes broadcast by Radio Rurale of the Republic of Congo, the sole owner of the recordings and of the rights on them. Radio Rurale provided them to CDS and gave its consent to their use in the MALOBA project and to their distribution in this dataset. The programmes cover everyday topics such as family, trade, travel, religion and health. CDS segmented the recordings and annotators transcribed them in Label Studio.
Reconstruction (V2). The corpus was rebuilt from the annotation bucket by the maloba-asr-rebuild service: audio re-extracted from the source recordings and stored losslessly (FLAC, 16 kHz, mono), with a PCM fingerprint and the source recording id for every segment.
Deduplication (V3). Identical audio ingested several times under different ids was deduplicated on the PCM fingerprint. The splits were redrawn on that fingerprint: no audio is shared between train, validation and test.
Harmonisation (v3.1, then v3.2). The transcriptions were harmonised with the same deterministic pipeline as the MALOBA translation, LLM and OCR datasets. The original transcription is kept unchanged next to the harmonised one.

What changed in v3.2. The Unicode look-alike ͻ (U+037B) is now mapped to ɔ (115 occurrences). Method, columns and evaluation rule are unchanged from v3.1.

Columns

Column	Content
`audio`	the segment audio (FLAC bytes, identical to V3: never decoded or re-encoded)
`text`	V3 transcription, unchanged
`transcription_v3`	copy of `text`, for explicit use in evaluation
`transcription_harmonisee`	transcription in the MALOBA spelling convention
`nb_modifications`, `modifications`	number and JSON list of replacements (`etape`, `avant`, `apres`)
`empreinte_pcm`	SHA-256 fingerprint of the decoded audio samples
`id_enregistrement`	source recording id
`duree_s`	segment duration in seconds
`vu_train_v2`, `vu_validation_v2`, `dans_test_v2`	the audio fingerprint appears in the V2 train, validation or test split
`eval_propre`	not in V2 train or validation: never seen by the fine-tuned Whisper model

The other V3 columns are kept as they are.

Harmonisation

The harmonisation is automatic and deterministic. No transcription is rewritten by a person or by a generative model.

Unicode. Look-alike characters are fixed (ε→ɛ, ↄ and ᴐ→ɔ, ͻ→ɔ, ꜫ→ɛ, Greek ο→o), then the text is normalised to NFC.
Tone marks are removed.
Spelling-guide rules are applied.
ɔ/ɛ spelling. Each word takes its majority spelling in the MALOBA corpora, from the frozen lexicon outils/lexique_majoritaire_v4.csv.
- The lexicon has 11,382 entries.
- The 6,322 entries added for this corpus were voted on the train and validation splits only, never on the test split.

French words from a protected vocabulary (outils/vocabulaire_francais.txt) and non-lexical tags ([…], (…), bruit) are left untouched.

Results:

10,528 transcriptions changed.
Words written in more than one way: 600 → 290.
Effect of the harmonisation alone on the test references: WER 5.27% raw and 4.86% normalised; CER 1.14% raw and 1.03% normalised. This is the share of error that comes from spelling variation alone.

The harmonisation module is outils/harmonisation_maloba.py, and it has no dependency on the rest of the MALOBA code. Applied to transcription_v3, it reproduces transcription_harmonisee exactly on all 18,767 rows.

Evaluation rule

To compare a model with the MALOBA results:

Keep split == "test" and eval_propre == True.
Harmonise the model outputs with the same code before scoring:

from harmonisation_maloba import harmoniser   # outils/harmonisation_maloba.py
prediction = harmoniser(raw_prediction)

Reading	Reference	Prediction
A	`transcription_v3`	raw output
B	`transcription_harmonisee`	raw output
C (reference reading)	`transcription_harmonisee`	`harmoniser(raw output)`

Report the WER and CER, raw and normalised. The test split contains only 19 source recordings. Segments of one recording are not independent, so compute confidence intervals by bootstrap over recordings, not over segments.

