1. Guided facets
Start with low-cognitive-load choices: gender, age group, and broad style. This is closest to your original idea and works well when users already have a voice profile in mind.
Mudo voice-picking UX
These are interaction patterns for the station builder. Each one uses the same Inworld voice catalog and preview audio, but makes the decision smaller: filter, match to station role, compare one at a time, or tune high-level traits.