Drop the same sentence into ElevenLabs and Gemini and you'll get two different-sounding results — but drop two differently written versions of the same sentence into the same model, and the gap is usually bigger. The model reads what's on the page literally. If the page reads like a legal disclaimer, it'll sound like one.
Short sentences beat long ones
Text-to-speech models generally place their intonation drops at sentence boundaries. A 40-word sentence with three clauses gives the model nowhere natural to breathe, so it either rushes through the whole thing in one flat run or inserts a pause somewhere arbitrary that a human never would. Breaking that same sentence into two or three shorter ones gives the model real places to land — which is also just how people actually talk.
Before: "Our new update, which took the team almost four months to build and test across every major browser, is now available to everyone starting today."
After: "Our new update is finally here. It took the team almost four months to build. And we tested it across every major browser before shipping it today."
Punctuation is a performance instruction, not just grammar
Commas, periods, and ellipses aren't just there to be grammatically correct — the model treats them as timing cues. A comma is a short breath. A period is a full stop. An em dash or ellipsis creates a longer, more deliberate pause than a comma does. If a line needs a beat before the punchline, an ellipsis often does more work than any amount of rewriting the sentence itself.
This also means over-punctuating backfires. A sentence with a comma every four words gets read in stiff little fragments instead of one flowing thought.
Write for the ear, not the eye
Sentences that read perfectly well on a page can sound stilted out loud, because written and spoken English aren't actually the same register. A few habits that consistently help:
- Use contractions. "It's" and "you'll" sound natural; "it is" and "you will" read as formal even when the meaning is identical.
- Read it out loud yourself first. If you stumble over a phrase or run out of breath reading it, the model will produce something that sounds the same way.
- Avoid stacked numbers and abbreviations. "3-5 business days" reads more naturally as "three to five business days" — most models handle spelled-out numbers more predictably than digits, especially for dates, phone numbers, and ranges.
- Vary sentence length on purpose. A short sentence after a couple of longer ones creates natural emphasis — it's a writing technique, not just a speech one, and it works even better once it's spoken aloud.
Match the model to the material
If you're switching between ElevenLabs and Gemini in 11API, it helps to know what each is actually good at. ElevenLabs' Multilingual v2 model handles emotional nuance and longer narration well — scripts with a real tonal arc benefit from it. Its Flash/Turbo models are faster and cheaper, and hold up fine for short, matter-of-fact lines like UI voiceovers or quick social clips where subtle emotional range doesn't matter as much. Gemini's text-to-speech tends to sound more even and neutral throughout — a reasonable default for instructional or explainer content where you want consistency more than performance.
Test the actual line, not just the script
The paragraph that reads beautifully in a document can still produce an odd result once it's actually spoken — a word gets unexpected emphasis, or a pause lands somewhere strange. This is normal, and it's exactly why regenerating a single line is worth the extra few seconds rather than assuming the first pass is final. Small rewrites — swapping one word, adding a comma, breaking a sentence in a different place — often fix an odd-sounding line faster than fully rewriting it.
The fastest way to get a better-sounding voiceover usually isn't a different model. It's a different sentence.
None of this is model-specific tuning — it's just how spoken language differs from written language, and every current text-to-speech engine responds to it the same general way. Once it becomes habit, scripts come out sounding right on the first or second try instead of the fifth.