Congolese Swahili · WHO DRC Ebola interviews · 15 October 2026

Speech recognition and translation for Congolese Swahili

How well current models transcribe Congolese Swahili speech and translate it into French, measured on public corpora. Every page here is built from public data only.

Speech recognition · summary

Do Gemma 4 or Whisper beat Omni?

Five ASR arms on 2,016 clips of public Congolese Swahili read speech. Omnilingual ASR 7B makes the fewest word errors.

Open →

Speech recognition · technical appendix

Method, error analysis and caveats

Setup, prompts, normalisation, paired bootstrap tests and where each model goes wrong.

Open →

Translation · dashboard

Which models translate Congolese Swahili into French best?

44 systems on 2,000 held-out sentences: frontier APIs, open-weight LLMs, dedicated MT and a LoRA fine-tune, with cost, prompting techniques and slices.

Open →