We voice reels for several brands with ElevenLabs. Early on, some takes sounded unmistakably synthetic. Changing the voice rarely fixed it. Changing how we produced it almost always did. These are the five changes that made the difference.
1. One continuous take
Generating each line separately and stitching them together sounds mechanical: every line starts with the same energy, the same pace, the same pitch. Generating the whole script in one take lets the voice carry energy from one sentence into the next, the way a person does.
2. Plain punctuation
Short sentences and a blank line between beats give the most natural pauses. We stopped using ellipses, stage directions and bracketed tags. In our tests they made delivery choppier, not more human.
3. Write for the ear
Spoken scripts need shorter sentences than written ones, and numbers written the way you say them. "Three to four weeks" reads better aloud than "3-4 wks". If a line is hard to say out loud, it will sound wrong when a voice says it too.
4. Get the levels right
A voice that is too loud against the music sounds like shouting, however good the delivery. We normalise the voice to around -18 LUFS and duck the music underneath it whenever the narrator speaks. Our first mixes at -15 sounded like the narrator was yelling.
5. Sync captions to the voice, not the script
Captions that appear on time make the whole piece feel deliberate. We detect where each phrase actually starts in the audio and bring the caption in just ahead of it, rather than guessing from the script.
Bonus: lock it once it works
When a client approves a voice, a pace and a mix, we write the exact settings down as that brand's recipe. Every future reel starts from the approved version instead of being re-tuned from scratch.