How Synthetic Voices Read a Script

A human narrator who sees "St. Mary's St." on the page pauses, frowns and says it right. A synthetic voice says something, instantly and with total confidence. The script is the only control panel. It has to work for the ear, and for a reader that has never been to your town and cannot ask.

Spell out what normalization gets wrong

Lesson 1 showed that normalization, the step that turns symbols into words, happens before the voice is involved. ElevenLabs' documentation says normalization is on by default and that larger models handle numbers better; its example has Flash v2.5 saying "one thousand thousand dollars" for $1,000,000. You can pick the bigger model, or you can take the guess away.

Acronyms deserve a rule of their own. If listeners say it as a word (NASA, FEMA), the voice does best when it is spelled as a word. If they say the letters (FBI, EPA), hyphens between the letters get them read as letters. If listeners do not know it at all, narration that skips the full name loses them.

Heteronyms

Some words have two pronunciations and the model has to guess from context: read, lead, live, wind, close, tear, wound, minute, record, present, object, desert. "The record shows" and "they will record the meeting" are different words to the ear. Models usually get these right in ordinary sentences and wrong in titles and fragments, where there is less context. On a short or odd line, a different word avoids the guess: "the vote count" instead of "the record," "will announce" instead of "will read."

Names: respell first

Place names and surnames are where synthetic voices embarrass you locally. Everyone in the valley knows how the creek is said, and the voice does not. The fix that works on every model and every tool is respelling: the word spelled as it sounds, with syllables hyphenated and the stressed one in capitals.

• Bois d'Arc Road becomes "BO-dark Road."
• Commissioner Nguyen becomes "Commissioner Win" if that is how she says it. Ask her. Do not guess for her.
• The invented town of Ruidoso Flats becomes "Rih-DOH-so Flats" in the script, and stays spelled correctly everywhere a reader sees it.

ElevenLabs documents other methods, each tied to particular models as of October 2026. Its newer Eleven v4 accepts IPA written between forward slashes. The older phoneme tags work only with the eleven_flash_v2 model and only on single words. Alias tags and pronunciation dictionaries cover other models and some studio products. Which of these reach you depends on the model a tool calls, and Voice Studio's pages do not list support for them, so treat respelling as the default and test anything fancier on one line before you rely on it.

Punctuation is direction

Voice Studio's help page says the model respects punctuation: periods, commas and line breaks control pacing, ellipses make longer pauses, question marks rise and exclamation marks add emphasis. ElevenLabs' prompting guide for its v3 and v4 models adds that capitals increase emphasis on a word. These are blunt instruments, and they work best used sparingly.

• Periods are the most reliable tool you have. Short sentences give the voice places to land.
• Commas create small breaths. A comma before the key fact gives it a beat of weight.
• Ellipses give a longer, hesitant pause. Good once in a script. Tiresome four times.
• Capitals on one word push emphasis on expressive models. On every third word they make the narrator sound furious.
• Exclamation marks almost never belong in factual narration.

What about explicit pause tags? ElevenLabs' controls page says SSML break tags allow pauses up to 3 seconds on some models, warns that too many cause instability, and says Eleven v3 and v4 do not support them. For v3, which Dialogue Editor uses, the page points to audio tags, punctuation and structure. In practice, a [pause] tag or a new sentence does the job.

Audio tags in narration

Tags like [sighs] or [whispers] belong in drama far more than in factual narration. A narrator who sighs before a budget figure is editorializing. Tags suit characters in clearly dramatized scenes; factual narration stays plain and lets the facts carry the weight.

Sentences built for breath

Synthetic voices do not run out of air, which is exactly why long sentences sound wrong from them. A human would have broken that 40-word sentence in two out of necessity, and listeners expect it. Sentences of 8 to 20 words, one idea each, with the subject and verb close together, sound most natural.

Timing a script

The Text2Speech counter estimates duration at 150 words per minute. That is a calm narration pace. Ads run faster; somber pieces slower.

The proof loop

Careful producers check a script in a loop before spending real credits:

1. A read-aloud at speaking pace, with a stopwatch. Anything a human trips over, the voice will too.
2. One draft on a cheap model.
3. One listen for meaning and one for names and numbers.
4. A transcript of the draft from Speech2Text, compared with the script. Mismatches mark the problem lines.
5. Fixes in the script, not the audio, with only the changed lines regenerated.