SAUTI ASR v1 is now open: Swahili speech recognition for everyone
Off-the-shelf ASR gets Swahili wrong one word in four. SAUTI ASR v1 cuts that in half. The model is open on HuggingFace with a live demo, Python and TypeScript SDKs, and API access.
Today we are making SAUTI ASR v1 available to everyone.
The model weights are published on HuggingFace under an open license. You can download them, run inference locally, fine-tune further, or integrate into your own applications.
What we are releasing
- **Model weights:** [Finiflowlabs/sauti-asr-v1](https://huggingface.co/Finiflowlabs/sauti-asr-v1) — fine-tuned Swahili speech recognition, ready to use
- **Live demo:** Try it in the [Speech to Text playground](/speech-to-text) or on [HuggingFace Spaces](https://huggingface.co/spaces/Finiflowlabs/sauti-demo)
- **API access:** Authenticated endpoints with rate limiting and observability
- **Client SDKs:** Python and TypeScript packages with sync and async interfaces
Why open-source?
African language AI has a distribution problem, not a technology problem. The best way to accelerate progress is to make our work freely available so researchers, developers, and companies across East Africa can build on it.
Open-sourcing means: - Researchers can reproduce and extend our results - Developers can integrate Swahili ASR into their products today - Other teams working on African languages can adapt our approach - The community can audit our work and help us improve
The result
| System | Word Error Rate |
|---|---|
| Off-the-shelf multilingual ASR | 27.2% |
| SAUTI ASR v1 | 13.5% |
A 50% reduction — moving Swahili speech recognition from unreliable to production-viable.
What is next
This release is the foundation. Here is what we are working on:
- **Voice Agent:** Full conversational voice AI — speak Swahili, get a response in Swahili
- **Real-time translation:** Live English–Kiswahili translation through earphones
- **More languages:** Expanding beyond Swahili to other African languages
If you are building with African language voice AI — as a researcher, developer, or product team — we want to hear from you. Try the demo, use the API, and tell us what you need.