Introduction
Start with three places, rather than the first wrong line
A subtitle appears two seconds before a speaker begins. Moving the track two seconds later seems to solve the problem, until the closing discussion is still several seconds wrong. That pattern calls for a timeline diagnosis before another adjustment. Choose a distinctive phrase near the start, another around the middle and another near the end. For each, record where the subtitle currently starts and where its corresponding speech belongs in the intended media. These are paired observations, not three unrelated times copied from different players. They reveal whether the discrepancy stays similar, grows with position or changes suddenly. The first result directs the next action: shift a shared offset, investigate a scale difference, or look for a version boundary. A single cue is useful for noticing trouble, but it does not tell you how the rest of a long recording behaves. Leave the original untouched while making this initial comparison.
Make the anchors recognizable enough that another editor can find them. A phrase such as “the next demonstration uses the blue sample” is usually easier to distinguish than a repeated “yes.” Note the surrounding sentence, selected audio language and media filename. Keep the definition of the matching event consistent: do not compare the first syllable at one anchor with the end of a sentence at another. Where a breath, soft consonant or background sound makes the start ambiguous, record that uncertainty instead of inventing precise milliseconds. Three anchors are an economical opening check, not a complete survey. Add observations around edits, interruptions and places where the mismatch seems to change. The objective is to learn enough about the relationship to choose a defensible correction, while leaving local wording and display decisions available for a later pass.
References: Aegisub: Timing subtitles to audio; Subtitle Edit: Point Sync.
Decide which clock and which recording the timestamps describe
A timestamp is a position on a particular timeline. It may describe an original session, a trimmed export, an uploaded video or a separate audio file. Matching names and approximately matching durations do not establish that two files share that timeline. A new opening slate creates a different relationship from a permanent speed change. An inserted explanation changes only the material after its insertion. Before adjusting anything, identify the media revision and the timing source. Ask whether the words themselves occur in the same order and whether the relevant recording contains the same introductions, pauses and closing material. A subtitle made for the full lecture cannot be assumed to fit a shortened teaching excerpt. If the pair is wrong, obtaining the matching media or captions can save more work than carefully repairing a relationship that was never intended.
Separate a recording's stored positions from a viewing adjustment. A player's temporary subtitle-delay setting may make one viewing session look correct while leaving the actual subtitle file unchanged. If that adjustment is already active, measuring new anchor times without noting it mixes two different states. Likewise, a streaming player and a local editor may be showing different editions. Establish a neutral comparison environment, document any intentional offset, and use ordinary playback to confirm what was measured after a seek. When the spoken audio itself is out of agreement with the picture, subtitle retiming is only one part of the problem. A text track can follow the speech while the media remains wrong. State which relationship you are fixing: captions to speech, speech to picture, or a source transcript to an edited recording. Each has a different acceptance test.
References: W3C: WebVTT cue timings; Aegisub: Working with Video.
1. Build a small timing record you can trust
1.1 Preserve both artifacts and label the direction of change
Keep an unmodified media original, an unmodified timing original and a clearly named working copy. Your timing record should identify a cue, its current start, the corresponding media event and the proposed correction. These fields prevent a common mistake: applying a number calculated in one direction as though it described the reverse direction. Suppose a subtitle starts at 60 seconds, while the matching speech begins at 62 seconds. Moving that subtitle later requires adding two seconds. If instead it starts at 62 and belongs at 60, subtract two. Write the sentence “current cue time becomes media time” beside the calculation. It is less elegant than an unlabeled number, but much harder to misunderstand during handoff. A clear filename should also reveal whether the output contains experimental timing or an accepted revision, so that a test export is not mistaken for the final caption track.
Record what remains fixed during the trial. If the media changes while you are measuring, the anchors no longer describe one problem. Save the exact export or otherwise identify its revision through your normal production process. Do not overwrite the source subtitle merely because the first experiment looks promising. A reversible repair lets you compare text, timing and cue count independently, return to the original when the model fails, and explain what changed. This matters particularly when one person prepares captions and another exports the video. An editor who later adds a title sequence has changed the timing relationship even if the spoken words are untouched. Agree on a delivery revision before the final timing pass, and treat a subsequent timeline edit as a reason to review the affected caption relationship again.
