AI subtitle timing check: inspect five cues before acceptance

An AI subtitle timing check can flag impossible intervals, overlapping cues and wording that differs from an approved reference. In this fifteen-minute exercise, you will review five supplied cues and save a correction sheet that another editor can reopen. A text-only review cannot establish that captions are synchronized with the recording.

Think of this as inspecting a delivery before accepting it. You already have subtitles; you are not asking a model to transcribe a recording, translate dialogue or generate a new video. A cue is one caption block with a start time, an end time and text. The useful output is a small list of decisions, not a confident statement that the whole video is ready.

Content updated: 15 October 2026. All names, words and times in the exercise are synthetic teaching material created for this article. They are not a customer project or measured model performance. The fifteen-minute target assumes an available Gemini account and a working text editor; account setup and final video editing are outside that estimate.

What can a text-only check actually prove?

Separate three checks before opening a tool. First, does each interval end after it begins? Second, do neighboring intervals overlap? Third, does each cue preserve the approved spoken reference? A text-only check cannot establish whether the speaker really begins at the listed time.

Artificial intelligence describes systems that perform a range of information-processing tasks. Here we use a conversational text assistant to inspect written inputs. We are not claiming that the assistant has heard an audio track. If your reference has not been checked against the recording, resolve that first with the workflow for transcribing audio to text.

Overlap needs context. Two speakers can appear simultaneously in a deliberate caption design. For this exercise, the declared rule is one speaker and one sequential caption stream. That makes an overlap a review item rather than an automatic instruction to move every later cue. Write your own stream assumption in a real handoff.

YouTube's help documentation explains that automatic captions can contain mistakes and should be reviewed. That supports checking the output, not assuming a particular error rate. W3C's caption guidance also covers non-speech sounds needed to understand the content. Neither source makes a five-row text inspection a complete accessibility evaluation or a guarantee of synchronization.

Prepare a five-cue inspection sample

Use the following data unchanged on your first attempt. Times use hours:minutes:seconds,milliseconds. The notation 00:00:02,500 therefore means two and a half seconds. The C labels are inspection identifiers, not a promise that this table can be imported into a particular subtitle format.

Approved reference: C1: Place the empty jar on the table. C2: Open the lid slowly. C3: Add one spoonful. C4: Then stir. C5: [timer sounds] Ready. For this exercise only, the reference is declared authoritative. There is no supplied recording, so you cannot independently verify what was said or heard.

The timer note matters because, in this fictional instruction, it signals completion. It is not spoken dialogue. In a real recording, identify sounds that carry meaning rather than writing every background noise indiscriminately. Someone watching without sound should not lose a necessary instruction simply because it was conveyed by a sound effect.

Create two labeled input blocks: DELIVERED SUBTITLES and APPROVED REFERENCE. Keep them separate so the assistant cannot silently treat the draft as the source of truth. For a real job, include the reference version and its human approver. If the owner cannot tell you which version is current, pause the comparison and resolve that question.

Use this synthetic exercise before pasting business material into an external service. Do not include private recordings or customer details merely to test the prompt. For a wider introduction to the technology, use the site's artificial intelligence guide; the task here remains a narrow inspection of text you are permitted to use.

Run the check without inventing replacement times

Open gemini.google.com in your browser. If necessary, choose Sign in and use a Google account that you are permitted to use. Start a new conversation and paste the instruction below into the message box. If access is unavailable in your account, stop here rather than assuming you must buy a plan or upload the recording.

Replace [DELIVERED SUBTITLES] with every C label, start time, end time and delivered sentence from the table. Replace [APPROVED REFERENCE] with the complete reference paragraph, including the timer note. Remove the two placeholder labels inside square brackets.

