MMS 1B is Meta's massively multilingual speech-recognition model. It already knows Lingala, but its alphabet holds neither ɔ nor ɛ, two letters of the Maloba writing convention. We trained a language adapter, 0.23 % of the model's parameters, on the 47.79 hours of the V3.2 training split, extending the alphabet to those two letters.
The same WER as Whisper, a lower CER
On the leak-free V3.2 test set, the MMS adapter reaches 44.12 % word error against 44.17 % for fine-tuned Whisper. The gap is not established. At character level, however, MMS does better: 16.26 % against 18.95 %. Its errors are lighter and its output more readable.
A specialisation, not a regression
On FLEURS, a set of read Lingala speech, the original MMS stays better (14.66 % against 22.48 %). The adapter specialised on the project's field speech, radio and service desk, at the expense of read speech. We publish both figures.
The adapter inherits the CC-BY-NC 4.0 licence of MMS. It is published on Hugging Face.
By The Maloba team, September 29, 2026.