All news
ResearchAugust 15, 2026·5 min read·By DevBlock Research
Ẹtí: An Open-Source Yoruba Speech Recognition Model
Ẹtí ASR Model
Free, open-source speech-to-text for Yoruba, built to run smoothly on standard computers.
Key Takeaways
- Free and open-source under the friendly MIT license.
- Lightweight 248 MB model runs efficiently on everyday computers without dedicated GPUs.
- Open benchmarks and training details published for researchers and builders.
- Accurately transcribes clear speech, with ongoing work on fast everyday conversations.
Voice interfaces are only useful if they understand the languages people actually speak. For Yoruba—spoken by over 45 million people—most global voice assistants still struggle with tones, local accents, and mixed English-Yoruba phrasing.
We released Ẹtí ("ear" in Yoruba) as a free, open-source voice model so developers, researchers, and local startups can build voice apps without high cloud costs.
Honest Results and Future Work
On clear speech and news audio, Ẹtí achieves high accuracy. On fast-paced casual conversations, speech can still be challenging.
We openly share both our successes and our limitations because African language AI can only improve when researchers have real, honest baselines to build upon.