AI Voice Pronunciation: Fix the Names It Gets Wrong

Fix a mispronounced name or word by climbing a short ladder from the cheapest fix to the most durable one: respell it phonetically in the script, then use spacing or letter-by-letter treatment for acronyms, then add it to a pronunciation dictionary, then use SSML phoneme tags if your tool supports them, and swap the voice only as a last resort. The choice between these is not about which "works," most of them work, it is about how far the fix travels. A respelling fixes one spot. A dictionary entry fixes the term everywhere it appears, forever. Pick the rung that matches how often the word recurs.

This is a method you apply yourself, drawn from vendor documentation and the speech-synthesis standard, not from tools we have personally run. The rest of this article is the ladder, in order, plus one trap that catches creators who reach for the quickest fix without thinking about their captions.

AI voice pronunciation: first, diagnose which word is failing

The right fix depends on the category of the problem, so name it before you reach for a technique.

AI voices reliably struggle with a few specific classes. Proper nouns are the biggest: personal names, place names, company and brand names, especially anything with an unusual or non-English spelling. Acronyms and initialisms are the second class, where the engine cannot tell whether to say the letters (F, B, I) or read it as a word (NASA). Homographs are the third, where one spelling has two pronunciations and the model picks the wrong one for your context, like "read," "live" or "lead." And there is plain emphasis, where the word is technically right but stressed on the wrong syllable. Knowing which of these you have tells you which rung to start on.

Rung one: respell it phonetically in the script

The fastest fix, and the one every tool supports because it needs no special feature, is to change the spelling until the voice says it right.

You are not correcting English, you are feeding the model letters that produce the sound you want. Descript's own guidance (help.descript.com, read 2026-09-11) puts this first: "Modify the word's spelling in the script. Experiment with phonetic spellings until you achieve the desired pronunciation." The classic examples are the awkward ones: writing "Worcestershire" as "Wuh-stuh-sher," or a brand like "UNICEF" as "yoonisef." You generate, listen, adjust the spelling, and generate again until it lands.

This rung is perfect for a word that appears once or twice. It is quick, it needs nothing, and it is under your control. Its weakness is that it does not travel, which is the whole reason the ladder continues.

Rung two: spacing, punctuation and letter-by-letter for acronyms

Acronyms have their own set of fixes because the problem is not the sound of a word, it is whether to treat it as a word at all.

If you want the letters read out, separate them so the engine stops trying to pronounce them together. Adding spaces, dashes or periods between the letters (F B I, or F-B-I) pushes most engines to spell it. If you want it said as a word, leave it solid and, if needed, respell it as on rung one. Punctuation also controls pacing more broadly, as Descript notes that punctuation affects the spacing of generated audio, so a comma or a period can fix a run-on that was mangling a name by rushing it. These are small, per-instance tweaks in the same family as respelling.

Rung three: a pronunciation dictionary, so the fix travels

When a term recurs, your brand name in every video, a product name in every episode, stop fixing it by hand and teach the tool once.

Many text-to-speech platforms offer a pronunciation dictionary, sometimes called a lexicon or custom pronunciations: a lookup table that maps a written term to the pronunciation you want. When the engine meets that term anywhere in your text, the entry overrides its default guess. This is the correct rung for anything that repeats, because you fix it a single time and it stays fixed across every future generation, instead of hoping you remember the respelling six months from now. Feature names and availability differ by tool and change over time, so check your platform's current documentation for what it calls this and how entries are formatted. The concept is the same everywhere: a durable, reusable override.

Rung four: SSML phoneme tags, for exact control

When you need precision rather than trial and error, and your tool accepts it, the speech-synthesis standard gives you a way to specify the exact sounds.

SSML, the Speech Synthesis Markup Language, defines a phoneme element that, per the W3C specification (w3.org/TR/speech-synthesis11, read 2026-09-11), "provides a phonemic/phonetic pronunciation for the contained text" through a required "ph" attribute holding the phoneme string. Instead of tricking the model with respelling, you state the pronunciation directly in a phonetic alphabet. The same standard defines a sub element, which uses an "alias" attribute to specify "the string to be spoken instead of the enclosed string," so the written text and the spoken text can differ deliberately. Not every consumer tool exposes raw SSML, and the ones that do vary in how much they support, so this rung is available when you are working somewhere that accepts it. Where it is, it is the most exact fix there is.

Rung five: change the voice

Sometimes the fastest real fix is not to fight the model but to pick a different one.

Voices differ in how well they handle unusual names and technical terms, so previewing your problem word across a few voices can occasionally solve it with no editing at all. Treat this as a last resort rather than a first move, because you usually choose a voice for its tone and brand fit, not its handling of one word. But if two voices are otherwise close and one says your brand name correctly, that is a legitimate tiebreaker.

The trap: respelling can wreck your captions

Here is the gap most quick-fix lists leave out, and it matters for anyone shipping video.

Rung one changes the actual text. If your workflow generates captions or a transcript from that same script, your beautifully respelled "Wuh-stuh-sher" will appear on screen exactly like that, which is worse than a mispronunciation. This is precisely why the higher rungs exist. A pronunciation dictionary and the SSML sub element both change what is spoken while leaving the visible or source text intact, which is the entire point of the sub element's separate "alias." So if your project has captions, prefer a dictionary or SSML for recurring terms and keep respelling for audio-only work, or remember to correct the display text separately. A fix that repairs the audio and breaks the subtitles is not a fix.

FAQ

How do I make an AI voice say a name correctly? Start by respelling the name phonetically in the script until it sounds right, which every tool supports. If the name recurs, add it to the tool's pronunciation dictionary so the fix applies everywhere automatically. For exact control where your tool supports SSML, use a phoneme tag.

How do I stop an AI voice from mangling acronyms? Decide whether you want the letters read out or said as a word. To spell it, separate the letters with spaces, dashes or periods (F B I). To say it as a word, keep it solid and respell it phonetically if needed. Punctuation also controls pacing around it.

What is a pronunciation dictionary and when should I use one? It is a lookup table mapping a written term to the pronunciation you want, which the engine applies wherever the term appears. Use it for any word that recurs, like a brand or product name, so you fix it once instead of respelling it every time. Features and names vary by tool, so check current docs.

Will respelling a word for the voice also change my captions? Yes, if your captions come from the same script, which is the trap. Respelling changes the visible text. For recurring terms with captions, use a pronunciation dictionary or the SSML sub element instead, since both change the spoken audio while leaving the written text intact.

The one habit that saves the most time

Match the fix to how often the word appears. Respell the one-offs, and put anything you will ever say again, your name, your brand, your recurring terms, into a pronunciation dictionary or SSML the first time it trips the voice up. The vendor documentation for your specific tool, plus the SSML specification at w3.org/TR/speech-synthesis11 (read 2026-09-11), are where to confirm exactly what your platform supports, because these features move.