How to Fix Text to Speech Pronunciation Problems

When a text to speech voice mispronounces a word, the fix is almost always in the text, not in the settings. Spell the word the way it sounds, write numbers as words, expand abbreviations and use punctuation to control pauses. These edits take seconds and work with any voice style.

This guide is based on how the Online Toolsz Text to Speech tool actually processes text. It has no SSML support and no custom pronunciation dictionary, so rewriting the input is the method. The good news is that rewriting is predictable: once you know what the voice does with a symbol or a number, you can steer it.

Diagnose first: one paragraph at a time

Do not generate a whole document and try to find the bad words afterwards. Paste one paragraph, generate, and listen with the text in front of you. Mark anything that sounds wrong. The problem words usually fall into a handful of groups: numbers and dates, abbreviations, symbols, names, foreign words and long sentences with no pauses.

For quick testing, lower the quality steps to around 4. Generation is faster and pronunciation issues are just as audible. Raise the steps again for the final version.

What the tool already changes before reading

Knowing the automatic clean-up saves you from fixing things twice, and explains some surprises:

Removed or turned into spaces: emojis, underscores, slashes, hash signs, square brackets and vertical bars. So "24/7" reaches the voice as "24 7", and "#1" becomes "1".
Replaced with words: "@" becomes "at", "e.g.," becomes "for example," and "i.e.," becomes "that is,".
Normalised: curly quotes become straight quotes, and long dashes become a simple hyphen.
Added: if your text does not end with punctuation, a full stop is added.
Skipped: characters the voice model does not recognise are dropped silently.

For Vietnamese, the tool also expands numbers, common abbreviations and symbols such as % and $ into spoken words. For the other languages, numbers and symbols are left to the voice model, which is where most of your manual fixes will go.

Numbers, dates and money

Digits are ambiguous. "1990" could be a year or a quantity; "3/4" could be a fraction or a date; "10:30" could be a time or a ratio. Write the number the way you want it said.

Written	Risk	Rewrite as
24/7	Slash removed, read as "twenty-four seven" or two separate numbers	twenty-four seven, or around the clock
3/4	Read as two numbers	three quarters, or March fourth
$1.5M	Symbol and suffix read oddly	one and a half million dollars
15%	Percent sign may be skipped	fifteen percent
v2.1	Read as letters and digits	version two point one
2024-25	Read as a subtraction or a long number	twenty twenty-four to twenty-five

Phone numbers and codes are easier to follow in groups. Write "zero two zero, seven nine four six" with commas between groups, and the voice will pause where a person would.

Abbreviations and acronyms

Decide whether the abbreviation is said as a word or as letters. NASA is a word; FBI is letters. The voice cannot always tell.

For letter-by-letter acronyms that come out as a word, add spaces or full stops: "F B I" or "F.B.I.".
For abbreviations that should be read in full, write them out: "approximately" instead of "approx.", "Street" instead of "St." when it might be read as "Saint".
Titles like Mr., Dr. and Prof. are recognised so the sentence is not split after them, but the voice still decides how to say them. If "Dr." sounds wrong, write "Doctor".
Units: write "kilometres", "megabytes", "degrees Celsius" rather than km, MB and °C.

Names and unfamiliar words: respell them

For brand names, surnames and technical terms, the most reliable fix is a phonetic respelling using ordinary letters of the language you are generating. The listener never sees the text, so it does not matter that it looks odd.

"Nguyen" in English audio: try "Win" or "Nwen", depending on the pronunciation you want.
"Siobhan": try "Shi-vawn".
"Kubernetes": try "Koo-ber-net-eez".
"GIF" with a soft G: write "jif".

Hyphens between syllables can help the voice keep them separate. If the stress falls in the wrong place, try changing the spelling of the stressed syllable to make it more prominent, then listen again. It may take two or three attempts. Keep a small list of your respellings so you can reuse them in every script.

Mixed-language text

Each generation is read in one language. The tool picks it from the dropdown, or detects it from the text when the characters make the language obvious. An English product name in a Spanish sentence will be read with Spanish rules; a French phrase in an English script will be read with English rules.

Two ways to handle it: respell the foreign word using the main language's spelling, or generate the foreign phrase separately in its own language and join the clips in an audio editor. The first is quicker; the second sounds more authentic for longer phrases.

Watch out for auto-detection on short snippets. If you paste a two-word phrase, set the language manually.

Pauses and pacing with punctuation

Punctuation is your pause control. A comma gives a short break, a full stop a longer one. If a sentence sounds rushed, the usual cause is length: split it into two.

Use a blank line between paragraphs to force a new chunk. The tool inserts a short silence between chunks, which is the longest pause you can create.
Long text is split automatically at sentence ends into pieces of up to about 300 characters (about 120 for Japanese and Korean). A very long sentence with no full stop cannot be split cleanly, which is another reason to keep sentences short.
Remove decorative punctuation: ellipses used for style, multiple exclamation marks and quotation marks around ordinary words.
For a list, put a comma after each item, or make each item its own short sentence.

If the whole piece feels too fast or slow, use the speed slider (0.80x to 1.30x) instead of adding extra punctuation everywhere.

Pre-export checklist

All numbers, dates, prices and units written as words where the reading could be ambiguous.
Acronyms spaced as letters or written out in full.
Names and jargon respelled and tested individually.
No emojis, hashtags or slashes carrying meaning.
Sentences short enough to say in one breath.
Language set manually for short or mixed text.
Final run at higher quality steps, listened to in full before download.

FAQ

Can I add a custom pronunciation dictionary?

No. The tool has no dictionary or SSML input. Phonetic respelling in the text gives you the same control for most words.

Why does the voice skip a character?

Characters the model does not recognise are dropped, and some symbols such as slashes and hash signs are converted to spaces. Replace them with words.

Does changing the voice style fix pronunciation?

Sometimes slightly, but rarely enough. Different styles share the same underlying reading of the text, so a respelling is more dependable.

Why does a word sound different in another sentence?

Context and the surrounding punctuation affect intonation. Test the word in the actual sentence you will use, not on its own.

Where can I use these fixes?

Anywhere you prepare narration, for example video scripts. See text to speech for YouTube video drafts and text to speech for language practice.