Write the way people talk
A script that reads well silently can still sound stiff aloud. Spoken language uses shorter sentences, contractions and the main point up front, because a listener cannot glance back at the previous line.
Read the script out loud once before generating it. Anywhere you run out of breath or stumble, the voice will sound rushed or flat too.
- Keep most sentences short; split anything with more than one comma-separated aside.
- Use contractions — "it's", "you'll" — unless the tone is deliberately formal.
- Lead with the point, then the detail; a listener remembers the first half of a sentence best.
- Cut words that only work on screen, like "see below" or "click the link".
Spell out what should be heard
Voices read ordinary numbers and dates naturally, but anything ambiguous is safer written the way it should be said: ranges like 3–4, fractions, version numbers, phone numbers, model names and units.
ToneCraft expands a handful of common abbreviations automatically — e.g., i.e., etc., vs., Dr., Mr., Mrs., No. — and leaves everything else as written. Write other abbreviations out, and leave raw links out of the script or write them the way you would say them.
- "3–4 days" → "three to four days"
- "v2.1" → "version two point one"
- "50%" → "fifty percent" when the exact wording matters
- "example.com/start" → "example dot com slash start"
Fix names and jargon once
Brand names, people's names and technical terms are where synthetic voices most often trip. Rather than respelling them in every script, add a pronunciation rule: "say this" → "as this", for example SQL → sequel.
In ToneCraft, rules are applied to every voice-over, literally and in the order they are listed, before the voice is generated. Because they match the exact text, add a rule for each spelling you actually use, and test it on a short line first.
Use paragraphs on purpose
Leave a blank line between paragraphs. Paragraphs give the voice natural breathing points, and they are also the unit you edit later.
Long scripts are split on sentence boundaries and stitched back into one seamless file, so length is not a problem. When you change a script and generate again with the same voice and settings, unchanged paragraphs are reused and only the paragraphs you edited are charged.
Direct the performance with tags, sparingly
Performance tags tell the voice how to deliver a line instead of being read aloud. On the standard voices that means tags such as [pause], [laugh] or [breath], and wrap tags like <whisper>…</whisper> around a short phrase. Premium voices use their own set of bracket tags.
A tag from the other voice family is not performed, and a tag needs real words next to it. Use them where a person would actually change delivery — a beat before a reveal, a laugh at a joke — rather than on every line.
Give each speaker a label
For dialogue, interviews or a two-host podcast, start each speaker's paragraph with an upper-case label and a colon, such as HOST: or GUEST:. Each label is cast to its own voice, and the labels themselves are neither spoken nor charged.
Labels have to be in capitals at the start of a paragraph: "Host:" is treated as part of the text, and the editor flags labels that look almost right.
Check the length before you generate
At about 150 words per minute, a one-minute voice-over is roughly 150 words; the free words-to-minutes calculator gives an estimate for your own pace. ToneCraft plans are sized at about 917 characters per minute of audio, and the free plan includes 25,000 characters a month.