ElevenLabs voiceover on Clipwave runs the Multilingual v2 model, and prompting it is script direction: the model reads exactly what you write, so pacing, pauses and emphasis are controlled inside the text — punctuation, capitalization and <break> tags — while the character of the read is set by two sliders, Stability and Similarity. What it is not: a tag-directed system. Bracketed directions like [laughs] or [whispers] belong to a different model generation and are read aloud here as literal words.
Write the way people talk: contractions, short sentences, concrete words. The model mirrors the register of the text — stiff prose produces a stiff read.
For deliberate silence use <break time="1s" /> (any value up to 3 seconds). For micro-pauses use punctuation: commas, dashes and ellipses — less precise than the tag, but gentler on the read.
CAPITALIZE the one word that carries the stress; exclamation marks lift energy and question marks produce a real rising intonation. Emphasis works by contrast — one word, not one sentence.
Write numbers, times, prices and acronyms as words ("ten thousand", "seven thirty", "A-P-I"). Numerals leave room for interpretation; words read one way only.
The delivery follows the content: an excited script reads excited, a solemn one reads solemn. For narration, framing lines like "she said, barely believing it" steer tone — but they are read aloud, so they only fit storytelling formats.
Voice choice plus Stability (consistency vs expressiveness) and Similarity (adherence to the voice's core timbre) are settings, not text. Set them per content type, not per sentence.
Template
[hook — one short spoken sentence] <break time="0.6s" /> [point one, spoken register, numbers as words]. [point two — CAPITALS on the single key word]... [payoff after the ellipsis]. [final line, short and directive].Okay so... I did NOT expect this. Two weeks with this serum, and my skin is honestly glowing. No filter. No tricks. Just... look at it!
Caps on "NOT" put the stress exactly where the joke of surprise lives; the ellipses read as authentic hesitation beats; the exclamation lifts the closer. Lower stability lets the energy through.
Here is how it works. <break time="1s" /> First, you connect your store. Takes about two minutes. <break time="0.8s" /> Next, the engine scans every product page... and flags the ones losing you money. <break time="1s" /> Last step. You approve the fixes, one click each.
Break tags carve the read into chapters the ear can follow. Keep them few — official guidance warns that many break tags in a single generation can destabilize the output.
Settle in. <break time="2s" /> Let your shoulders drop... slowly. <break time="2.5s" /> Nothing to do now. Nowhere to be. <break time="3s" /> Just this breath.
Three seconds is the maximum a break tag accepts — for longer silence, add it in the edit. High stability keeps the pace even and hypnotic instead of drifting.
FORTY-EIGHT hours. That is all that is left! Every plan, half off — every single one. When the timer hits zero... it is gone. Go!
Low stability plus caps and exclamation marks produce urgency without sounding shouted; the ellipsis before "it is gone" is the dramatic beat. Short imperative closer lands hard.
The dashboard gives you three numbers that matter: revenue today, refund rate, and inventory at risk. Click any of them, and you get the full breakdown — by product, by channel, by week.
Higher stability produces the steady, corporate read a demo needs. The colon-and-list structure creates natural rhythm; the em-dash holds a beat before the enumeration.
Welcome back to Margin Notes — the show about the boring businesses that quietly print money. Today... self-storage. Why is it everywhere? And who is actually getting rich off it?
Questions in the text produce genuine question intonation — no direction needed. Mid stability keeps it conversational: consistent enough for a series, loose enough to sound human.
Doors open at seven thirty. Tickets are forty-five dollars at the gate — cheaper online. The show runs about two hours, with one intermission.
Multilingual v2 handles numerals reasonably well, but writing them out removes all ambiguity — "7:30" and "$45" each have several possible readings; "seven thirty" and "forty-five dollars" have one.
Benvenuto. Questa è la parte facile: carichi il video, scegli lo stile, e il resto succede da solo. Due minuti, davvero.