References: Aegisub: Shift times; Subtitle Edit: Visual Sync.
1.2 Choose anchors that measure the same audible event
Use distinctive wording, a reasonably clear speech onset, sufficient timeline separation, consistent audio selection and a recorded measurement uncertainty. A repeated greeting or a phrase obscured by overlapping speakers is a poor starting anchor. Find something that can be located without guessing which occurrence you heard. The subtitle start is not automatically an exact transcription of the first audible sample; a caption editor may deliberately allow a small lead or adjust a boundary for reading. When diagnosing a large mismatch, compare equivalent events and preserve the editorial allowance rather than treating every difference as a clock error. Near a long pause, inspect the actual onset of the next phrase instead of assuming that the largest waveform peak marks it. A loud chair movement can be easier to see than the speech you need to time.
Anchor spacing determines what you can learn. Two observations only a few seconds apart may show an offset while hiding a small discrepancy that becomes obvious over an hour. Use widely separated passages to examine a possible scale difference, then inspect additional material between them. Include a point immediately before and after a suspected insertion or missing section. If the words cannot be matched confidently, stop trying to improve the numerical fit. Check the selected language, edition and subtitle text first. An accurately measured time for the wrong utterance is still a bad anchor. Keep a short explanation of why each observation was chosen, especially if its boundary is approximate. That note helps the reviewer distinguish a meaningful residual from a difference caused by two people marking slightly different parts of the same word. If two candidate phrases have substantially different uncertainty, prefer the clearer event for estimating the relationship and retain the ambiguous one as a local review item. More decimal places in a note cannot compensate for not knowing which sound was measured.
References: Aegisub: Timing subtitles to audio; Subtitle Edit: Point Sync.
1.3 Describe the error pattern before naming its cause
Let c mean the current subtitle time and m the corresponding position in the final media. Define the observed correction as m minus c: positive means the cue must move later. Compare that value across the recording. A similar correction suggests an offset; a steadily changing correction suggests a scale component. A sudden change suggests a different local relationship. These are descriptions of the observations, not diagnoses of the software that produced them. A linear-looking pattern does not prove that a recorder's clock failed. Different exports, speed adjustments, mistaken assumptions about frame positions and other processing choices can produce relationships that deserve investigation. State what the data supports first. “The required correction grows between the first and last checked phrases” is useful even when the production history is incomplete.
Treat irregular residuals as a reason to inspect the material, not an invitation to keep adding arbitrary decimal places. They may reflect poor anchors, local caption decisions, a different cut, missing speech or several problems at once. A sparse record cannot distinguish every possibility. Add observations where the behavior changes and listen to the surrounding passage. If a phrase is absent from the media, there is no valid playback time to assign it in that edition. If the same phrase has been moved to a different location, preserving subtitle order may no longer be appropriate. These findings change the task from correcting a common timing relationship to rebuilding specific portions of the track. Recognizing that early is a practical result: it prevents a mathematically smooth adjustment from concealing an editorial mismatch.
References: Subtitle Edit: Visual Sync; Subtitle Edit: Point Sync.
2. Test an offset and a scale against the observations
2.1 See when a shared shift is enough
Consider an authored fixture with current starts at 60, 900, 1,800, 2,700 and 3,660 seconds. Every corresponding media event is exactly two seconds later. Here the constant correction is supported at all five observations. A whole-track shift of plus two seconds preserves the spacing and duration of each cue, provided both its start and end move together. The selected range and affected boundaries still matter. Shifting only starts changes durations; shifting only a subset leaves the rest on the old timeline. Aegisub's documentation distinguishes these controls, so check their meaning in the editor you use. Preview a few affected cues before committing the complete operation. A correct number applied to the wrong rows or to only one boundary produces a different result from the intended repair.
If you use WhisperAITranscribe, an online audio and video transcription tool, to prepare a source transcript, carry the media identity forward into this timing review. The presence of a readable transcript does not establish which later edit its time references describe. Use a supported timing artifact or review the relevant passages in an appropriate editor; a particular retiming control or export should not be assumed from the product name. Where a shared shift is supported, preserve the reviewed wording and change the time relationship deliberately. Where it is not supported, resist regenerating all the text as the first response. New recognition introduces another wording revision without necessarily resolving the mismatch between source and delivery. Identify the problem's layer before deciding whether you need text work, timing work or a different media pair.