Inspect this synthetic single-speaker, sequential caption stream. No audio or video is supplied. DELIVERED SUBTITLES: [DELIVERED SUBTITLES]. APPROVED REFERENCE: [APPROVED REFERENCE]. For each C identifier, check that the end is later than the start, inspect neighboring intervals for overlap, and compare wording with the reference. Show the time arithmetic. Identify any missing meaningful non-speech cue separately. Return Cue | Evidence | Issue | Next action. Do not invent corrected timestamps or silently rewrite the captions. Mark synchronization and whether sounds are actually present as “requires video review.”

Read the answer as a proposed inspection sheet. Each finding should name a cue and point to supplied evidence. “Improve accessibility” is not a usable correction note. “C5 lacks the reference's timer cue; confirm the sound and placement in playback” gives the editor a specific next action without pretending the model heard anything.

If the answer supplies a repaired timestamp for C2, reject that part. The written inputs prove that its duration is impossible, but they do not reveal the correct replacement. Ask the editor to inspect the recording.

Compare the answer with the expected findings

C1 lasts two seconds and matches its reference wording. Record “no discrepancy in supplied text; synchronization unverified.” Do not shorten that to “passed.” A valid interval and matching sentence do not show whether a viewer can read the line or whether the sentence occurs at that point in the recording.

C2 has a negative duration: 2.500 minus 3.000 equals minus 0.500 seconds. This interval cannot stand as written. Flag it for correction using playback, without selecting a new end time. There is no evidence that all subsequent captions should shift by half a second; one bad interval does not establish a global offset.

C3 ends at 6.000 seconds, while C4 begins at 5.500. Their intervals overlap for 0.500 seconds. Under the declared single-stream assumption, that requires an editing decision. Separately, C3 says “two spoonfuls” while the approved reference says “one spoonful.” That is a quantity change, not a harmless stylistic alternative.

C4's words match the reference, but its overlap with C3 remains open. C5 omits the timer sound cue from the declared reference. Restore the meaning only after checking how that event should appear in the actual video. The exercise does not determine an exact start time for the sound label.

C4 ends exactly when C5 begins, at 8.000 seconds. Those intervals touch at the boundary; there is no positive-duration overlap to report. Playback may still reveal a reading or presentation issue, but do not invent a mathematical error to make the inspection look more thorough.

Compare these findings manually with the model's answer. If it misses the quantity change, add the reference quote beside the draft quote. If it calls the touching boundary an overlap, show the zero-length intersection. Keep correction notes factual. For earlier script development, YouTube writing prompts are a separate workflow, not a reason to rewrite an approved instruction during acceptance.

Save a handoff another person can reopen

On Windows, open Notepad from the Start menu. Paste the supplied inputs and your checked inspection table into a new document. At the top, write “Text inspection complete; video review pending.” Include the stream assumption and reference version.

Choose File → Save as, select Documents, and save the file as subtitle-check-01.txt. If an encoding option is displayed, retain UTF-8. Close the document. Open File Explorer, go to Documents, and double-click subtitle-check-01.txt. This is the acceptance step for the exercise: verify the evidence survived saving rather than assuming the chat history is your handoff.

In the reopened document, locate C2's negative duration, the C3/C4 overlap, the one-versus-two wording difference and the missing timer cue. Your practical result is a durable correction sheet with four distinct review items, not a claim that every line requires a change.

Next, the person editing the video should play the relevant section with sound, inspect caption entry and exit, and check readability on the intended viewing device. Confirm that text does not cover necessary visual information. If video assembly is a separate assignment, the guide to freelance video editing can help define that handoff. Do not treat the text file as an import-ready caption master.

In my earlier writing about choosing effectiveness over perfection, I described how chasing every detail could delay useful work. The application here is deliberately modest: finish one verifiable correction sheet instead of requesting an undefined “better video.” This is a working principle drawn from that writing, not a measured subtitle case study or a promise of time saved for a client.

If another language track is involved, keep its meaning review separate from timestamp inspection. The guide to AI video dubbing covers that different output. End this exercise when the reopened file is complete; release the video only after a human has checked the actual recording and the final captions together.