Producing twenty-nine voiced locales offline

An app for children under six is a voice-first product: the audience cannot read, so every prompt, number and reward has to be spoken. Localising it the usual way means a text-to-speech vendor, a per-character bill that scales with every content change, and a dependency that has to stay funded for the lifetime of the app.

We produced the entire package offline instead — two speakers across twenty-nine locales, plus a separate, gentler set of takes for two- and three-year-olds on the counting and colour lines, fifteen music tracks and eight sound effects. Every line has a manifest entry recording its text, file, the model that generated it and its QA result, and every music and sound file carries a per-file licence record.

What the automation could not do is judge whether a line sounds right to a child in a language nobody on the team speaks. That part stayed human, and the package ships with a listening checklist that says so explicitly rather than implying the QA was complete.