You can now describe a voice in one sentence and get a few AI-generated versions of it back. Something like "calm, mid-30s, Irish accent, front desk at a dental clinic." You don't need a recording or a voice actor. That makes trying voices cheap. Picking the one that holds up on real calls is harder, and that's what our new guide covers: 1. Write the prompt like a casting brief. Cover age, accent, pace and energy, and end with what the voice is for. A receptionist and a collections agent shouldn't sound alike. 2. Adjectives don't translate. In @GradiumAI's rating data, US listeners kept 76% of the voices made from a "warm" prompt. French listeners kept 53%. Write a separate prompt for each language and use case, with a native speaker. 3. Don't pick by gut. Play listeners two voices at a time without saying which is which, and count which one wins more often. 4. Test it on real calls. A voice that sounds great in a demo can still stumble over a card number, an email address, a caller who cuts in, or a bad phone line. Run the same test calls through each finalist and change only the voice. 5. Test again whenever you change the voice model, its settings or the voice ID. The guide closes with a seven-step checklist to run before launch: cekura.ai/blogs/voice-design…