References: Aegisub: Shift times; W3C: WebVTT cue timings.
2.2 Calculate a two-point relationship with the axes named
For a possible speed difference, use m = a × c + b as a candidate relationship. The scale a describes how spacing changes; the offset b describes the origin difference. From two distinct anchors, calculate a as the media-time difference divided by the cue-time difference, then obtain b from either pair. In the authored linear fixture, c values of 60 and 3,660 correspond to m values of 62.060 and 3,665.660 seconds. Their differences are 3,600 and 3,603.600 seconds, respectively. The resulting scale is 1.001 and the offset is two seconds. These values map existing subtitle positions to the target media. Reversing the direction requires a different expression; applying the same scale and offset again would move in the wrong direction. Keep the labeled anchor pairs beside the calculation so the transformation remains understandable after the immediate editing session. To map a target media time back to the original subtitle timeline, subtract b and divide by a. In this fixture, subtracting two from 62.060 and dividing by 1.001 returns 60. Applying the forward transformation to 62.060 instead answers a different question. That small reverse check is useful when a handoff supplies a speed ratio without explaining its direction.
Notice why the first cue alone gives an incomplete answer. Its required correction is 2.060 seconds, which includes the two-second origin difference and the scale effect accumulated by 60 seconds. Applying plus 2.060 everywhere makes that anchor exact. At c = 1,860, however, the intended media position is 1,863.860; the shifted cue reaches only 1,862.060, leaving it 1.800 seconds early. At c = 3,660 the same approach leaves it 3.600 seconds early. Repeating a manual shift after every viewing pass can therefore feel like chasing the error down the recording. The changing discrepancy is the clue that a shared offset is inadequate. Calculate a scale only when the observations support one, then test its prediction elsewhere before trusting the apparent elegance of the two-point answer.
References: Subtitle Edit: Visual Sync; FFmpeg: timestamp offset and scaling options.
2.3 Keep a middle passage out of the calculation
Two points define an affine relationship, meaning a scale plus an offset. They do not provide an independent validation of that relationship. Use another passage to calculate a residual: the predicted corrected cue time minus its observed media time. In the linear fixture, the midpoint at c = 1,860 maps to 1,863.860, matching the authored observation. A separate holdout at c = 960 maps to 962.960. These checks are exact because the fixture was constructed from the stated relationship; they demonstrate the calculation, not a measured transcription repair. In a real recording, the residual also reflects measurement uncertainty in locating speech and the existing caption boundary. Inspect whether remaining differences are compatible with those observations rather than demanding a fictitious precision that the source cannot support. Anchor separation also affects how measurement error influences the scale. If the cue positions are fixed and each measured media onset is uncertain by 0.100 seconds, their difference can be uncertain by 0.200 seconds. Across 3,600 seconds that corresponds to a scale uncertainty of one part in 18,000; across only sixty seconds it is one part in 300. This simplified bound explains why a short opening sample is weak evidence for an hour-long scale correction. It is not a universal timing tolerance or a statistical confidence interval.
Keep validation passages away from the two fitting anchors and spread them across the material. If every observation comes from the opening minute, they provide little evidence about the closing section. A successful midpoint does not prove the absence of a short insertion elsewhere, either. The useful question is whether the model continues to predict timings in the places that matter. Test both ordinary passages and known trouble spots. If an anchor fails, verify the phrase and media version before changing the formula. When the failure persists, retain the observation as evidence against the model instead of quietly replacing it with a more convenient point. A repair process should expose mismatches, not select observations until a preferred correction appears to work. That discipline is particularly valuable when the final track contains hundreds of otherwise plausible lines.
References: Subtitle Edit: Point Sync; Aegisub: Timing subtitles to audio.
