AI video captions: change the style, verify the sound stays the same

A caption revision sounds harmless: keep the video, change how the words look. But a useful editing workflow should let you check both halves of that request.

We ran a small, reproducible example through Valmera's MCP connector: select a dialogue excerpt, add sentence-style captions, then revise only the caption treatment. The two previews used the same source range and had identical decoded audio samples. The full-resolution export added branding and was longer than the edited timeline.

This article shows the result and a method you can reuse. It is a first-party Valmera demonstration recorded on 7 November 2026, not an independent editor comparison or a general caption-accuracy benchmark.

The result: one excerpt, two caption treatments

Both frames below come from the downloaded 854×480 previews at program second 0.4.

Documentary preset — the question appears as a two-line phrase:

Clean preset — the opening appears as a short word group:

Footage: © copyright Blender Foundation | durian.blender.org, used under CC BY 3.0. We selected an excerpt of the Sintel trailer, added captions and transcoded it; the final export also has Valmera branding. Blender Foundation does not endorse Valmera. Attribution and modification details.

Use GitHub's download control if the file page does not play the MP4. The two previews are historical review drafts; the final export is a separate file.

Make the revision specific

Caption wording, timing, grouping and appearance are different decisions. A request to correct a name should not silently become a new cut. A request for shorter groups should not automatically become a request to remove pauses.

For a presentation change, start with this brief and replace the placeholders:

In project [ID], revise the captions in approved version [version]. Use [style] with [desired grouping]. Preserve the source selection, order, speed, framing and original sound. Show the same passage before and after, identify the new version, and report any other change.

For this run, the first caption preset was documentary; the revision used clean. Reading the saved edit back showed that the revised preset grouped captions into at most two words.

The screenshots let you compare that presentation change. They do not establish that either style is ideal for your footage. At the intended viewing size, inspect the fastest phrase, a face near the caption band and a busy background. Shorter groups reduce the words visible at once, but also change how often a viewer must follow a new group.

Read the actual cut back

The requested source selection was 12–21.5 seconds, with word snapping enabled. The tool returned 11.9–21.5 seconds after moving the opening boundary to a quieter point.

That is a 9.6-second excerpt. The distinction between requested and saved boundaries matters: your caption comparison must use the actual edit.

For this single, normal-speed section:

program time = source time − 11.9

The transcript's question at source 12.00–14.40 maps to program 0.10–2.50. The answer at source 18.45–19.89 maps to program 6.55–7.99.

Those are transcript-derived speech windows, not a separate subtitle file. Caption grouping can hold words differently. An edit with several retained sections needs a section-by-section map; see source and output time.

Check the unchanged part as carefully as the changed part

The recorded revision called add_captions with the clean preset, rendered a complete preview and read the EDL back. The saved selection stayed at [11.9, 21.5], with no added music, sound effects, voiceover or speed changes.

Both preview files had 230 video frames at 24 fps and a 9.600-second container duration. To compare their audio independently of the MP4 container, we decoded the first audio stream to the same PCM format and hashed it.

With FFmpeg installed, these read-only commands reproduce that check on the downloaded previews:

ffmpeg -v error -threads 1 -i documentary-v3-preview.mp4 \
  -map 0:a:0 -c:a pcm_s16le -f hash -hash sha256 -

ffmpeg -v error -threads 1 -i clean-v4-preview.mp4 \
  -map 0:a:0 -c:a pcm_s16le -f hash -hash sha256 -

Both returned:

SHA256=6c92f86529b67b6bb9548e9da0ff4dde5cd299a8404d1bd1f2e839343cf78217

What this proves: the decoded audio samples match between these two preview files under the same decoding procedure.

What it does not prove: source-to-output audio fidelity, transcription accuracy, intelligibility, or a match with the longer final export. If hashes differ in another run, investigate the actual difference; a mismatch alone does not tell you whether the sound is perceptually worse.

This is a useful pattern for AI editing integrations: specify the invariant, save the two versions and measure the relevant output. Do not treat the assistant's statement that it preserved something as the entire check.

Treat export as its own result

The final MP4 was 1920×1080 H.264 with stereo AAC at 48 kHz. It contained 351 video frames at 24 fps and lasted 14.625 seconds, including this run's appended end card. Its opening sample also showed a corner mark.

The 9.6-second edited timeline and 14.625-second file are therefore different measurements. Check the downloaded delivery rather than assuming the preview's duration and appearance describe it completely.

The final file's SHA-256 matched the product's export receipt. All three outputs decoded without FFmpeg reporting an error. The evidence record and artifact hashes identify the files.

Keep the limits visible

The source fades toward black at the end of the excerpt. The preview reported a 0.08-second dark-ending warning; the final export reported about 0.17 seconds. We retained that ending to compare captions on the same selection. It was not repaired.

A separate listening-review operation was unavailable in live dispatch and was recorded for engineering follow-up. No listening-model review succeeded. The audio findings here come from transcript timing, saved edit state and deterministic file measurements.

This demonstration does not test overlapping speakers, multilingual captions, filler-word removal or long-form drift. It demonstrates a focused style revision with inspectable output.

Try it on a passage you know

Choose a short section whose words and intended boundaries you can verify. Approve the cut, save its preview, request one caption change and compare the same moments. Finally, review the exact exported file at normal playback speed.

You can connect a compatible assistant to Valmera or open Studio. Account creation and uploads are free; AI editing requires an eligible subscription. See current plans.

The complete recorded method includes the tool arguments and version history. For captions that are late, drifting or wrong after a cut, use the caption-revision guide.