Why code-switching is hard for AI voice

What "code-switching" actually means here

Code-switching is mixing two languages naturally within a single sentence or conversation — not translating a whole paragraph into one language and the next paragraph into another, but genuinely blending them the way people actually talk. A real, everyday example: "Abeg, just send the file when you're less busy" — English and Nigerian Pidgin in one breath, with nobody stopping to think about it.

This is completely normal in real speech across Nigeria and much of Africa. It is almost never how AI voice systems are actually built.

Why most AI voice models can't do this at all

A standard text-to-speech model is trained on — and configured for — one language at a time. Feed it a sentence that switches languages mid-stream, and one of two things typically happens: it mispronounces the second language badly, using the phonetic rules of the first, or it simply fails outright. Neither is usable for a real narration a viewer is meant to actually understand.

This isn't a minor edge case for an Africa-first platform — it's close to the default way people actually write and speak. A tool that can't handle it isn't missing a feature; it's built for a different linguistic reality than the one it's being used in.

What ToonMate actually does — stated honestly

We want to be precise here rather than oversell it: ToonMate does not use a single model that natively understands mid-sentence language switching the way a specialized, proprietary system might. That's a genuinely different, harder technical achievement, and we don't claim to have built it.

What we actually built is a real, working system that gets to the same practical result a different way:

1. A script is marked up by real language segment — a section in English, a section in Yoruba, and so on, in whatever order the actual sentence needs.
2. Each segment is synthesized separately, using whichever real voice provider is actually best suited to that specific language — different providers are genuinely stronger at different Nigerian languages, so this isn't a single, one-size-fits-all call.
3. The resulting audio segments are stitched into one continuous track, so what a viewer actually hears is one seamless narration, not audibly separate clips bolted together.

The honest way to describe this: it's a real, working answer to the practical problem — a viewer hears natural code-switching — built through careful segmentation and provider selection, not through a single model that inherently "understands" multiple languages at once the way a human speaker does.

Why this distinction actually matters

We're naming this limitation directly because we think it matters more to say plainly than to blur it in marketing language. If you need a single model with genuine, native multilingual understanding for reasons beyond narration — say, a real-time conversational agent — this specific approach isn't that. If what you need is a finished video where the narration genuinely sounds like natural, mixed-language speech, this is a real, working answer to that problem today, not a future promise.