Why prototype with real voices
Voice copy that reads well on a page often sounds wrong out loud: too long, too formal, or with the important information buried in the middle. Synthetic test voices that sound robotic make it hard to judge the copy itself, because testers react to the voice. Hearing prompts in a natural voice, early, shows what actually needs fixing.
- Find long prompts: anything over a few seconds tests listeners' patience.
- Hear the persona: is it warm, neutral, brisk? Try two voices and compare.
- Test with users using Wizard-of-Oz style sessions, where a researcher plays the right file at the right moment.
- Hand over finished copy with audio references, so production voices match the intent.
How to set it up
- One project per product, and a script per flow: "Booking", "Errors", "Onboarding".
- Cast the persona: one voice, or two or three candidates to compare, from short samples recorded by people who've agreed.
- One prompt per line, in the order the flow reaches them. Number your prompt IDs in a sheet that matches the line order.
- Write numbers and times as words, the way the finished system would say them: "your table for four is booked for seven thirty". See numbers and abbreviations.
- Generate, listen, cut words, generate again.
- Download and load the files into your prototyping tool, a soundboard for user tests, or a slide deck.
Variable parts of a prompt
Real systems fill in names, numbers and dates. For a prototype, generate a few realistic versions of each templated prompt ("Your order of three items arrives Tuesday", "Your order of one item arrives tomorrow"), or split the prompt into a fixed part and a variable part on separate lines, the way many phone systems actually play them.
Reviewing with the team
Share the project with researchers, writers and stakeholders. With Can view they can listen to every prompt and leave comments; with Can edit they can rewrite and re-take prompts themselves. See collaborating on a project.
What it isn't
abcVoiceover makes audio files ahead of time from a script; it isn't a live speech engine your product can call while it runs. For prototypes and tests, pre-generated prompts are usually all you need, and for production the same copy, voice direction and audio references carry straight over.
Responsible voices
If the persona's voice is cloned from a real person, get their written consent, and make sure test participants know they're hearing a synthetic voice. See using AI voices responsibly.