3. Recognize patterns that one straight correction cannot explain
3.1 A jump after an insertion can imitate overall drift
The second authored fixture has current starts at 60, 900, 1,800, 2,700 and 3,660 seconds. The first two events need no correction; the remaining three occur twenty seconds later in the target media. The observed pattern is a step change, not a gradually accumulating difference. Nevertheless, fitting only the first and last anchors produces a straight relationship that matches both perfectly. Its scale is 181 divided by 180, with an offset of minus one third of a second. At the 1,800-second observation, that prediction misses the intended media time by about 10.333 seconds. The interior residual rejects the overall model even though its endpoints look excellent. This is why an editor should distrust a two-point success when the recording may contain added or removed sections. The middle observations distinguish explanations that the endpoints alone cannot separate.
The numbers do not reveal where an insertion begins, what it contains or whether it is actually an insertion. Those details require the production history or direct comparison of the editions. They do show that the correction should not be spread smoothly over every earlier cue. Find the boundary through nearby phrases, establish which material is shared, and divide the review accordingly. A cue entirely after a verified twenty-second insertion may need a corresponding offset while the earlier portion remains unchanged. A cue spanning the boundary deserves individual review, because its words may no longer occupy one continuous interval. If sections were reordered, an offset for “everything after this point” may also be wrong. Use a local mapping only after identifying its actual scope; do not turn a convenient pattern into an invented edit history.
References: Subtitle Edit: Point Sync; Subtitle Edit: Visual Sync.
3.2 Frame-rate labels require a question about elapsed time
An apparent rate mismatch is worth investigating, but different frame-rate labels do not by themselves prove that speech has changed speed. Compare elapsed event times. As a mathematical example, the same sequence of 86,400 frames lasts 3,600 seconds at 24 frames per second and 3,603.600 seconds at 24,000 divided by 1,001 frames per second. If the sequence is genuinely played at those different rates, positions measured on the faster timeline map to the slower one by a factor of 1.001. This gives a concrete source for a 3.600-second difference over that hour of the faster timeline. It is not a diagnosis of any uploaded video. Another conversion might duplicate or drop frames while preserving the intended elapsed duration, so its changed frame-rate label would not justify the same subtitle transformation.
FFmpeg's documentation distinguishes frame-rate modes and timestamp processing; that distinction explains why a blanket “change the subtitles to the video's fps” instruction is inadequate. A time-based subtitle specifies when text appears in the media's playback timeline. It should follow the actual timing relationship, not a ratio guessed from two metadata displays. Variable frame-rate material makes the shortcut especially risky, because an average rate is not a complete account of every frame's presentation time. Examine the source history and the same audible events before selecting a rate conversion. If the events already occur at the same elapsed positions, leave their times alone and investigate the playback or pairing issue. If they do not, establish the direction and magnitude from the observations, then check whether the documented export process explains them.
References: FFmpeg: frame-rate modes and timestamp handling; Aegisub: Variable framerate video.
3.3 Local caption choices can survive a correct global model
After a supported shift or scale, some cues may still need local alignment. A subtitle can start near the right moment yet end before the speaker completes the phrase. Another may contain an unrelated sound, combine two speakers or extend through a long silence. These are cue-level decisions, not necessarily evidence that the global relationship failed. Separate them during review. Inspect a passage's words and interval together, and distinguish a systematic timing discrepancy from a caption that was segmented poorly in the first place. The original track may already have local imperfections; transforming all its times faithfully carries those imperfections into the new timeline. Conversely, a deliberate reading allowance should not be erased just because it differs from an exact speech boundary. Decide what each interval is meant to communicate before making finer changes.
When the starting transcript comes from WhisperAITranscribe, keep this caption-editing pass separate from reviewing the recognized wording. A fluent sentence does not locate its own syllables, and a structurally valid interval does not establish that the sentence matches the audio. For an uncertain passage, return to the relevant media and correct the appropriate layer. If speech is missing from the transcript, retiming alone cannot restore it. If the words are correct and only their interval is wrong, preserve the wording while fixing that interval. Where the final recording contains additional meaningful sound, consider whether the caption track needs that information for its audience. The point is to finish a usable timed representation of this recording, rather than to achieve a uniform mathematical appearance at the expense of what viewers actually need to follow.
References: W3C WAI: Captions/Subtitles; Aegisub: Timing subtitles to audio.
