In this article
Prepare an AI voiceover by making the intended reading explicit before producing the full narration. Keep the audience-facing spelling separate from the text sent to the speech tool, especially for names, dates and abbreviations. Clear decisions and a short listening review are more useful than repeatedly changing the whole script without knowing what went wrong.
Keep display copy and spoken copy separate
Your on-screen copy may correctly use digits, a brand's capitalization or a compact abbreviation. The spoken version may need those items written differently to make the intended delivery clear. Preserve both versions and link them to the same approved meaning. Do not silently change a name or factual value just to obtain a smoother reading.
ElevenLabs recommends spelling out numbers, dates, symbols and acronyms in the way they should be delivered. Treat that as preparation guidance, not a guarantee that every voice or model will produce the same result.
Resolve dates, numbers and units before generation
Consider this fictional display line: “The event is on 05/06 at 09:30 in Room 12.” Before narration, confirm the date convention. If the intended date is June 5, an approved English spoken version could be: “The event is on the fifth of June at nine thirty in room twelve.” The example is not an actual event announcement.
Do the same for prices, measurements, telephone numbers and product codes. Should a number be read as a quantity or as individual digits? Should an abbreviation be a word or separate letters? Record the decision instead of expecting the engine to infer it. Keep the underlying value unchanged and review each language's own conventions.
Build a small approved pronunciation sheet
For each sensitive term, record the display form, language, intended reading and who confirmed it. Ask the name owner or project owner when the correct pronunciation is unknown. A rough phonetic spelling can be a working note, but do not publish it as the official spelling of a person's or product's name.