Speech recognitionUsable commercially
Whisper Small + QLoRA
Transcribe Lingala speech (radio, service desk, consultation).
Models
Five models adapted to Lingala
Each starts from an open foundation model, adapted with the Maloba corpora rather than trained from scratch: the right choice for low-resource languages. Each page says what the model does well, what it does badly, under which licence, and links to the report that establishes it.
Speech recognitionNon-commercial
MMS 1B + adaptateur lingala
Transcribe Lingala speech, with output more readable at character level.
Document reading (OCR)Usable commercially
Qwen2-VL 2B + QLoRA
Read a photographed Lingala document, printed or handwritten.
French → Lingala translationNon-commercial
NLLB-Maloba-h1
Translate an instruction, an administrative document, a message.
Conversational assistantSubject to licence terms
Llama 3.1 8B + QLoRA
Answer, summarise and explain in Lingala.
Licences, in plain words
A model keeps the licence of the foundation model it comes from, whatever the licence of the data. Whisper and Qwen2-VL (Apache 2.0) may be used commercially. NLLB-Maloba-h1 and the MMS adapter (CC-BY-NC 4.0) are restricted to non-commercial use. The Llama model is subject to the Llama 3.1 Community License.