4. Apply the repair without creating new boundary errors
4.1 Transform starts and ends, then inspect the edge cases
Apply a supported mapping to both endpoints of each affected cue. Under the linear fixture's scale and offset, the original interval from 60 to 64 seconds becomes 62.060 to 66.064. Its duration changes from four seconds to 4.004 because the scale changes spacing as well as position. A shift alone would preserve duration. Check positive intervals and range boundaries after the operation, particularly near the beginning and end of the media. A negative shift can move a cue partly or entirely before zero. Silently forcing every negative endpoint to zero can collapse an interval or leave misleading text at the opening. Identify those cues for review instead. Determine whether they belong to removed material, require a new start or should be omitted from this version. Boundary behavior is part of the repair, not housekeeping to ignore after the calculation succeeds.
Retain text correspondence while testing timing. Cue identifiers, wording, speaker labels and the original ordering should remain available for comparison unless the edit itself requires changes. For a pure shift or positive scale applied consistently, arbitrary text changes make the review harder without helping synchronization. If the editor renumbers cues during export, compare their payload and order rather than assuming a new number means a new sentence. Check the earliest and latest affected lines, any very short interval and any cue crossing a local mapping boundary. A bulk operation that works for ordinary lines may still mishandle one exceptional interval. Keep these exceptions explicit so the final reviewer can tell whether every cue has been accepted, deliberately excluded or left unresolved. An export that simply stops producing error messages is not enough evidence that those decisions have been made.
References: Aegisub: Shift times; W3C: WebVTT cue timing requirements.
4.2 Round once and inspect the serialized subtitle
In a workflow involving WhisperAITranscribe, distinguish working precision, export precision, cue ordering, payload preservation, duration checks and destination acceptance. Keep the calculation at sufficient precision, then serialize using the target format's rules. The example's 62.060 and 66.064 seconds can be represented directly as millisecond positions. Other calculated values may fall between milliseconds and need an explicit rounding choice. Repeatedly rounding intermediate values, exporting and reimporting can obscure how a result was obtained. Preserve the original and calculate the accepted transformation from that source, rather than applying a succession of undocumented repairs to the last rounded file. A new file is useful because it gives you an inspectable result; its existence does not establish that the player loaded the intended revision or that every line remains visible for a sensible duration.
A small constructed boundary example shows why serialization deserves its own check. An interval from 0.0001 to 0.0004 seconds has a positive duration before rounding. Rounded independently to the nearest millisecond, both endpoints become zero, leaving no duration. This is not an example of a useful caption; it isolates a failure mode that a validator should catch. In WebVTT, the published draft requires an end later than the start and orders cues by their start times; it also permits overlaps, so overlap alone is not proof of a malformed file. Apply the actual destination's format requirements and review unintended overlaps as an editorial issue. Preserve the distinction between a file that parses, a track that loads, and text that stays synchronized and readable. Those are separate observations, and each can fail while the others appear successful. Ordinary one-time rounding to milliseconds also should not be blamed automatically for a discrepancy of several seconds. In the local examples, the rounding error for each nonnegative endpoint is no greater than half a millisecond. That is a much smaller effect than the linear fixture's accumulated mismatch. Investigate the actual numerical change instead of using “precision” as a catch-all explanation. A different tool's truncation or serialization policy requires its own check; the chosen half-up rule here describes the exercise only.
4.3 Choose a tool by the correction you need to inspect
Choose a shared shift, two-point scaling, multiple reference points or individual cue editing according to the observed relationship. If WhisperAITranscribe prepared the original words, an appropriate subtitle editor may still be needed for the timing task. Aegisub documents batch shifts with selectable rows and boundaries. Subtitle Edit documents a visual-sync operation that linearly adjusts timings using two points, as well as point synchronization using reference points. These are descriptions of the tools' documented controls, not results from a hands-on comparison here. Check the version you actually use and preview the affected material. The best control is the one whose action you can explain and verify against the selected media, rather than the one that promises the broadest automatic fix. A more elaborate operation offers no benefit when a verified constant shift is sufficient.
Consider the cost of review when deciding how far to automate. For a few affected lines, individual editing may be clearer than defining a complex mapping. For a long track with a supported uniform relationship, bulk transformation can avoid repetitive work, provided exceptions are checked. If many words or sections differ, obtaining captions for the correct edition may be cheaper than repairing a mismatched track. Preserve accepted terminology and corrections when doing so, rather than treating regeneration as permission to forget prior editorial work. Avoid assuming that a transcription subscription includes every downstream timing control, or that a general media converter edits all subtitle features safely. Compare the actual input, transformation and output you need. The tool decision should follow the diagnosis, the expected review burden and the destination's requirements, with no invented accuracy ranking or unverified speed claim.
