Fix Subtitle Timing and Drift

Open your existing SRT or WebVTT in the subtitle timing editor and sync fixer, then optionally load the matching local video. You choose the reference moments yourself: the editor does not listen to audio, recognize words, or automatically align speech. Its global shift and two-anchor linear retiming apply explicit timing changes that you can preview before accepting. The examples below are invented teaching cases, with seconds written to three decimal places where useful for checking the arithmetic.

Identify a constant delay, growing drift, or a changed edit

Begin with three listening points: an early sentence, a sentence near the middle, and a late sentence. For each, note the subtitle reference time and the time where that same event belongs in the video. Compare like with like, such as the onset of the same clearly spoken word. A vague impression that the text feels late is useful for noticing a problem, but it is not precise enough to choose a whole-track correction.

Calculate the required correction as target video time minus current subtitle time. If a cue starts at 20.000 seconds but its spoken phrase begins at 21.200 seconds, it needs +1.200 seconds. If similar checks elsewhere also need about +1.200, the evidence suggests a constant offset. The entire track may have been prepared against a version with a different opening, while the speech itself retained the same pace.

Uniform drift has a different signature. An early reference might need a small correction while later references need progressively larger ones. The difference grows smoothly rather than jumping at a particular scene. Two anchors can describe that straight-line relationship, but two observations alone cannot prove that the whole recording follows it. A middle reference supplies an independent test instead of merely confirming the two points used to calculate the result.

An edited video often produces steps: the first minute matches, a section after a removed pause is suddenly early, and another cut changes the error again. Reordered scenes can also make wording appear in the wrong sequence. Neither pattern is solved by stretching every cue equally. Mark the first point where the relationship changes and compare the versions. Correcting a single region manually can be more appropriate than forcing unrelated sections onto one global timing model.

Build reliable reference points from the actual video

Use the same final media file that will accompany the exported subtitles. A preview copy with a different introduction, trimmed silence, or altered playback speed is a different timing reference even when its filename looks familiar. Write down enough of the version name to distinguish it later. Also confirm that the subtitle language and spoken take match; a correct clock cannot compensate for a sentence recorded in a different order or omitted entirely.

In an invented pottery demonstration, the presenter says a distinctive instruction just after lifting a clay bowl. That spoken onset makes a better anchor than the bowl movement alone: the presenter might move before speaking. Replay a short stretch at normal speed, then seek around the word to refine your choice. Record which word or sound you used so a second reviewer can locate the same event without guessing from the timestamp.

Choose anchors well apart, within material you can verify confidently. Two neighboring cues may both sound clear, but small measurement errors between them can produce a misleading scale over a long track. An early and a late reference test more of the recording. Avoid a long fade, indistinct background speech, or a cue whose start was intentionally placed before the voice. An anchor should represent the relationship you want to preserve throughout the track.

Check whether your chosen subtitle reference is a cue boundary or an event within the cue. Either can describe a time relationship, but label it accurately and use the corresponding event in the video. Do not pair the cue start with the last word of its sentence. When loading optional local media, confirm it has finished loading and that playback works. If it cannot decode in the browser, use a supported local review copy with an identical timeline.

Apply a global shift when every cue needs the same move

A global shift adds the same signed amount to every start and end: new time = old time + shift. Positive values move subtitles later; negative values move them earlier. The sign describes movement on the timeline, not whether the current file is described as delayed. When subtitles are already late, they usually need a negative shift. State the intended movement in plain words before entering the number to avoid reversing the correction.

For a separate invented offset example, a cue currently runs from 8.400 to 11.100 seconds, while the matching phrase belongs at 7.200 to 9.900. Apply -1.200 seconds: both boundaries move earlier, and the duration remains 2.700 seconds. A second cue at 15.000 to 17.500 becomes 13.800 to 16.300. The gap between those two cues also remains unchanged because the same amount was added to every boundary.

Preview the proposed values before applying them. Check an early cue as well as the reference that prompted the adjustment. A shift of -1.200 applied to a start at 0.800 would produce -0.400, which is not a valid output time for this editor. The whole transformation should be rejected, not partially applied or clipped to zero. Clipping would shorten that cue and hide the fact that the proposed model does not fit the opening.

