Diagnose fixed offsets, gradual timing drift and edit-related jumps in a transcript, then choose a reversible correction checked against several audio anchors.
Direct answer: If the mismatch grows gradually, the transcript and media probably do not share the same timing scale; a single time shift will not repair the whole recording. If the mismatch stays constant, inspect a fixed offset, and if it jumps suddenly, look for an edit or a different source version. Measure several matching passages before changing anything, keeping the original media and timing file intact.
Moving every timestamp back two seconds can make the first sentence look correct while leaving the final section badly misaligned. That is not evidence that you need to repeat the same adjustment. It means you have not yet identified the shape of the error.
The useful diagnostic compares where a passage actually occurs with where the transcript says it occurs. Work from matching content, not from file duration alone.
Run the timing drift diagnosis
The timing drift diagnosis is an editorial troubleshooting method for distinguishing a fixed displacement, a changing scale and a discontinuity. It helps you choose the type of correction; it does not identify every possible encoding or transcription fault automatically.
Before starting, keep an unchanged copy of the recording and transcript or caption file. Record their filenames, versions and durations. Work on copies in an authorised location, particularly for private interviews. Do not upload sensitive recordings to a new synchronisation service merely to test whether it can fix the problem.
If the transcript has no timed entries, this is a different task: you need to establish timing rather than repair it. If it is linked to a different edit, obtain the matching source if possible before attempting mathematical corrections.
Accurate timing is one requirement of a useful searchable media archive, but the evidence remains the recording. A timecode is a locator, not proof that the text or speaker label is correct.
Pick content anchors across the recording
Choose clear spoken phrases or visible events near the beginning, middle and end. Add anchors immediately before and after any suspected edit. Avoid vague moments such as a long pause whose boundaries are difficult to identify.
For each anchor, record the actual media time and the corresponding transcript time. Define the sign of your error consistently. In this article, error equals transcript timestamp minus actual media time, so a positive value means the text is scheduled too late.
Listen or watch the matching content at normal playback speed. Do not compare a displayed clock in one app with an unrelated wall-clock time in another. If the player displays several time systems, establish which one represents elapsed media time before collecting values.
Success means you have several pairs describing the same passages. If you cannot match the wording, inspect whether the transcript was made from a different recording or contains substantial text errors. Retiming cannot repair a transcript of the wrong material.
Read the shape of the mismatch
| Observed pattern | Likely class of problem | First useful action |
|---|---|---|
| Similar error at every anchor | Fixed offset | Check for added introduction or leading silence |
| Error grows steadily across the file | Timing-scale difference | Compare matching durations and test proportional correction |
| Error changes abruptly after one point | Edit or discontinuity | Locate the cut and compare source versions |
| Error varies irregularly | Local alignment errors or several causes | Inspect affected sections individually |
These are diagnostic hypotheses, not guaranteed causes. A smooth-looking difference at three anchors may conceal an edit between them. Add more anchors when the proposed correction would affect a long or consequential recording.
Do not assume that two files with identical titles contain identical media. A trimmed introduction, removed pause or replacement section can change the relationship between time and words. Compare the actual content around the first point where alignment fails.
My recommendation is to reject a global correction unless several distributed anchors support it. A local repair takes more work, but it is better than making previously correct sections wrong in pursuit of a tidy one-click fix.
Calculate a gradual drift example
Consider a fictional 30-minute recording with a marker at its start. The marker occurs at media time 0 seconds but its transcript cue begins at 2 seconds. At media time 1,800 seconds, the matching cue begins at 1,832 seconds. These are illustrative observations, not a measured product failure.
The starting error is 2 − 0 = 2 seconds. The ending error is 1,832 − 1,800 = 32 seconds. Error has increased by 32 − 2 = 30 seconds over 1,800 seconds of media.
The added error rate is 30 ÷ 1,800 = 0.0166667 seconds per second, approximately one extra second per minute. Under a strictly linear assumption, transcript time equals media time multiplied by 1.0166667, plus two seconds.
To recover media time from a transcript timestamp, subtract two seconds and divide by 1.0166667. For the final cue: (1,832 − 2) ÷ 1.0166667 is approximately 1,800 seconds. Simply subtracting two would leave it at 1,830 seconds, still 30 seconds late.
A midpoint check is essential. Under this model, media time 900 seconds should correspond to transcript time 900 × 1.0166667 + 2, approximately 917 seconds. If the observed midpoint is far from that, the simple model is not supported.
This calculation illustrates the correction type; it is not an instruction to apply that ratio to your file. Measure your own anchors, include both start and end times for caption cues, and verify the resulting playback before replacing any distributed version.
Apply the smallest reversible correction
Use the documented retiming feature in an editor compatible with your file format, or ask an experienced editor to apply the measured mapping. This article does not prescribe menu paths for an unspecified app. Preserve the original and export the result under a new name.
