Speaking Kiswahili Back: Swahili Text-to-Speech

After we opened Swahili speech → text, the other half of the pair was clear: machines that can speak Kiswahili back, not only hear it. Millions of people already write and read Swahili. Too often a spoken interface still arrives in another language, or not at all.

We are opening Swahili text → Swahili speech. The public checkpoint is nileagi/nileagi-sw-tts. Write a clear line of Kiswahili below and listen. For Kisukuma, use Sukuma text-to-speech.

What to expect

It is a narration voice for Kiswahili written in the usual Latin spelling. Output is 24 kHz speech. Digits and amounts are spoken as Swahili words (so 2500 is heard as elfu mbili na mia tano). A longer paragraph is spoken in pieces and joined, so a full message can still come back as one recording.

It suits clear Swahili lines and short things you might hear at a counter or in a notice. Have a person listen before anything goes out as a public recording.

Applications

The useful case is Swahili that should be heard, on a device people already have, without sending the text through another language first.

• Spoken notices and service lines

  Read greetings, balances, and short instructions aloud so a listener does not have to stare at a screen.

• Learning and accessibility

  Let students and readers hear Kiswahili they just wrote, and give people who prefer listening a way through the same text.

• A Swahili speech pair

  Speak text with this voice, or write speech down with the companion recogniser, without leaving Kiswahili.

For builders

Load the model from Hugging Face with Transformers. The live demo keeps requests modest, about 2,000 characters. Sample audio ships in the Hub repository and in the try box above.

Collaboration

If you are evaluating on new Swahili text, adapting this voice for your setting, or exploring what comes after synthesis, write to hi@nileagi.com. The weights are released under CC BY-NC 4.0; commercial use needs a written agreement with NileAGI. Use this system for clear Swahili research synthesis, keep legal, medical, or emergency use out of scope for now, and do not treat Sukuma or English as input.