Guides

Writing Scripts for AI Avatars: Pauses and Pacing

Writing scripts for AI avatars that sound human: where to pause, how to pace, how to handle names and numbers, and how to template personalized scripts safely.

PitchTwin team, , 5 min read

An avatar can look exactly like the person and still feel off if the script rushes through a name, never takes a breath, or reads "Q3 ARR" letter by letter. Most of that is fixable in the text you send, not in the video. This guide covers how to write and structure scripts for AI avatars so the result sounds like someone talking, not someone reading.

Two ways to send words

PitchTwin's POST /generations accepts exactly one of three inputs. Two of them carry words that PitchTwin voices in the avatar's cloned voice:

  • text: a single string. The natural choice for short, simple messages.
  • script: an array of { text, pause_after_ms } segments. Use it when timing matters.
{
  "avatar_id": "av_3f9c1e7a2b4d6e80",
  "script": [
    { "text": "Hi Mark,", "pause_after_ms": 400 },
    { "text": "saw Acme just raised. Congrats!", "pause_after_ms": 800 },
    { "text": "I have a quick idea for your onboarding." }
  ]
}

pause_after_ms inserts an exact silence after a segment, up to 10 seconds. The person keeps moving naturally during the pause; it does not freeze the frame. Segments between pauses are voiced together, and each part is voiced with its neighbors as context, so intonation still flows across a pause instead of resetting like a new recording.

The third input, audio, is for when you produce the speech yourself. Pacing then lives in your own TTS or recording; the rest of this post still applies to how you write the words.

Where pauses belong

Pauses are the main tool you have, and the most common mistake is using none.

After the name

A short beat after a greeting makes personalized video feel deliberate. "Hi Mark," followed by 300 to 500 ms reads like someone looking up from their notes. Without it, the name blurs into the next sentence, which is exactly where a viewer is listening hardest.

Between ideas

Put a longer pause, around 600 to 1,000 ms, where a human would change topic: after the hook and before the pitch, or after the problem and before the solution. These are starting points, not rules. Listen and adjust.

Before the call to action

A pause before the ask gives it weight. "So here's what I'd suggest." pause. "Let's take fifteen minutes next week."

What not to do

  • Do not pause after every sentence. Uniform gaps sound mechanical.
  • Do not use long pauses as filler. A 3-second silence feels much longer on video than in a script.
  • Remember the limits: up to 50 pauses and 200 segments per generation.

Pauses add seconds to the video, and you pay per second of finished video (1.5 tokens per second when PitchTwin voices the script). That is rarely a reason to cut them, but it is a reason not to sprinkle them everywhere.

Pacing

Use sentence length first

The cheapest pacing control is the writing itself. Short sentences read faster and punchier. Long sentences with commas slow down and sound more explanatory. Mix them.

Punctuation also shapes delivery: commas create small breaks, a full stop ends a thought, a question mark lifts the end of a line. Use them the way you would when speaking, not the way you would in a formal document.

Then adjust speed

voice.speed accepts values from 0.7 to 1.2, with 1.0 as default.

{
  "avatar_id": "av_3f9c1e7a2b4d6e80",
  "text": "Welcome to module three. Today we cover expense approvals.",
  "voice": { "speed": 0.9 }
}

Slightly slower works for training content and dense explanations. Slightly faster can suit short social clips. Large changes in either direction start to sound unnatural, so change speed in small steps and listen.

Write for the ear

Scripts are heard once, at the speaker's pace, with no way to re-read. A few habits help.

  • Spell out what should be said. Abbreviations and symbols can be read in more than one way. Write "third quarter" if that is what you mean by "Q3", and "annual recurring revenue" the first time you mean "ARR". Write "$1.2M" as "one point two million dollars" when the exact reading matters.
  • Handle tricky names. If a company or person's name is mispronounced, try a phonetic spelling in the script. Keep a small dictionary in your app that maps names to spoken forms, and apply it before sending.
  • One idea per sentence. If a sentence needs to be read twice on paper, it will not land on video.
  • Use contractions. "We're" and "you'll" sound like a person; "we are" and "you will" sound like a policy document.
  • Lead with the point. Viewers decide quickly whether to keep watching. Put the reason for the video in the first sentence after the greeting.

Language

If the script is not in the language the voice would otherwise pick, set voice.language to an ISO 639-1 code, such as en for English or de for German. Write the whole script in one language per generation, and have a native speaker review templates before you scale them. For more on languages, see multilingual avatar video.

Templating personalized scripts

Personalized video usually means one template and many variables. Build the script array in code rather than concatenating strings, so pauses survive personalization:

def build_script(first_name: str, company: str, trigger: str) -> list[dict]:
    return [
        {"text": f"Hi {first_name},", "pause_after_ms": 400},
        {"text": f"I saw that {company} {trigger}.", "pause_after_ms": 700},
        {"text": "We help teams like yours cut onboarding time.", "pause_after_ms": 600},
        {"text": "Worth a quick chat next week?"},
    ]

payload = {
    "avatar_id": "av_3f9c1e7a2b4d6e80",
    "script": build_script("Mark", "Acme", "just opened a Berlin office"),
}

Guard the variables before they reach the API:

  • Empty values. "Hi ," is worse than no video. Fall back to a generic greeting when a field is missing.
  • Length. A long company name or an LLM-generated "trigger" can break the rhythm. Cap variable length, and keep total text under the 5,000-character limit.
  • Odd casing and symbols. "ACME CORP, INC." reads differently from "Acme". Normalize names before inserting them.
  • LLM output. If an LLM writes the variable parts, ask it for short, spoken-style sentences and validate the output before generating.

The personalized video API page covers the larger workflow around templates.

Check the result with word timings

Every generation returns words, word-level timings of the final soundtrack, along with duration_ms. Use them to review scripts at scale: flag videos that run much longer than the template's median, or check that the gap after the name actually landed. The same data drives captions; see captions from word timestamps.

Try it

Sign up at /signup for 60 free tokens, no card required, and paste a script into the Playground to hear how pauses and speed change the delivery. The docs list every field used above.

Keep reading