How we are building an African music model that sings like home, and where Kivio goes next.
Our vocal filter was quietly binning nearly 1,500 real vocal tracks as instrumentals, because the old transcriber could not hear African languages. Whisper large-v3 recovered them. Here is what that taught us about building for Naija sound.
Every time you pick the better take on our rate page, GWEN moves a little closer to real African taste. We explain the preference-tuning loop, and why an African panel is a moat global tools cannot copy.
Yoruba, Igbo and Hausa are tonal. Most AI singers pronounce them like tourists. We dig into why authentic African pronunciation is the hardest and most valuable thing to get right.
GWEN is built only on music we own or licensed, with AI-training consent. In a year of AI-music lawsuits, clean provenance is not just ethics. It is the product.
Occasional notes on African AI music. No spam.