Why Text-to-Speech Mispronounces Words (and How to Fix It)

You’re listening to something, everything’s going fine, and then the voice says “the author has reed the manuscript” or renders a perfectly ordinary surname as an unrelated word. It’s jarring out of proportion to its importance, and it’s the most common complaint people have about text-to-speech.

It’s also usually not the voice’s fault. Understanding where in the pipeline it goes wrong tells you which fixes will work and which are a waste of time.

The step before the voice

A speech system isn’t one thing. Before any audio is generated, your text goes through normalisation — converting written text into a sequence of words the model can pronounce. This step decides:

• Is Dr. “doctor” or “drive”?
• Is 1996 “nineteen ninety-six” or “one thousand nine hundred ninety-six”?
• Is 3/4 “three quarters”, “three fourths”, “March fourth” or “three slash four”?
• Is St. “street” or “saint”?
• Is NASA a word or four letters?
• Is $1.5M “one point five million dollars”?

None of that is in the text. All of it is inferred. And most mispronunciations you hear are this step guessing wrong, not the acoustic model failing to say a sound. That’s why switching to a fancier voice so often doesn’t help — you’ve replaced the singer, not the sheet music.

The five things that break

1. Homographs

Same spelling, different pronunciation, disambiguated only by grammar:

I will read the paper I already read. The lead engineer found lead in the pipe. They live near a live wire. A tear in her eye, a tear in the page. The bass player caught a bass. Wind the clock before the wind picks up. He’ll object to the object.

Getting these right requires parsing the sentence’s grammar. Systems that synthesise in small chunks — which is common for streaming, because it reduces latency — have less context to work with and get them wrong more often. It’s a genuine engineering trade-off between responsiveness and accuracy, and it’s why the same sentence can be read correctly by one app and wrongly by another using the same underlying voice.

2. Proper nouns

The hardest category, and the one no amount of model quality fixes. Given the letters Siobhan, Nguyen, Worcester, Cholmondeley, Xiaomi or Nietzsche, there is no rule derivable from English orthography that produces the right answer. The system either has the name in a lookup dictionary or it guesses from spelling-to-sound rules that were never designed for it.

This is why place names, author names and drug names are reliably the worst part of listening to academic or medical material.

3. Acronyms and initialisms

Three behaviours, and the system has to pick:

• Say it as a word: NASA, NATO, laser
• Spell it out: FBI, HTML, PDF
• Hybrid: JPEG (“jay-peg”), SQL (either)

Novel or domain-specific acronyms default to whichever the model’s heuristics favour, which is why a paper full of specialist abbreviations can become genuinely hard to follow. Worse, some acronyms collide with real words — US, IT, WHO, AN — and the system has to decide from context whether you meant the country or the pronoun.

4. Numbers, symbols and units

Version numbers read as decimals. Ranges read as subtraction. Ratios read as dates. Units expanded inconsistently — 5 km might be “five kilometres” or “five kay em”. Currency, percentages and scientific notation all have multiple valid renderings.

5. Citations and notation

The special hell of academic and legal material: § 12.3(b)(ii), id. at 47, (Smith et al., 2019, pp. 45–47), Fig. 3b, p < .05. Even when read correctly, these are unfollowable by ear — they’re visual shorthand that was never meant to be spoken. How to listen to research papers covers working around them, and text-to-speech for lawyers covers the legal version.

What actually fixes it

Use a pronunciation dictionary if your app has one. This is the real solution. A custom lexicon lets you map a word to your preferred pronunciation permanently, so the surname in your thesis or the drug name in your notes is right every time. NaturalReader offers this; several other tools do too. If you’re listening to a lot of material in one specialist domain, this feature is worth more than a bigger voice catalogue — see the Frateca vs NaturalReader comparison.

Respell it in your own text. For material you control — your notes, your drafts, your scripts — just write it phonetically. Nietzsche → Neecha. Siobhan → Shiv-awn. Ugly on the page, correct in your ear. This is the standard trick for anyone producing audio from their own writing; text-to-speech for content creators covers script preparation more broadly.

Slow down. A mispronounced word at 1.2× is decodable from context; at 2× it’s noise that derails the whole sentence. If a document is full of unfamiliar terms, drop the speed rather than fighting the pronunciation.

Learn to auto-correct. After ten minutes with any technical text, your brain starts silently fixing a consistent mispronunciation. Consistency helps here — a voice that says a name the same wrong way every time is far easier to listen to than one that varies.

Punctuate your own text properly. For material you write, punctuation is the main signal the system has for phrasing and pause length. A missing comma is a missing breath.

💡 What doesn’t work: cycling through voices hoping one gets the name right. Voices from the same provider usually share the same front-end normalisation, so they’ll make identical mistakes. You’re changing the timbre, not the decision.

When mispronunciation means “read this with your eyes”

There’s a useful signal buried in all this. When a passage dissolves into notation soup — dense citations, equations, tables, code — that’s the document telling you it isn’t audio material. The right response isn’t a better voice; it’s to switch modes.

That’s the core discipline of listening to hard material: listen for structure and argument, read for notation and detail. Use the audio pass to find out which paragraphs need your eyes, then spend your desk time only on those.

The realistic expectation

Modern neural voices handle ordinary prose extremely well — well enough that the gap with human narration has genuinely narrowed for most content. What they don’t do is know how to say a surname they’ve never encountered, and no version bump is going to fix that, because it isn’t a modelling problem.

Set your expectations at the level of the material: flawless on articles and books, occasionally odd on academic and technical text, and reliably wrong on unfamiliar names. Then use a dictionary for the names you hear often.

If you want to know what separates a voice that’s pleasant for an hour from one that isn’t, our tested guide to text-to-speech voices covers it. And to try natural voices on your own material, Frateca is free to start — free plan, no credit card required.