In development
K-TTS
A text-to-speech model built for three languages that general-purpose voices handle poorly together. Yoruba is tonal and diacritic-sensitive, Arabic needs correct vowelling and right-to-left handling, and most systems apply an English cadence to both. K-TTS is trained to avoid that.
- Status
- In development
- Availability
- Not yet released
- API
- In development
What makes three languages hard
Each of these is handled poorly by general-purpose voices, and handled worse when a sentence moves between them.
Tone-aware Yoruba
Tone marks change meaning, so they are trained as signal rather than stripped as noise.
Diacritic-sensitive Arabic
Correct vowelling, and right-to-left text handled without mangling the output.
Code-switched input
A sentence that moves between languages mid-way, spoken without an English accent applied to all of it.
Per-language evaluation
Scored separately in each language, because one aggregate number hides the weakest of the three.
Where this stands today
K-TTS is in active training and is not available yet. We are working from native-speaker recordings across all three languages rather than adapting an English model and hoping the prosody transfers. There is no public API and no waitlist theatre. If this is something you need, tell us what for, and we will come back to you when there is a build worth your time.
Tell us what you need it for
We would rather hear the use case early than guess at it. It shapes what we train for.