After accepting a valid shift, listen through several transitions and inspect the final cue. The operation fixes timing placement, not wording, line length, or text density. It also leaves existing overlaps intact, so those still need review. Export a checkpoint only after checking the result, and retain the original separately. If a second adjustment seems necessary, calculate it from the current track or return to the original deliberately; do not accidentally apply the original correction twice.

Use two anchors to calculate a linear drift correction

Two-anchor retiming maps two known subtitle times to their matching video times. Call the subtitle references s1 and s2, and the target video references v1 and v2. The scale is a = (v2 - v1) / (s2 - s1). The offset is b = v1 - a * s1. Every boundary then follows new time = a * old time + b. This is a straight-line mapping, not an audio analysis.

In the invented pottery example, the first subtitle reference is 12.000 seconds and belongs at video time 12.600. The second subtitle reference is 312.000 and belongs at 315.600. The subtitle interval is 300.000 seconds; the target interval is 303.000. Therefore a = 303 / 300 = 1.01 and b = 12.600 - 1.01 * 12.000 = 0.480. The completed mapping is new time = 1.01 * old time + 0.480.

Check the arithmetic against both anchors: 1.01 times 12.000 plus 0.480 is 12.600, and 1.01 times 312.000 plus 0.480 is 315.600. A cue at 112.000 to 115.000 becomes 113.600 to 116.630. Its original 3.000-second duration becomes 3.030 seconds. Unlike a shift, linear scaling changes durations and gaps in proportion to the scale. That makes a fresh reading review necessary even when the anchor alignment is correct.

An independent middle reference at subtitle time 162.000 is predicted to land at 164.100. Listen for its matching event there. If it lands correctly, the linear model has passed a useful additional check. If it misses substantially while both anchors fit, investigate a cut, an uncertain anchor, or a changing speed relationship. Do not add an unrelated offset to make the middle fit, because that would move the two verified endpoints away from their targets.

Catch invalid bounds before accepting the transformation

Enter two distinct subtitle reference times in chronological order, with their matching video references also increasing. Identical subtitle times make the scale denominator zero. Reversed targets imply an unusable backwards timeline, and a zero-length target interval collapses every cue. Even sensible-looking anchor entries can map an earlier cue below zero. The preview must evaluate the entire track rather than checking only that the four anchor fields contain numbers.

Use the earliest cue as a boundary test. Suppose a different teaching mapping is new time = 1.02 * old time - 2.000. A cue starting at 1.000 would map to -0.980 seconds. Do not approve that result just because later material looks correct. Revisit the references and the opening edit, or make a deliberate content repair. An invalid global transformation is refused without changing the accepted cue list; it does not silently preserve only the later cues.

A valid track also needs each end to follow its start. Tiny positive durations can still become troublesome when exported at millisecond precision, especially after a strong compression. Inspect the proposed serialized boundaries rather than relying only on an attractive scale value. Keep full precision while calculating, then review the rounded output. A warning about fast text is different from a zero or negative duration: the former invites editorial judgment, while the latter blocks a timed export.

If a transformation is rejected, keep the explanation with your review notes and correct the underlying entry. Avoid repeatedly changing random numbers until an export becomes available. Undo can reverse a recent successful edit or timing operation, but a rejected operation should not create a partially changed state to undo. After undoing, verify which version is on screen before entering fresh anchors. Exported checkpoints and a short record of accepted adjustments make that distinction easier to maintain.

Repair cuts and local errors without stretching unrelated cues

When the error changes suddenly, locate the transition before editing individual times. Compare the sentence immediately before the suspected cut with the first complete sentence after it. A removed introduction can explain one constant offset for everything that follows; several internal removals create several relationships. The global controls act on the whole loaded track. They are not a segment selector, scene detector, or automatic way to reconstruct the editorial history of a video.