Timed text has structural requirements as well as numerical values. WebVTT, for example, defines cue start and end offsets and requires an end later than its start. Correcting timestamps should preserve valid timing structure, not merely move visible text until one sentence looks right. W3C WebVTT specification.
For a fixed offset, check the first cue after shifting so it does not land before the recording begins. For a scale change, check cue durations as well as starts. For an edit-related jump, use separate sections or obtain a matching transcript rather than forcing one ratio across the whole file.
If you are unsure which file format or timing system you have, inspect its documented format and the editor's supported import types. Do not repeatedly save through incompatible tools while hoping the numbers will settle.
Verify the corrected file in the destination player
Open the exported file with the intended media, not only inside the editing view. Check the original anchors and several passages between them. Watch for overlaps, missing cues, truncated words and a late final section.
Keep a correction note stating the source versions, measured errors, applied adjustment and remaining limitations. If another person will use the transcript, identify which file is current. A successful import is not enough; the words must appear with the right content.
Diagnose before spending another hour retiming
- Spend five minutes preserving files and confirming the source versions.
- Use ten minutes to measure at least three distributed anchors and any suspected edit boundary.
- Spend ten minutes classifying the pattern and testing the proposed correction on a copy.
- Reserve ten minutes for destination playback checks before accepting the result.
Stop global adjustments when the pattern is inconsistent or the source versions differ. Obtain the matching recording, repair sections individually or get experienced help. Repeating a convenient correction is not progress when it keeps moving the error elsewhere.
Related guides
Frequently asked questions
Why does shifting every timestamp fix the start but not the end?
Because the mismatch may change over time rather than remain constant. A fixed shift changes every timestamp by the same amount, so it cannot remove an error that grows through the recording. Measure matching passages near the start, middle and end and compare their errors. If the growth is reasonably proportional, a timing-scale correction may be appropriate; if it jumps at an edit, use a different approach. Do not infer a scale difference from the first and last cues alone when the recording may contain cuts. Check intermediate evidence before applying a global adjustment.
Can a transcript from an earlier edit be repaired automatically?
Sometimes, but only after establishing how the edits changed the relationship between text and media. A single added introduction may create a fixed offset, while several cuts or rearranged sections require more local work. Automatic alignment may offer a candidate result, but inspect its output against the recording and preserve the original. Do not assume the text remains complete if words were added or removed. If you can obtain the transcript made from the exact final version, that may be the simpler and safer starting point. The correct choice depends on the edit history, not merely the file names.
Does changing playback speed cause permanent timestamp drift?
Do not assume that it does. Distinguish a player's temporary playback control from a recording that has actually been exported with altered timing. Compare elapsed media positions at normal speed and inspect the source versions before diagnosing a permanent change. If the exported media has a different duration for the same uninterrupted content, a timing-scale difference is one possibility to investigate. This article does not claim identical behaviour across players or editors. Record the specific app, file format and export settings if you need further support, rather than treating every speed-related symptom as the same problem.
Should I correct only caption start times?
No, not without considering their end times and intended durations. A cue that begins correctly but ends too soon or remains on screen into the next speaker's turn can still be unusable. Apply the appropriate correction to the timed structure, then inspect the result in the destination player. Some formats permit overlaps for particular purposes, so do not remove all overlaps mechanically without understanding them. Preserve the original file and follow the editor's documented behaviour. The target is synchronised, readable text with valid timing, not a list of start numbers that happens to match a few anchors.
What if the timing is wrong only in a few places?
Investigate those sections locally rather than imposing a global shift or scale. Check for omitted words, merged transcript blocks, pauses or edits around each problem. Use clear anchors before and after the affected region to establish where alignment becomes correct again. A local repair should leave already verified sections unchanged. If many irregular errors appear, the transcript may need a fresh alignment pass or more extensive review. Keep the original and record which regions were corrected. Do not let one accurate section persuade you that an entire file has been verified, or one bad section justify changing everything.
When should I ask an experienced editor for help?
Ask when you cannot identify the timing system, the mismatch is irregular, the file contains several edits, or the transcript will support consequential publication or accessibility work. Provide the exact source versions, file formats and a few measured anchor pairs, with private material shared only through an authorised arrangement. That evidence makes the problem easier to diagnose than a vague report that the subtitles are wrong. You can still perform the initial comparison yourself. Stop experimenting when repeated saves risk damaging the only usable copy or when you cannot explain why the next correction should work.
Sources and verification
- W3C WebVTT specification, checked 11 September 2026 for cue timing structure and start/end requirements. No particular editor interface or product failure is asserted.
- The supplied parent was read locally because its public route was unavailable. The timing diagnosis and linear example are editorial reasoning using hypothetical observations.
This article is practical guidance. Apply it in proportion to your tools, evidence, risks, and responsibilities.