Usage

from datasets import load_dataset

ds = load_dataset("Congo-digital-service/audios-lingala-annotatees-v3.2")
test = ds["test"].filter(lambda x: x["eval_propre"])
print(test[0]["transcription_harmonisee"])

Access is gated: accept the conditions on this page, then log in with huggingface-cli login.

Speakers and privacy

The recordings contain human voices. Voice recordings are personal data.

The recordings are radio broadcasts. The people heard in them (presenters, guests, callers) spoke on air; CDS did not contact them individually.
Radio Rurale of the Republic of Congo, the sole owner of the recordings, consented to their use and distribution in this dataset.
The dataset contains no speaker names or contact details. Annotator identities are not included.
Users must not try to identify speakers, and must not use the recordings for voice cloning or speaker identification. This is a condition of access.
Anyone who recognises their own voice can ask for the recording to be removed by contacting CDS (below). Removed recordings are left out of the next version.

Limitations

Accents removed from French words. The harmonisation removed é from 4,209 occurrences of French words (613 forms, for example émission → emission, difficulté → difficulte), while è, à and ç are kept. The protected French vocabulary was too small. This has no effect on evaluation in reading C, where both sides are harmonised, but a model trained on transcription_harmonisee will write these words without é.
Leftover characters. A few characters have no rule and are left as transcribed: ∫ (33 occurrences, probably a mistyped ʃ), ‘ (30), ɲ (25), ӡ (7) and some single occurrences.
Radio speech only. All recordings come from broadcasts: presenters and studio speech are over-represented compared with everyday conversation. The number of distinct speakers has not been established.
Few source recordings. 172 in train, 19 in validation, 19 in test. Results on the test split have a small effective sample size.
Spelling convention. The harmonised convention is consistent across the MALOBA datasets. It has not been validated word by word by linguists.
Code-switching. Transcriptions mix Lingala and French, as in everyday speech in Brazzaville and Kinshasa.

Versions

Version	Change
v2	Corpus rebuilt from the annotation bucket; audio stored in FLAC 16 kHz mono with fingerprints. The fine-tuned Whisper model was trained on this version.
v3	Deduplicated on the audio fingerprint; splits redrawn on the fingerprint.
v3.1	Harmonised transcriptions; V2 overlap columns; harmonisation module.
v3.2	Look-alike ͻ (U+037B) mapped to ɔ.

The V3 splits are not the V2 splits. A V3 training segment may come from the V2 test split. The overlap columns show where each segment was in V2.

Licence

This dataset is licensed under the Nwulite Obodo Open Data License 1.0 (NOODL-1.0). The licence text is authoritative; the summary below is for convenience only.

Users in Africa and in developing countries: royalty-free use, modification and sharing.
Users from high-income countries, and commercial entities: use is subject to benefit or value sharing with the dataset providers, agreed before use. Contact CDS (below).
Everyone: keep the attribution below, including when sharing adapted versions.

Attribution:

MALOBA project: Lingala speech data recorded and annotated by Congo Digital Services (CDS SARL, https://www.congo-digital.com/) and the MALOBA community of annotators, with UNDP Republic of Congo (language digitalisation initiative). Licensed under NOODL-1.0.

Scope notes:

Models trained on this dataset keep the licence of their base model. For example, a model fine-tuned from MMS remains CC-BY-NC 4.0 and cannot be used commercially, whatever the user's tier under NOODL.
The speaker conditions in Speakers and privacy apply in addition to the licence.

Citation

@misc{cds2026malobaspeech,
  title  = {audios-lingala-annotatees-v3.2: annotated Lingala speech from the MALOBA project},
  author = {{Congo Digital Services}},
  year   = {2026},
  note   = {MALOBA project, UNDP Republic of Congo. Version 3.2. Licensed under NOODL-1.0.}
}

Contact

Congo Digital Services (CDS SARL), Brazzaville, for the MALOBA project: contact@congo-digital.com. Card updated on 30 September 2026.

Downloads last month: 21

Total file size:

4.78 GB

Models trained or fine-tuned on Congo-digital-service/audios-lingala-annotatees-v3.2