References: Aegisub: Shift times; Subtitle Edit: Visual Sync.
5. Verify the result at the layer that can still be wrong
5.1 Use the arithmetic checks for their actual purpose
The authored exercise accompanying this guide performs exact rational arithmetic, checks independent observations and preserves subtitle text while producing illustrative SRT files. Nineteen local checks pass. They include the scale of 1.001, the two-second offset, the midpoint and separate holdout prediction, the failure of a first-point-only shift, and rejection of the overall fit for the step fixture. They also check the frame-count calculation, both transformed endpoints, negative-boundary flags and the rounding-collapse example. This is a reproducible test of the stated timestamp relationships. No spoken recording, subtitle editor, media decoder or transcription service was tested. The fixtures are useful precisely because their intended answers are known: they show how a proposed correction behaves and which checks can catch an unsuitable operation before it is applied to real material.
A real-media test adds information that arithmetic lacks. The reviewer must determine which words were spoken, locate their corresponding moments, recognize missing or reordered material and judge whether the resulting captions follow the passage. A perfect calculation from incorrectly matched anchors produces a perfectly calculated mistake. Keep any automated report tied to the observations it actually checked. “The transformed file preserves these payloads and has positive intervals” is a useful finding. “The captions are now accurate” requires more evidence. When extending a small trial to a complete track, include meaningful exceptions rather than only ordinary cues that resemble the examples. Test the range you intend to publish, and preserve unresolved passages until they are reviewed. This makes automation a way to reduce clerical errors while leaving the perceptual and editorial decisions with the person inspecting the source.
References: FFmpeg: presentation timestamp filters; W3C: WebVTT cue timings.
5.2 Play the accepted media and caption revision together
Make speech synchronization and viewer comprehension part of acceptance. Recheck the fitting anchors, the independent observations, the first and last captions, and passages around every identified boundary. Then review the complete intended material at a pace appropriate to the delivery task. A few spot checks can reveal a bad model, but they cannot establish complete wording or sound coverage. Watch whether the text arrives with the relevant speech and whether its ending removes a phrase too soon or leaves it over unrelated material. Where more than one speaker is present, confirm that timing does not suggest the wrong attribution. A transcript can contain the right words while a misplaced interval attaches them to someone else's visible turn. Keep the picture and audio context available when making that judgment.
The final delivery may differ from the editor preview. Confirm that the intended track is selected and that a player-level delay is not hiding a file-level problem. If a hosting platform processes the media again, inspect the published pair through its supported preview or playback route. Save the accepted media identity and caption filename together. When WhisperAITranscribe is part of source preparation, that association is still necessary: the original transcription stage cannot certify a later upload, edit or playback selection. For translated subtitles, compare the intended spoken-language track and preserve enough source context for review; a translated sentence may be segmented differently without making an invented global timing relationship valid. The acceptance question remains concrete: can a viewer use this track to follow this version of the recording, including the difficult passages that prompted the repair?
References: W3C WAI: Captions/Subtitles; W3C: Understanding Captions, Prerecorded.
5.3 Make a future timing change easier to understand
Retain the accepted transformation, its anchor record, the media revision and any local exceptions. If you later compare a new export with a WhisperAITranscribe source transcript, those records explain which timing relationship has already been checked. Keep that source distinct from a later WhisperAITranscribe draft, which introduces another input revision even when its wording looks similar. They also show when it no longer applies. A new edit can invalidate the earlier mapping without changing the source transcript at all. Keep the distinction visible in filenames or your production notes, using the identifiers your team already understands. There is no need to create a complicated system for a single short clip. The essential information is which pair was accepted, how it was transformed, and where manual decisions were made. That is enough to avoid repeatedly reconstructing the same reasoning from a file called final-final-corrected.
