Research
Speech recognition and language models for low-resource languages, starting with Kinyarwanda, still in progress.
The programme includes a text pipeline, a 183-million-token corpus, and continued pretraining and supervised fine-tuning of open language models. It also includes KinyaBench, a public-track evaluation harness for Kinyarwanda slices of existing multilingual benchmarks, and a pre-registered ASR experiment programme on formal and casual speech. I am now converting about 200,000 hours of raw audio, mostly Kinyarwanda with some Kirundi, into speech recognition training data.
There is no public model or paper yet.
Word error on formal and casual Kinyarwanda
Open and closed recognizers on formal parliamentary speech and casual conversation, including a four-system vote.
KinyaBench: open language models on Kinyarwanda
Preliminary chat, 0-shot, single-prompt scores for AfriqueQwen3.5 4B and 9B on the Kinyarwanda slice of public benchmarks.
Building a Kinyarwanda corpus and training stack
The February–April 2026 programme that built a Kinyarwanda text pack, measured tokenizer cost, trained models from the pack, and distilled a 32,433-row supervision set.