Multilingual v2 speaks 29 languages and reads the language from the script itself — there is no language setting. Keeping the same voice across markets preserves brand continuity; keep one language per generation.
The letter arrived on a Tuesday. She turned it over twice before opening it... and when she finally did, she laughed — a short, surprised laugh — and read it again.
No [laughs] tag here — the laugh is narrated, not performed, and the model colors the delivery from the mood of the words. This is how emotion is directed on Multilingual v2: through the content itself.
Stop scrolling. This took me four years to learn, and it will take you thirty seconds. The fastest way to grow on any platform is... consistency you can actually keep. Here is the system.
Front-load the hook in the first three words. The ellipsis before the payoff is the retention beat — a fraction of silence that makes the answer feel earned. Pair the finished track with visuals cut to it, not the reverse.
Twelve curated voices; Rachel is the default. Audition your actual first line on two or three voices before committing — the voice changes perceived brand more than any slider.
0-1, default 0.5. Lower is more emotional and variable (0.3-0.45 for UGC and energy), higher is more even (0.6-0.8 for tutorials and long narration). At 1.0 the read turns monotone; very low can wander.
0-1, default 0.75. How tightly output sticks to the voice's core timbre — 0.7-0.85 is the working range. Raise it if the character drifts between takes.
There is no speed slider on Clipwave — pace is written, not set. Shorter sentences and fewer commas read faster; ellipses, dashes and break tags slow the read down.
The exact syntax is <break time="1.0s" /> with any value up to 3 seconds, placed where the silence should fall. It is the most precise pause control the model has — and the one with a documented failure mode: too many tags in a single generation can make the output unstable. The working discipline is to reserve tags for structural pauses (between sections, before a reveal) and let punctuation carry the micro-pauses: a comma breathes, a dash holds a beat, an ellipsis adds weight. If a script needs silence every few words, that is an editing problem, not a tagging problem.
Multilingual v2 has no performance tags, so emotion is directed through three channels. First, the content itself — the model mirrors the mood of the words, so a script that reads urgent sounds urgent. Second, punctuation as an energy dial: exclamation lifts, question rises, ellipsis weighs, dash suspends. Third, for storytelling formats only, narrative framing ("she whispered", "he said, too quickly") — remembering it is read aloud as part of the text. The stability slider then decides how much of that direction comes through: lower values give the model room to perform, higher values flatten toward consistency.
The track you generate here is a building block for the rest of Clipwave. Attach it as the audio for a HeyGen talking avatar and the lips follow this exact performance — voice direction and presenter become two separate, controllable steps. Or lock the voiceover first and cut b-roll and generated footage to it: audio-first editing keeps pacing decisions in the script, where you control them, instead of in the timeline. The reliable order is script, then voiceover, then visuals — not the reverse.
No. Those are audio tags from a newer model generation; the Clipwave integration runs Multilingual v2, which reads bracketed text as literal words. Direct emotion with wording, punctuation and the stability slider instead.
Insert <break time="1.5s" /> exactly where the silence goes — any value up to 3 seconds works. Use it sparingly; for micro-pauses, commas, dashes and ellipses are softer and more stable.
29, including English, Spanish, French, German, Italian, Portuguese, Hindi, Japanese and Chinese. There is no language setting — the model reads the language from the script text. Keep one language per generation for the cleanest read.
There is no speed setting on Clipwave — pace lives in the writing. Shorter sentences and lighter punctuation speed the read up; ellipses, dashes and break tags slow it down. If the pace is wrong, rewrite the sentence rather than fighting it.
Generation is variable by design, and Stability is the dial. Push it toward 0.7 for repeatable takes; keep it low when you want performance and are willing to pick the best of a few runs.
Respell it phonetically in the script: "SEE-kwel" for SQL, "loo-MAY-ah" for Lumeya. Phoneme tags are not supported on this model, so ear-spelling is the reliable fix — apply it at every occurrence.