What should accompany an AI voiceover?
A finished narration needs a way to follow its meaning without hearing it. For a video, add synchronized captions for speech and important sounds. A transcript provides a separate readable version; a descriptive transcript also explains essential visuals. For audio on its own, publish a transcript next to the recording.
This guide starts after you have an approved recording from an AI voice tool or a video editor. It covers delivery, rather than how to generate a voice. Use the Gemini voice guide for recording and the Flow guide for creating clips. The same checks apply to a human narration.
Match each version to the information it must carry
Captions follow the audio timeline. Include a speaker label when the speaker would otherwise be unclear, and identify a meaningful sound when it changes the scene. Translated dialogue alone may leave out those details. Closed captions can be switched on or off; open captions are part of the picture.
A transcript should remain understandable as a document, without constant playback. Give it a clear title and logical paragraphs. Remove caption timing markers if they only interrupt reading. Keep relevant speaker names and sounds, and put the transcript near the media with an obvious label.
Visual description solves a different problem. If the lesson says “choose this one” while pointing at a button, the audio omits an instruction. Name the button in the narration, or provide an appropriate described version. Captions that repeat “choose this one” cannot supply the missing visual information.
Prepare captions from the final audio
Freeze the spoken content before timing captions. Check the actual export against the approved facts; a speech generator may skip a word or pronounce a number incorrectly. If the recording is wrong, repair the recording as well as the captions. A corrected text track should not silently contradict the voice.
YouTube warns that automatic captions can mishear speech because of factors such as pronunciation, background noise and overlapping speakers. In YouTube Studio, open Subtitles, select the video and review the relevant language track. Correct the words and timing while listening to the final audio. Check the published track afterward, not only the editing preview.
If you already have reviewed text, YouTube also supports caption-file uploads, manual entry and an auto-sync workflow. Use a supported format and language. Auto-sync does not remove the need to check timing, and a translation needs a separate meaning check. Never assume captions are ready just because the upload finished.
Worked example: three errors in a twelve-second lesson
Imagine a fictional software tutorial. From 0–4 seconds, the narrator says “Open Settings, then Export.” From 4–8 seconds, the narrator says “Choose CSV, then Download.” From 8–12 seconds, a completion chime sounds and the screen shows “Export ready: 12 rows.” These are planning intervals for this invented example, not measured caption timings.
The automatic draft reads “Open settings then expert” and “Choose see us V then download.” It also omits the chime. Correct the two phrases to match the recording, and add a brief sound cue if the chime communicates completion. Adjust cue boundaries to the spoken delivery rather than blindly using the four-second planning blocks.
The screen's twelve-row result is still missing from the audio. A useful revision is to narrate “The export is ready: twelve rows,” leaving enough time for it to be understood. Re-render that ending, then regenerate or retime the affected captions. The descriptive transcript should also state the completion message. Do not insert an unspoken result into a dialogue caption and pretend the narrator said it.
The acceptance question is concrete: can someone explain which menu to open, which format to choose and what confirms completion using the alternatives provided? A transcript can answer that question even when the video cannot be played.
Use AI for a draft, then compare it with the recording
An AI tool can organize a rough transcript, flag unfamiliar terms or suggest where a sentence might be divided. It cannot establish accurate timings from a script alone. If the tool cannot access the audio or visuals, tell it exactly what evidence it has and ask it to mark unresolved items.
Use a fictional practice clip before uploading a customer's recording. Remove unnecessary personal information and confirm permission to use the material. Keep the final audio, caption file and transcript together with a version label so a later voice edit does not leave an obsolete text track.
PROMPT TO ADAPT
Review this rough transcript against the supplied final recording. Preserve the spoken meaning. Flag uncertain words, names, numbers and speaker changes; do not guess. Note meaningful sounds. Suggest readable caption divisions, but do not invent timestamps if you cannot measure the audio. Separately list essential on-screen instructions absent from the narration. Return a correction list and the evidence needed to resolve each uncertainty.
Check the player and the phone-sized export
The media player must let readers reach and operate playback and caption controls with a keyboard. Check visible focus, understandable control labels and the ability to read controls when enlarged. W3C recommends choosing a player with accessibility support instead of casually building a custom one.
Preview the actual captioned export on a narrow screen. Ensure text is legible and does not cover the part of the demonstration it explains. Compare the beginning, middle and ending for drift after edits. With a separate caption track, test the intended language and the on/off control at the publication destination.
Keep a readable transcript available without requiring the video download. For a website, use deliberate playback and a lightweight poster rather than starting large media automatically. Test the delivered page as a visitor, because an editor preview cannot prove that the published track loaded.
Five-minute practice: repair the missing instruction
Take the invented twelve-second lesson above. Write three columns: what is heard, what is visible, and what a reader needs to know. Mark “Export ready: 12 rows” as visual-only information. Then write one narration revision and one descriptive-transcript sentence that convey that result.
Answer: the instruction sequence is Settings → Export → CSV → Download, followed by an explicit twelve-row completion result. Correct “expert” and “see us V” using the recording, not by changing the intended task. After revising the ending, the caption timing needs another check.
Finally, cover the video and read your transcript. Then play the lesson muted with captions. Record any unanswered question in each mode. These quick exercises reveal missing information; they do not certify that the media meets every accessibility requirement.
- Names, figures and negations match the final approved audio.
- Essential visual instructions have an appropriate description.
- Caption timing was checked after the final edit.
- The transcript is easy to find and read without playback.
- The published player works with the keyboard and at a narrow width.
Sources & further reading
- W3C WAI: captions and subtitles
- W3C WAI: transcripts
- W3C WAI: describing visual information
- YouTube Help: review automatic captions
- YouTube Help: add subtitles and captions
- W3C WAI: accessible media players
Educational content. Examples are illustrative. Protocols, products and rules can change; consult the linked documentation for current details.