How to Write Numbers, Dates, Money, and Units for Better TTS Pronunciation

Numbers are compact for readers and ambiguous for speech. A human can look at “1/5,” “2026,” “$14.99,” or “3–30 s” and infer the intended reading from context. A text-to-speech model has to choose a spoken form.

For production work, the safest strategy is not to assume that every numeric notation will be interpreted the way you expect. Decide how the number should sound, test a short line, and write ambiguous forms more explicitly when needed.

This guide covers the cases that cause the most avoidable pronunciation errors.

Decide the spoken form before formatting the text

Start by asking a simple question: if a human narrator were reading this line, what exact words should the listener hear?

Consider:

You may want “fourteen dollars and ninety-nine cents,” or you may be happy with a model that says “fourteen ninety-nine.” The written symbol alone does not communicate your production preference.

When the exact wording matters, write the spoken form explicitly:

This is especially useful for ads, tutorials, financial explanations, product pricing, and any line where a misread number damages trust.

Years are not always ordinary cardinal numbers

A year such as 2026 may be spoken as “twenty twenty-six” in English. A model could also produce a more literal form depending on context and language.

If the year is important to the script, test it in a complete sentence:

If the result is not the intended reading, write:

Do not globally replace every year with words without listening first. Numeric notation is often handled correctly, and explicit wording can make source scripts harder to maintain. Use explicit forms where consistency matters.

Dates should avoid locale ambiguity

A date such as 08/09/2026 is ambiguous even to humans. It can mean August 9 or September 8 depending on locale.

For speech scripts, prefer a form that communicates the intended date directly:

or:

For localized scripts, use the conventional written order for the target language and listen to the generated result. Do not assume that changing the interface language alone makes a numeric date unambiguous.

When a date appears in a legal, billing, release, or event context, clarity is more important than compactness.

Currency symbols need context

$ can refer to several currencies, and a listener cannot see the symbol after it has been spoken. If the currency matters, name it.

Instead of:

Use:

when that is the intended meaning.

For a product price such as $14.99, compare a natural numeric input against the explicit spoken form. Keep the version that sounds correct and remains understandable to the audience.

For multilingual content, avoid copying an English currency formatting pattern into every language. Number grouping, decimal separators, and spoken currency order vary.

Percentages should be tested with surrounding words

A simple percentage such as 25% is often handled well. Problems appear when a sentence contains several percentages, decimals, or ranges.

Instead of presenting a dense line:

consider whether the spoken content needs all three numbers. If it does, split the statement:

For narration, comprehension is a quality metric. A technically correct but overloaded sentence can still be poor audio.

Decimal points can sound unnatural in long values

A value such as 3.14 is straightforward. A long decimal such as 0.037582 may sound tedious or may be read digit by digit.

Ask whether the listener actually needs the full precision. If not, round the number in the script rather than expecting punctuation to solve the pacing.

If exact precision is required, test how the model reads the decimal and write the intended spoken sequence explicitly if necessary.

Do not alter scientific, financial, or medical precision merely to make the voice sound smoother. In high-stakes content, preserve accuracy and use a human review process.

Ranges are easy to misread

A compact range such as 3–30 seconds can be read correctly, but punctuation and dash type may change behavior.

For a stable voiceover line, write:

when the exact reading matters.

This is particularly helpful in user instructions. IndexTTS Online, for example, describes a short reference-audio range in product guidance. The reader needs to understand the duration, not admire compact typography.

Use the same approach for ranges of dates, prices, or quantities.

Units and symbols should match the audience

5 km, 10 GB, 44.1 kHz, and 250 ms combine numbers with abbreviations. Technical audiences may prefer the abbreviations. General audiences may understand the spoken expansions better.

A tutorial might say:

rather than:

Both can work. The choice depends on the output, not on what looks shorter in the source text.

If an abbreviation is pronounced badly, expand it. If the abbreviation is an established term your audience expects to hear, test whether the model already handles it naturally before changing it.

Ordinals need special attention

Strings such as 1st, 2nd, 3rd, and 21st can appear in dates, rankings, and instructions.

Test the complete line. If “21st” is unstable, write “twenty-first.” For dates, prefer the natural date form for the target language rather than manually combining a number and suffix.

The same rule applies to versions:

may be clearer than:

when the listener is not looking at a screen.

Large numbers should be edited for listening

A large value such as 1,250,000 can be read accurately and still overwhelm a listener when surrounded by other figures.

You can improve comprehension by combining precise numbers with verbal framing:

If exactness is required:

Choose based on what the audience must remember.

A voice script is not a spreadsheet. It should preserve the information that matters without forcing the listener to parse a wall of digits.

Acronyms next to numbers create extra ambiguity

A line such as API v2.5 supports 5 languages contains an acronym, a version number, and a count. Test the line as a unit.

Possible production form:

Whether you should expand “API” depends on the desired pronunciation and audience. The key is to diagnose each ambiguous token instead of adding punctuation around the whole sentence.

Multilingual scripts need language-specific review

A numeric format that works in English may be wrong or unnatural in Spanish, Japanese, Chinese, or Arabic.

Before generating multilingual content:

1. localize the number format, not just the surrounding words;
2. confirm decimal and thousands separators;
3. use an unambiguous date format;
4. state the currency when necessary;
5. test units and abbreviations in the target language;
6. listen to the complete sentence, not the number in isolation.

Cross-language voice cloning does not remove the need for language-aware script editing.

Build a small numeric regression set

If your project repeatedly contains prices, dates, version numbers, percentages, or measurements, keep a small test script.

For example:

Generate this test whenever you change the model, language, or major formatting rules. It gives you a quick way to detect unexpected pronunciation changes before a long recording session.

Use one-variable tests when a number sounds wrong

If the line fails, do not immediately change the reference voice, pace, punctuation, and model.

Test two text versions first:

A:

B:

Keep everything else unchanged. If B fixes the issue, the problem was text normalization, not the voice reference.

This is the same controlled-testing principle used in Voice Lab comparisons.

Final checklist

Before final generation, scan the script for:

• numeric dates that could be interpreted in two ways;
• currency symbols without a named currency;
• long decimals or unnecessary precision;
• ranges written with ambiguous dashes;
• compact units the model may not pronounce correctly;
• ordinals and version numbers;
• large numbers that are hard to understand by ear;
• identifiers that should be read character by character;
• locale-specific separators copied from another language.

The goal is predictable speech, not maximum compactness.

Whenever a number is important, write for the listener. A few extra words in the script can prevent a confusing or unprofessional line in the final audio.