Commercial-safe models
What your contributions are building
We fine-tune only permissively-licensed bases — Whisper (ASR), Llama/Aya (translation), Piper (TTS) — so these models can ship in real African products. Lower WER/CER is better; higher F1/MOS is better.
Swahili Kiswahili
Bantu · ~80M speakers
1
recordings
0
validated
2
MOS ratings
Amharic አማርኛ
Semitic · ~57M speakers
0
recordings
0
validated
0
MOS ratings
Benchmark runs
| Task | Lang | Model | WER | CER | Intent F1 | MOS | n | Date |
|---|---|---|---|---|---|---|---|---|
| ASR | AM | whisper-small (baseline, zero-shot) | 1.25 | 1.57 | — | — | 20 | 2026-06-14 |
| ASR | AM | whisper-small (Dimtse FT, target) | 0.38 | 0.21 | — | — | 20 | target |
| TTS | AM | piper-am (contributor voices, MIT) | — | — | — | 3.9 | 40 | target |
| NLU | AM | xlm-roberta-base + heads | — | — | 0.91 | — | 120 | 2026-06-14 |
Baseline = zero-shot before fine-tuning; "target" rows are the goals once contributor + Dewul data lands on the sponsor GPU.