AI models for African languages

Dyula to French translation (JoeyNMT).

Mooré ASR (~23k pairs).

Mooré-French/English translation (~205k pairs).

Fon ASR (~15% WER).

Tunisian Arabic ASR (~21h LoRA).

First Tunisian Arabic BERT.

Darija ASR (57h, WER 8.5%).

English to Darija translation (NLLB-1.3B tune).

First Darija foundation model.

Moroccan Darija chat model (Gemma-2-9B tune).

BERT model for Moroccan Darija (~100M tokens).

Bambara TTS with 10 voices.

Bambara ASR (~40h).

Wolof TTS (~40k samples).

Wolof ASR trained on ALFFA (WER 21.3%).

Wolof-French translator (BLEU 13).

First open Wolof LLM (8B).

Masked language model covering 23 languages.

Masked language model; Swahili plus 10 other languages.

Masked language model covering Wolof, Bambara, Lingala, Swahili, French and Arabic variants.

Named-entity recognition covering 21 African languages.

Lingala, Bambara and Swahili via API; Wolof and Fon in the consumer app since June 2024.

Supports Swahili, Lingala and Arabic (no Wolof/Bambara).