Good explainer voice over sounds like a colleague explaining what's on screen: complete sentences, plain words, one short paragraph per scene, read at about 1.9 words per second. It gets written last, after the visuals are locked, because a line written to a finished picture can land its key word on the key visual moment, and words written first never sync.
The running example is a 60-second explainer for an AI support assistant: a customer question arrives, the assistant searches the help center while a retrieval panel fills with the three pages it pulled, and the answer streams out with citations. Every rejected narration draft in our production record failed in one of two ways: performing the topic like a trailer, or compressing sentences into fragments.
The voice is subordinate to the picture
The authority rule from our handoff sheets, verbatim: "Visuals are authoritative: re-time VO to the picture, not the reverse." The picture shows state, and the narration says what the state means — the why, the concept names, and only what the picture can't show.
If a line could play over any video, it isn't doing its job. In the assistant video, "it searched the help center before it answered, and these are the three pages it pulled" only makes sense over the frame where the retrieval panel fills, and that anchoring is what makes it narration instead of copywriting.
The same rule is why you never bolt new narration onto an animation timed to different words: it won't sync, and it reads as floating. When the picture changes, the words get rewritten to it (the full authoring order).
The register that survives review
The grading bar, in the words of the director who graded our scripts: full sentences that are "compact and dense in meaning — they fit in the allotted time and don't run on — but easy to understand."
- Write complete sentences with connective tissue. "Every time you add a page to the help center, the assistant keeps itself current by re-indexing it" survives review. "Pages in. Index synced." does not.
- Explain the screen, including the interaction the viewer is watching: "selecting a citation opens the exact passage the answer came from."
- Name concepts in a fixed cadence of term, role, then derived term: "Behind the assistant is a knowledge base. Inside it, the key unit of information is the document, and a document gets processed into smaller pieces called chunks."
- Make the viewer the subject where you can. "You probably already have a help center."
- Walk the naive path first when it helps. "The naive solution is to paste the whole help center into the prompt, but you'd hit the context limit and answers degrade."
- Calibrate every claim. If you can't support "more informed," write "informed"; one of our gold scripts makes exactly that downgrade.
- End on behavior or payoff, never a slogan. "When the assistant gets an answer wrong, the citation trail is the first place to look."
The final gate is the read-aloud test: any line you'd be embarrassed to say to a colleague across a desk gets rewritten.
Failure mode 1: trailer voice
The first draft of one of our videos came back with a one-line rejection: we are not doing marketing, we are doing explainers. The draft had performed the topic instead of explaining it — scene-setting, suspense beats, punchlines where facts should have been. The rejected lines and their shipped rewrites are quoted in full in the hooks guide.
Applied to the assistant video's retrieval scene:
- ❌ "A question lands. Somewhere in a thousand pages, the answer is already waiting."
- ✅ "When a question comes in, the assistant searches the help center first. Here it finds the three pages that mention refunds."
The bad version is fiction with a suspense beat, and it never states the lesson. The same trade shows up at the payoff scene:
- ❌ "No tickets. No queue. Just answers."
- ✅ "The customer gets a sourced answer in a few seconds, and nothing lands in your ticket queue."
Founder-coined aphorisms get the same treatment: we built a film around them once, and spoken aloud they read as grandiose even though the founder wrote them. Coined phrases appear only when the founder supplies them for that specific video, verbatim.
Failure mode 2: amputated fragments
The opposite drift comes from over-applying "brief and condensed" until the prose stops being speech. From a produced video about loops:
- ❌ "For each item of this collection. Three items — three passes. One. Two. Three."
- ✅ "The loop is configured with a collection to iterate over — three items here, so everything inside will run three times."
"For each item of this collection." is a fragment pretending the viewer heard its first half, and "One. Two. Three." narrates a count the picture already shows. Counting beats belong to the visuals; the voice says what the count means.
- ❌ "The loop exits once. Loop dot results: every pass's output, one array."
- ✅ "After the last pass, the loop exits once. Downstream, loop dot results is an array holding every pass's output, in order."
A label with a colon is a caption. The rewrite is a sentence with a subject and a verb, and it keeps "in order," which is half the point of a loop.
One test separates the two modes: condensing removes fluff and keeps grammar, while compressing removes grammar and keeps keywords. When a scene runs long, trim fluff from its sentence. Amputating it into "Three sources. Ranked. Cited." is never the fix.
The mechanics of a clean synthesized read
Timed against their scripts, our production reads come out at about 1.9 words per second, recorded or synthesized. Synthesis changes the workflow enough to need its own rules:
- Pace at ~1.9 words per second. A 9-second scene holds one short sentence of 15–18 words; a 12-second scene holds two.
- Leave a breath. About 0.7 seconds of silence follows each scene's line; the pad keeps the read from feeling wall-to-wall.
- The voice may stretch a scene, never squeeze it. A scene's final length is the larger of its visual minimum and the audio plus the breath pad.
- Key each paragraph to the scene's exact name. A typo'd scene name silently gets no audio.
- Convert dashes to commas in the speakable copy. A comma produces the pause you want.
- Keep code tokens out of the read. The voice says "loop dot results" and never meets backticks.
- Synthesize per scene, and keep the takes. A one-line fix in scene 4 then never touches the approved reads of scenes 1–3, and since synthesis isn't deterministic, the audio you approved is the asset.
- Silence is legal. A scene with no narration keeps its authored duration.
Done this way, a narration fix is a sixty-second loop: edit the prose, re-synthesize the one changed scene, re-time, done. Words stay the cheapest thing to change in the whole production, which is exactly why they're written last.
FAQ
Synthesized voice or human recording? Both ship, and the question is taste rather than technology: a flat human read loses to a well-directed synthetic one, and vice versa. If a founder records their own, hand them a per-scene sheet with the timing and what's on screen at the key moments.
How many words fit in a 60-second video? Around 100–110 spoken words: 60 seconds at 1.9 words per second, minus a breath pad per scene. Explainer video length covers how runtime is chosen.
What if the narration runs longer than the animation? Let the scene extend, so the extra time holds the scene's settled frame. Never compress the picture below what it needs, and never amputate the sentence into fragments to force a fit.
Can I write the voice over first and animate to it? It's the most common order and the most common mistake. Visuals animated to fixed words float, because the key visual never lands on the key word. Lock the picture, then write to it.
To see this on your own product, send us the URL. Twenty short candidate videos come back, and you pick the ones worth finishing.