Maloba in LoBAI
The Maloba models are presented in conversation in LoBAI, the Congo Digital Services interface, and through an OpenAI-compatible API. The access point is being commissioned.
Local Language Accelerator programme · UNDP Congo
Lingala speaks to machines
With its speakers, Maloba is building the data and the models Lingala was missing: recognising speech, reading documents, translating and answering. Everything is published, measured and open to the community.
Delivered by Congo Digital Services for the WYMEE / CDS / SACLI consortium, under the authority of the Ministry of Posts, Telecommunications and the Digital Economy.
58h
of annotated and published speech
58,13 h published after segmentation and deduplication, from 210 broadcasts
Corpus V3.2 dataset sheet §6.218 767
transcribed speech segments
15 175 for training, 1 795 for validation, 1 797 for test
Corpus V3.2 dataset sheet5
models adapted to Lingala
Speech (Whisper, MMS), documents (Qwen2-VL), translation (NLLB), assistant (Llama)
Deliverable 4 §1310
samples judged by three Lingala-speaking validators
930 judgements; meaning preserved in 96,5 % of cases
Deliverable 3 §6.788,7%
correct characters in document reading
92 % of outputs judged recoverable by human validators
Report E01Figures as of September 30, 2026. Every value links to the report that establishes it.
The data
Four reference corpora: speech, documents, text, translation. Identified sources, harmonised spelling, original text preserved.
See the corporaThe models
Five models adapted to Lingala from open foundations. Each page gives the measured result, the licence in plain words, and the limits.
See the modelsThe results
Automatic measurements that can be replayed, review by three Lingala-speaking validators, and the fifteen evaluation and training reports.
See the resultsTake part
Lend your voice, annotate, review, translate, open a site. The corpus is built line by line, with its speakers.
Join the workGovernance
Who may use the corpora, who should benefit from them, how consent is obtained, and how to have a recording withdrawn.
Read the governanceResources
Lingala spelling conventions, the harmonisation module, the evaluation pipeline, public repositories, reports and press kit.
See the resourcesWhat works, and what remains hard
We publish both. That candour is what makes the rest credible.
What works
Document reading reaches 88,7 % correct characters, confirmed by human validators, and is the only result whose inter-validator agreement was judged sufficient. Translation rises from 23,37 to 32,00 BLEU, a significant gain.
What remains hard
Speech recognition still stands at 44,17 % word error on field speech. The amount of audio (58 h against the 80–100 h expected) and unsettled spelling account for part of the gap.
Who owns the voice of a community
The Maloba corpora belong first to the community that produced them. They are published under the Nwulite Obodo Open Data 1.0 licence, which does not grant the same rights to an underfunded African laboratory and to a technology multinational.
African researchers, developers and organisations
Free and open access, with attribution to the Maloba project.
No royaltyUsers in other developing countries
Free and open access, with attribution, for any use.
No royaltyCompanies and high-income countries
Benefit sharing to be agreed with CDS before use; the proceeds return to the project and to the Lingala-speaking community.
Benefit sharingLatest publications
What the team has done, measured and published.
Three validators, 310 samples
What human validation says about our data: meaning is preserved in 96.5 % of cases, but validators agree little on what counts as “correct” or “partial”.
MMS learns to write ɔ and ɛ
Trained in 1 h 50 on 0.23 % of its parameters, a language adapter for MMS reaches the same word accuracy as Whisper, with lighter errors at character level.
Who does this work
UNDP, Local Language Accelerator programme
Funding, steering, technical review
Funding partnerMinistry of Posts, Telecommunications and the Digital Economy
Oversight of the project
OversightSIL Congo
Document collection: 22 digital documents (1 008 pages) and 110 manuscripts
Data partnerRadio Rurale · ORPR
54 h of digital archives; about 150 h of tape still to digitise
Signed agreementThe community of speakers
Health Voice-a-thon, at the Grandes Endémies centre
ContributorsWYMEE, CDS, SACLI
Consortium holding the contract, delivered by Congo Digital Services
Consortium