How does AI sound in Kikuyu?

November 12, 2026 · Allan Karanja

After working on speech-to-text for Kikuyu, I wanted to try the other direction. I fine-tuned Meta’s text-to-speech model to generate Kikuyu speech. At 159 MB, it’s small enough to run on a CPU.

This is the future of AI: AI in your local language, running locally.

Here’s a Kikuyu proverb generated in three voices:

Ũgĩ wa mũndũ ũmwe ndũrĩmaga.
One person’s ability cannot till all fields.

The model performs well on short sentences but might struggle on longer sentences that weren’t seen in the original dataset.

Model: Model on Hugging Face

Dataset: Filtered dataset