Sauti TTS

We built a Swahili text-to-speech system. It's the best one that exists.

I'll explain what it does, why we built it, and where it's going.

200 million people speak Swahili. Almost none of the speech technology built in the last decade works well for them. The existing Swahili TTS systems are either twenty years old, robotic-sounding, or not publicly available. This felt like an important problem.

We fine-tuned F5-TTS - a flow-matching architecture - on Google's WaxalNLP dataset, which has 180+ hours of studio-quality Swahili recordings. The result is natural-sounding Swahili speech with correct prosody and rhythm. You can give it a 5-second voice sample, and it clones that voice. We call it Sauti TTS. Sauti means voice.

The most interesting part of this project, technically, was not the model. It was the data.

Neural TTS models are extremely faithful to their training data. If your audio has tiny clicks and pops from mic handling, digital glitches, or sampling rate conversion, the model learns to reproduce them. Most of these artifacts are inaudible to humans. machines hear everything.

So we built a pipeline to fix this. It detects clicks using derivative analysis of the audio signal, repairs them through interpolation, replaces the standard resampler (which introduces its own artifacts) with a higher-quality one, and gates every sample on a quality score before it enters training. This is not swahili-specific. It works for any language.

I think data quality is the most underrated problem in speech AI right now. Everyone talks about model architecture. Almost nobody talks about whether their training audio has clicks in it.

We also built a full Swahili text normalizer. When you write "Ksh 5,000," a Swahili speaker says "shilingi elfu tano." When you write "Dkt. Amina," they say "Daktari Amina." A TTS system that can't do this will sound broken every time it encounters a number or an abbreviation, which is constantly.

This sounds trivial. It is not. Getting it right is one of those things that's invisible when it works and devastating when it doesn't.

Sauti TTS is one half of what we're building. The other half is Sauti ASR - our speech recognition system, which achieves a word error rate of 3.24% in Swahili. That's the best published result for the language.

Having both a TTS and ASR system for the same language turns out to be more valuable than the sum of the parts. The ASR system's errors tell you exactly which sounds it confuses. The TTS system can then synthesize targeted training audio for those exact confusions. retrain the ASR. repeat. We call this Sauti-Loop. A paper is coming.

Compounding loops like this are rare and powerful. They're how you get from good to great on a problem.

We're also building AkiliX, a multimodal language model for seven East African languages. Sauti is the voice layer. Akilix is the brain. They will work together.

Sauti TTS is open-weight. MIT license on the code. We built it, no large compute cluster. No corporate budget. just a clear problem and the engineering to solve it.

I believe the best AI for Africa will be built in Africa. not because of some abstract principle, but because the people closest to the problem understand it best. Swahili prosody, Swahili morphology, and Swahili text conventions are things you have to live to get right.

We're releasing everything. If you're building voice technology for Swahili or any East African language, we want to work with you.

— Kiplangat Korir, MsingiAI