Should the agent write and voice in one step?

No. Separate them and persist the script between. Combining them means you cannot re-voice without re-writing, and your captions will drift out of sync with the audio.

Is synthetic narration acceptable to audiences?

For explainer and information formats, generally yes. For anything where personality carries the piece, a real voice still wins clearly.

How do I get captions that match?

Transcribe the generated audio rather than captioning the script. The read never matches the written text exactly, and viewers notice.

Video agent use cases

Can an AI agent write and voice a narration track?

Yes Reliable to automate today. Last checked 2026-10-14

Yes. Script-to-voiceover is the most dependable stage in any AI video pipeline — the text is cheap, the audio is fast, and the output is deterministic enough to build an edit around. The failure mode is not quality, it is pacing: synthetic reads do not breathe where a human would.

What you are actually trying to do

You want narration for an explainer, a faceless channel or a documentary-style piece, generated from a topic or an outline without recording anything.

Done properly, narration becomes the clock the whole edit runs on, which removes an entire class of sync problems.

What works today

Script generation against a format brief is the single most reliable thing a text model does. Give it structure — hook, beats, word budget — and it hits it.

Voice synthesis is past the uncanny line for narration. Most viewers do not identify it as synthetic in a narration context, though they often do in dialogue.

Per-line audio with returned durations is the key capability. Generate one file per line, measure each, and build the timeline from real numbers instead of estimates.

Multi-language dubbing from one script is now routine and genuinely useful for reach.

Where it breaks

Pacing is wrong by default — rushed through the point, pausing in the wrong places. Punctuation, line breaks and explicit break tags fix most of it: AI voiceover pauses and pacing are wrong.

Numbers, acronyms and units get mangled. Write them as they should be spoken in the script — 'twenty twenty six', 'A P I' — rather than hoping the model normalises them your way.

Long single-block generations drift in tone and energy and give you nothing to re-cut. Generate per line or per paragraph.

Loudness varies between lines, which is audible on a cut. Normalise the whole track once at the end.

The agent silently rewrites the script between stages if script and voice generation are not separated, so the audio no longer matches your captions.

How to wire it up

Have the model write to an explicit structure and word budget. Vague briefs produce 400 words where you wanted 150.

Format: 150-170 words. Hook in the first sentence.
Three beats. One line per beat, max 25 words.
Write numbers and acronyms as they should be SPOKEN.
Return JSON: [{"id":1,"line":"..."}]

Freeze the script before synthesis. Write it to disk, and have the voice stage read that file — never a freshly generated version.
Synthesise one audio file per line and record the real duration of each. These durations are now your timeline.
Insert pauses deliberately with punctuation or break tags where the meaning needs air, rather than accepting the default read.
Concatenate and normalise loudness across the whole track in one pass.
```
ffmpeg -i vo_raw.wav -af loudnorm=I=-16:TP=-1.5:LRA=11 vo.wav
```
QC the result: total duration within tolerance, no silent file, peak level sane. A zero-byte or silent voice track is the most common way a pipeline publishes a video nobody can hear.

What to use

Any text model — script. Cheap, reliable, the easiest stage to get right.
ElevenLabs and peers — narration with per-request audio you can measure.
ffmpeg — concatenation, loudness normalisation, duration probing.
A forced-aligner or word-level transcriber — to get caption timings back from the audio you just generated.