Four habits fix most acronym and jargon problems in AI narration. Write letter-by-letter acronyms with spaces, write word-like acronyms the way they sound, spell out units and version numbers, and describe code instead of reading it. A voice reads the text you give it, so the text has to say what you want to hear.
This post is about everything except names. If a person or place is what's going wrong, fixing a mispronounced name covers that. Here: SQL, GDPR, v2.4.1, getUserId, 5 GB and H2O.
Make text to speech read acronyms letter by letter
An acronym can be said two ways. GDPR is spelled out, NASA is a word, and SQL is spelled out by some people and "sequel" by others, so nobody is wrong, which is exactly the problem: the voice has to guess, and a guess can land differently from one sentence to the next, even inside the same lesson.
We can't promise how a given voice will read any one acronym, and you should doubt anyone who does. What you can control is the text. Decide the reading you want, write it that way, and listen to one line before you commit.
- GDPR, spelled out — G D P R
- SQL, spelled out — S Q L
- SQL, as a word — sequel
- NASA, as a word — Nasa, or NAH-suh if it still stumbles
- PDF, API, CEO — P D F, A P I, C E O
- e.g. and i.e. — for example, and that is
- vs. — versus
Spaced capitals are the usual way to push a voice into letter-by-letter reading. They don't always behave identically on every voice, so test the real sentence. Plurals are the awkward case: "A P Is" looks odd and can be read oddly, so reword to "A P I endpoints" and sidestep it.
Under GDPR, your PII must be stored per the SOP, e.g. in the CRM.
Under G D P R, your P I I must be stored per the S O P, for example in the C R M.
Pick one reading per acronym and keep it
In a course or a series, consistency beats correctness. If lesson 2 says "sequel" and lesson 9 says "S Q L", learners wonder whether you mean two things. Write the decision into a short glossary and apply it with find-and-replace before each generation.
Version numbers, dates and numbers with dots
Symbols and dots in numbers are ambiguous out loud. Is 3.11 "three point eleven" or "three point one one"? Software people say the first, a general-purpose voice may say either, and a learner who hears the wrong one will go looking for a release that doesn't exist. Write out what you mean.
Dates deserve the same care, since 11/12 is a different day in different countries. Spell the month as a word and the problem goes away.
- v2.4.1 — version two point four point one
- Python 3.11 — Python three point eleven
- 2026-11-24 — November twenty-fourth, twenty twenty-six
- Ticket #4471 — ticket number four four seven one
- 1/3 of users — one third of users
- 10x faster — ten times faster
The long form looks silly on the page and sounds right in the ear, which is the only place it's used. Keep the clean version in your source document and generate from a separate narration copy, the same way you would for a respelled name.
Text to speech technical jargon: narrating code
Reading code aloud is nearly always a mistake. Underscores, camelCase and brackets turn into a recital that nobody follows, and the learner who is trying to type along falls a line behind within seconds. Show the code on screen and narrate what it does. Say the exact name only when the learner has to type it.
Call getUserId() and store the result in user_id.
Call the get user I D function and store what it returns in a variable called user underscore I D.
Describe, don't recite
"The function that fetches the user's ID" is easier to follow than five syllables of camelCase.
Say symbols only when they matter
A learner typing a command needs "dash dash version". A learner understanding a concept doesn't.
Spell tool names the way you say them
"N P M install" if you say the letters, "nginx" respelled as "engine X" if that's how your audience says it.
Technical lessons also benefit from a precise, even voice, since a showy delivery fights the material and makes dense terms harder to catch. The lesson voiceover tool is built around that kind of steady narration, and it's a good place to audition voices on your own wording.
Pronounce abbreviations: units, chemistry and medical terms
Units are where the shorthand is densest. "5 GB" reads better as "5 gigabytes", "mg/dL" as "milligrams per deciliter", and "km/h" as "kilometers per hour". The slash is the usual culprit: it has no sound, so spell out the word it stands for.
- 5 GB — 5 gigabytes
- mg/dL — milligrams per deciliter
- H2O — H two O, or just water
- CO2 — C O two, or carbon dioxide
- NaCl — sodium chloride
- 10 µm — 10 micrometers
One thing you don't have to do by hand: temperature symbols. VocalLab spells out ℃, °C and °F before the text reaches the voice, so "25°C" is sent as "25 degrees Celsius". It does this for English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese and Korean. Every other unit is on you.
For anything medical or safety-related, be stricter than you think you need to be. Write "five milligrams twice a day" in full rather than "5 mg BID", and listen to the finished clip against the source before it goes to learners.
Write for the ear, not the page
Introduce an acronym the way a person would say it. "General Data Protection Regulation, or G D P R" works once, then G D P R is enough. Parentheses, slashes, ampersands and "and/or" don't survive being spoken, so replace each with a plain word.
Keep sentences short when a lot of jargon is packed in. A line with four terms in it is hard to hear even for a human listener, because nothing on the page marks where one idea ends and the next begins. Split it into two sentences and the voice gets room to breathe, and so does the learner.
Test one line, then keep a pronunciation list
Generate a single sentence containing the trickiest term before you generate the whole lesson. If it reads wrong, change the text and try that same sentence again. A short test costs a few points; a re-run of a full lesson costs the whole lesson.
Spaced letters do add characters, but not many. A 1,700-character script costs ⌈1,700 ÷ 17⌉ = 100 points in the app. Spread 20 acronyms of 4 letters through it, add 3 characters for each, and it's 1,760 characters, or 104 points. Four extra points to get 20 acronyms right is a good trade.
Keep the list in one file: written form, narration form, and the lesson where you confirmed it. Explainer and training videos benefit most, and the explainer video voice generator page covers that format.
Articulate Male Explainer Voice
Technical lessons heavy on terms
- Neutral American
- Articulate, precise
Software and IT tutorials
- Neutral American
- Precise
Frequently asked questions
How do I make AI narration spell out an acronym letter by letter?
Write it with spaces between the capitals, for example G D P R. Test the real sentence, because the result can vary by voice and by position in the sentence.
How do I get an acronym read as a word, like NASA?
Write it the way it sounds. Try "Nasa" first, and if that still stumbles, respell it as "NAH-suh" with the stressed syllable in capitals.
Does VocalLab handle degree symbols automatically?
Yes. ℃, °C and °F are spelled out before the voice reads them, so 25°C becomes 25 degrees Celsius. It works for English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese and Korean.
How should I narrate code or a version number?
Describe code instead of reciting it, and write version numbers in words, such as version two point four point one. Show the exact text on screen.
Do respelled acronyms cost more points?
Slightly. Points are charged on the characters you submit, so spaced letters add a few characters. Twenty four-letter acronyms in a 1,700-character script add about 4 points.