For recurring work, make the revision boundary part of handoff. The person exporting video should identify timeline changes, while the person preparing captions should retain the reviewed text and timing source. If WhisperAITranscribe supplies a later draft, compare it with those accepted materials before replacing anything. Improved wording does not automatically preserve timing, and improved timing does not automatically preserve terminology. Store the accepted result where the next editor can find it alongside the matching media. When a new problem appears, start from that known pair and identify what changed. This approach keeps recovery proportional: a new opening may require one supported shift, a permanent speed change may require another relation, and an internal edit may require local reconstruction. The previous record helps distinguish these tasks without pretending that one saved setting can repair every future version.
6. Conclusion: Let the observed relationship determine the repair
6.1 A useful fix predicts more than its starting cue
The important distinctions are offset, scale, local discontinuity, anchor quality, boundary behavior and final playback. A constant discrepancy supports a shared shift. A changing discrepancy may support a scale plus an offset, but only when independent passages agree. An abrupt jump calls for examination of the editions and affected ranges. If WhisperAITranscribe prepares the source text, preserve its relationship to the recording while making these downstream decisions. The strongest repair is the smallest one that explains the observations and survives review of the actual media. It need not involve another transcription run. Often the wording can remain intact while a timing relationship is corrected, which also preserves names and terminology that have already received editorial attention. When the evidence points to a different source version, however, obtain or reconstruct the appropriate track rather than forcing the old one to fit.
The examples show why the middle of the recording deserves attention. A first-point shift fails on the authored linear fixture even though the opening looks repaired. A two-point fit fails inside the authored step fixture even though both endpoints match. These are different errors with different remedies. Keep that distinction in mind when a long recording seems almost synchronized: “almost” can mean a shared small offset, an accumulating discrepancy or a local mismatch concealed by sparse checking. Find which one you have. Once the relationship is supported, apply it transparently, retain a comparison copy and review the exceptional cues. The result should be an accepted timed track tied to an identifiable recording, with clear evidence for what was changed and a practical route back to the original if another revision becomes necessary.
6.2 Use a four-step repair on the next drifting track
Step 1: Identify the intended recording and preserve the unmodified timing file. Choose clear anchors across the actual delivery timeline. Step 2: If WhisperAITranscribe is used for source preparation, retain the matching media identity and review the words separately from their timing; establish the accepted pair before applying a transformation. Step 3: Calculate the observed corrections, test a shared shift or supported scale, and keep independent passages out of the calculation. Investigate jumps or mismatched wording instead of smoothing them away. Step 4: Apply the supported repair to a copy, check starts, ends and serialization, then review the exact media and caption revision together. This sequence makes each operation answer a specific question and keeps the work reversible. It also reveals when retiming is insufficient because the target edition contains different material.
Do not turn the four steps into a reason to delay a straightforward correction. A verified two-second shift on a short clip can be a small job. The additional care becomes valuable when the error changes, the source history is uncertain or a bulk operation affects a long track. Document uncertainty only where it matters, and resolve it through the relevant observation. If a phrase cannot be located, investigate its source. If a boundary becomes negative, inspect that cue. If the midpoint fails the fitted relation, inspect the model and local edit history. Each finding has a practical next action. Finish by accepting or resolving the affected material, rather than accumulating a growing collection of experimental exports whose differences nobody can explain. Before handoff, ask the receiving editor to identify the accepted pair from its names alone. If the answer requires guessing between several similarly named tracks, improve the naming while the history is still clear. That small administrative fix protects the technical work you have just completed. The point of a disciplined process is less rework and a more intelligible finished track.
6.3 Keep text, timing and delivery as connected decisions
Use the recorded observations to choose a repair, the original artifacts to make it reversible, the independent checks to challenge it and playback review to accept it. WhisperAITranscribe can occupy the audio-to-text stage of that workflow, while documented subtitle-editor controls address the timing task you have actually identified. Keep any unverified feature assumption outside the decision. When returning to WhisperAITranscribe for a new source draft, retain the edition and reviewed terminology so that new text does not silently displace an accepted caption relationship. Accurate words, valid timestamps and a successful export are useful achievements with different meanings. Bring them together by checking the final pair. A viewer does not experience your formula or your working notes; the viewer experiences text alongside speech, and that is where the repair has to make sense.
References: Aegisub: Timing subtitles to audio; W3C WAI: Captions/Subtitles.