For a shortened demonstration, review the subtitle content against the surviving speech first. A cue that describes a removed step may need deletion rather than a new timestamp. Preserve a source copy before removing it, and check that the remaining instructions still make sense. A timing repair should not leave a command displayed over a different action merely because a calculated boundary fits within the video. Meaning and sequence matter as much as numerical validity.

Split a cue when its two meaningful parts now need separate intervals. Choose both the split time and the text or line boundary so words are neither duplicated nor lost. Check the new durations against the spoken phrases. Merging with the next cue can help when a short response belongs with its lead-in, but review the combined text and any intervening silence. Neither operation should be used simply to make a warning count smaller.

For extensive rearrangements, return to the video editing workflow and prepare a track for the final sequence, then use this editor for focused cleanup. There is no need to treat a nonlinear problem as a failed linear calculation. Keep a note identifying which region changed and why a global fix was unsuitable. That explanation lets the next reviewer start from the correct assumption instead of repeating a two-anchor experiment that cannot fit the material.

Review playback and reading quality at the beginning, middle, and end

After a successful transform, replay the beginning, middle, and end without adjusting anything during the first pass. Note where a subtitle appears before the relevant speech, disappears too soon, or remains over the next speaker. Then inspect those cues individually. Pausing at an exact reference verifies a boundary, while normal playback tests the reading experience. Both checks are necessary because a cue can have mathematically correct endpoints and still divide a sentence awkwardly.

Read through warning categories separately. An overlap may be accidental, or it may represent two voices heard together. An out-of-order cue needs investigation before any manual reordering. Long lines, brief durations, and fast text point to possible reading pressure rather than automatic failures. Do not delete a bracketed sound note or a speaker label just to lower a number. Those details may explain a response that otherwise becomes confusing when the audio is unavailable.

The editor’s approximate reading speed uses Unicode code points including spaces per second. It is not words per minute, a measurement of rendered line width, or an accessibility certificate. As an invented arithmetic check, 36 counted code points over 2.400 seconds gives 15 code points per second. If linear compression reduces that duration to 1.920 seconds, the same text reaches 18.75. Use editable review preferences to flag material for listening, not to declare universal platform limits.

The preview is plain text and cannot certify the final appearance of styled captions. Inspect the exported track in the intended destination as well, particularly when it contains supported VTT settings or markup. Keep line breaks meaningful, verify names and technical terms, and listen across cue boundaries rather than reviewing isolated cards. Finish with a continuous watch of the repaired area so a correct anchor does not distract you from a neighboring cue that still needs attention.

Export the corrected track with a reproducible handoff

Choose the output needed for the next step: SRT or VTT for timed text, and a plain TXT transcript for wording review. A transcript does not retain the timing model and cannot replace the corrected subtitle file. Keep a native-format working copy when supported metadata matters. Before a conversion that would lose identifiers, settings, or formatting information, read the warning and acknowledge the loss deliberately. A change of extension alone does not perform a valid conversion.

This focused editor accepts basic text tracks. Advanced VTT STYLE or REGION blocks and live timestamp maps are refused rather than imported with their extra information silently removed. If you encounter one, obtain a suitable basic-track export from the source workflow or continue in an environment that supports those features. Retain the original advanced file. A successful timing calculation is not permission to discard information that affects how the source track was meant to behave.

Use filenames that distinguish the source, final video version, language, and timing revision. In a short handoff note, record the accepted shift or the two anchor pairs and scale. Include which middle reference you checked and any remaining deliberate overlaps. If a PDF review report is available, treat it as a summary of changes and checks, not a subtitle file or a replacement for playback review. The receiving editor still needs the actual timed-text export.

Reopen the exported file and compare cue count, the first and last boundaries, and several distinctive phrases. Check speaker names and sound notes as well as the anchor cues. Deliver the corrected track with the exact matching video reference and identify any conversion losses. The result is a repair another person can reproduce and inspect, not simply a file labelled fixed. Save your downloads before leaving the page because the editor does not keep a saved browser history of the session.

Questions about Fix Timing and Drift

Continue with another subtitle workflow

Use the free subtitle editor to apply the ideas to your track, or explore another guide below. Browse all free tools for other preparation tasks.