I have a folder on my desktop called “Hall of Fame.” It is full of audio files where my AI voiceover did something so wrong it looped back around to art.
There is the one where the narrator solemnly informed viewers that honesty is the best policy… and pronounced policy with the confidence of a man saying “pol-see-ee.” There is the classic where a sentence containing “3.5 million” became a brief mathematics lecture. My favorite is still the time a wellness script told me to take a deep breath “in through the nose, out through the mouth,” except the machine heard something else entirely and produced a sound I can only describe as a haunted accordion.
I save these because they taught me more about how TTS actually works than any documentation ever did. Every failure has a pattern. Once you know the patterns, your blooper reel gets boring fast — which, professionally speaking, is exactly what you want.
Here are the ten failure categories I run into most, with the fix I apply every single time.
Why Ambiguous Words Break AI Voices
English is a minefield for text-to-speech. Read the word “read” out loud — was it past or present tense? What about “lead,” “bass,” “wind,” “tear”? Humans resolve these instantly using context. Machines resolve them using probability, and probability is computed from training data, not from the specific joke you are trying to tell.
The result: the rarer your phrasing, the worse the guess. Write a script full of idioms, invented compounds, and niche jargon, and you are essentially asking the model to freestyle. Sometimes it nails it. Sometimes you get a haunted accordion.
The good news is that roughly nine out of ten failures fall into ten fixed categories.
The Top 10 TTS Fails, Ranked by How Hard I Laughed
1. Homographs read the wrong way “I lead the team” versus “the pipe is full of lead.” Fix: rewrite to remove the ambiguity. “I head the team” says the same thing and cannot be misread. You are not dumbing down your script; you are removing a coin flip.
2. Numbers with decimal points “3.5 million users” becoming “three point five… million users” with an awkward pause where the sentence loses confidence. Fix: write it the way a human would say it — “three and a half million users.” Spelled-out numbers almost always produce better pacing.
3. Acronyms pronounced as words — or letters — at random Fix: be explicit. “A-P-P” with hyphens reads letter by letter in most engines. If you want the word (NASA, NATO), write the word. Decide which one you want before you generate, not after.
4. URLs and handles read aloud Nobody has ever wanted to hear “h-t-t-p-s-slash-slash.” Fix: delete them from the script. Say “find us at our website” or “at Soundwaver on social.” If a handle is unavoidable, write it as “at” plus the plain name.
5. Perfect pronunciation, dead expression This is the sneakiest fail because nothing is technically wrong. The voice reads your joke like a tax form. Fix: punctuation is your emotion controller. Exclamation points, em dashes, and short punchy sentences give the model somewhere to put energy. One long sentence with one comma at the end is a one-way ticket to Monotone City.
6. Numbers in currency and dates “$19.99” read as “nineteen point nine nine dollars” instead of “nineteen ninety-nine.” Fix: write money the way cashiers say it. Write dates the way humans say them. Your script is a performance, not a spreadsheet.
7. Homophones “Their team” vs “there team” — the machine often does not care, but your listeners will, slowly, in the back of their minds. Fix: read your script aloud yourself. Your ear catches swapped words your eyes have approved a hundred times.
8. Foreign words and names Half my clients have last names the engine treats like a roulette wheel. Fix: add a phonetic hint the first time a name appears, or accept a simplified pronunciation and move on. Perfection here costs more time than it is worth.
9. Markdown and formatting symbols Hash tags, asterisks, underscores — invisible on screen, chaos in audio. Fix: strip all formatting before generating. Plain text in, clean voice out. I keep a one-click “clean for voice” pass in my notes for exactly this.
10. The run-on sentence Three clauses, four commas, one breath — and the voice sprints to the finish and falls over the line. Fix: cut it into two sentences. Shorter sentences are not lesser writing; for audio, they are the writing.
My 60-Second Preflight Check
Before I generate anything, I run the same four passes:
- Numbers to words. Every decimal, every currency figure, every year.
- Ambiguity scan. Lead, read, present, lead, wind, tear — the usual suspects.
- Symbol purge. Links, handles, hashtags, asterisks, all gone.
- Read it out loud. If your mouth stumbles, the model will too. This step is embarrassingly effective.
Then I listen to the finished audio with attention on two hot spots: proper nouns and anything numeric. Those two categories account for most of what goes wrong. Spot-check them and you catch the rest of the reel before your audience does.
Boring Bloopers Are the Goal
The point is not to fear mistakes. It is to know where they live. A voice tool that misreads an idiom is not broken; it is doing exactly what probability does when handed a joke it has never seen. Your job is to hand it jokes it can survive.
Do that consistently and your “Hall of Fame” folder starts gathering dust. Which is, oddly, the best compliment a voiceover workflow can get.
I still keep mine around, though. Sometimes you need to hear a haunted accordion to remember why you wrote the checklist.
