AI models for African languages
Does it actually work for Wolof, Bambara, Lingala or Darija? Benchmarks rarely say. The community tests these models on real speech and text, and reports back.
dyula-french-nmt-150
JuliusFx
Dyula to French translation (JoeyNMT).
nllb-600M-moore-lora-v0
Rekin226 / Mooré-Voice
Mooré-French/English translation (~205k pairs).
whisper-tunisian-dialect
TuniSpeech-AI
Tunisian Arabic ASR (~21h LoRA).
Darija ASR (57h, WER 8.5%).
Terjman-Ultra-v2.0
AtlasIA
English to Darija translation (NLLB-1.3B tune).
Al-Atlas-0.5B
AtlasIA (Morocco, community)
First Darija foundation model.
Atlas-Chat-9B
MBZUAI-Paris
Moroccan Darija chat model (Gemma-2-9B tune).
DarijaBERT
SI2M Lab INSEA + AIOX Labs
BERT model for Moroccan Darija (~100M tokens).
wav2vec2-xls-r-300m-wolof-lm
abdouaziiz
Wolof ASR trained on ALFFA (WER 21.3%).
wolofToFrenchTranslator_nllb
GalsenAI (Senegal)
Wolof-French translator (BLEU 13).
AfroLM
B. Dossou et al.
Masked language model covering 23 languages.
AfriBERTa-large
Castorini (Ogueji et al.)
Masked language model; Swahili plus 10 other languages.
AfroXLMR-large-76L
Davlan (D. Adelani et al.)
Masked language model covering Wolof, Bambara, Lingala, Swahili, French and Arabic variants.
MasakhaNER (afroxlmr-large-ner)
Masakhane
Named-entity recognition covering 21 African languages.
Google Cloud Translation
Lingala, Bambara and Swahili via API; Wolof and Fon in the consumer app since June 2024.
Whisper large-v3
OpenAI
Supports Swahili, Lingala and Arabic (no Wolof/Bambara).
NLLB-200
Meta AI