The process
Nothing about this is a black box. Here's exactly what happens between someone recording a phrase and it becoming part of the Fiji Baat voice.
Every contributor joins with a code — either one shared by someone already involved, or one requested directly. This keeps the project invite-based rather than wide open, so quality and consent stay easy to manage as it grows.
Before recording anything, every contributor confirms they're okay with their voice being used to train the model. No recording happens without that confirmation.
A simple hold-to-record button, on any phone or computer. Each phrase is short — numbers, common words, everyday sentences — designed to cover the sounds and rhythms of Fiji Baat. Contributors can also upload an existing recording instead, and can re-record anything they're not happy with, any time.
Contributors aren't limited to a fixed list — anyone can type in and record a phrase they think is missing from the set. Every suggestion is reviewed before it's added, so the phrase list keeps growing with real community input.
Nothing goes straight into the training set. Each recording is checked by hand before it's approved — this is what keeps the eventual voice sounding natural and accurate, rather than being trained on rushed or unclear takes.
Once there's enough clean, approved audio, it's used to train a text-to-speech model — the actual "voice" of Fiji Baat. This is done locally, not outsourced to a third-party cloud AI service, which matters for keeping the community's voices under the project's own control.
The trained voice powers free tools — text-to-speech, translation, dubbing — and can also support paid work through Nighthawk Productions and licensing partnerships. See For Business for that side of things, or For Contributors for what it means for the people who record.
