When an AI voiceover sounds robotic, people blame the voice. Nine times out of ten it's the script. Modern TTS models follow the text's rhythm — so if the text reads like a press release, the audio sounds like one. These seven tricks fix that.

1. Write for the ear, not the eye

Sentences that look fine on a page often run too long out loud. If you can't say a sentence in one comfortable breath, split it. Contractions ("it's", "you're", "don't") are not optional — written-out forms are the fastest way to sound like a terms-of-service reading.

2. Punctuate the pauses you want

Commas produce small pauses, periods bigger ones, and paragraph breaks the biggest. An em dash — like this — creates a natural beat that works well before a reveal. If a line rushes past too fast, the fix is usually punctuation, not speed.

3. Use explicit pause and breath marks

uttrd's editor understands two bits of markup: [pause:2s] inserts a deliberate silence, and [breath] adds an audible intake of breath. A pause before your key point does more for emphasis than any exclamation mark; a breath at the top of a long paragraph makes the whole read feel recorded rather than rendered.

4. Give the voice a role, not adjectives

In the Instructions box, "warm, professional, engaging" does very little — every model has heard those words a million times. A role works better: "a documentary narrator letting the story build slowly" or "a friend explaining something they're genuinely excited about". Roles carry pacing, tone, and emphasis all at once.

5. Vary the energy across the script

Human speakers don't hold one energy level for three minutes. Write the variation into the script: a short punchy line after two long ones, a quiet aside, a question. The voice will follow the shape you give it.

6. Match speed to content

The speed slider isn't just for fitting a time slot. Tutorials hold up at 0.9–1.0×, energetic shorts can take 1.1–1.25×. Always listen at the speed you'll ship — pacing that feels right at 1× can smear at 1.5×.

7. Generate twice, keep the better take

Generation isn't deterministic — the same script produces slightly different reads. For anything important, generate two or three takes; they all land in your recent-takes history, so you can replay them side by side, name the keeper, and download it. That's what the naming feature is for.