Fix Audio Sync Drift by Mapping Where Timing Goes Wrong
Diagnose constant offset, progressive drift, and sudden sync jumps before changing audio speed.
The question that matters
The opening clap lines up perfectly. Ten minutes later, a presenter closes their mouth while the last word is still playing. Sliding the audio fixes that sentence and breaks the beginning. This is the moment to stop nudging and measure how the error changes.
Fixing audio sync drift starts with a pattern, not a preferred setting. A fixed delay, a steady timing divergence, and a sudden jump need different repairs. The workflow below uses an illustrative interview recorded on a camera and a separate recorder. Its numbers are hypothetical teaching examples, not measured product results.
Establish one reference before touching timing
Duplicate the sequence and preserve the original files. Choose the picture or other recording that defines the intended timeline. For an interview, that will usually be the approved camera footage, but do not assume a proxy or previously retimed export is authoritative. Compare it with the source first.
Find events that exist in both recordings: a clap, a distinct consonant, or an object contacting a table. A visible gesture without a sharp sound is a weaker anchor. Avoid comparing distant microphones as if sound reached both at exactly the same instant. Keep the same recording pair throughout the diagnosis so a change of reference does not masquerade as a change of drift.
Put it into practice
- Name the reference and suspect file.
- Keep untouched copies and record existing speed changes.
- Choose several recognizable events across the take.
Build a small error map
Place markers near the beginning, roughly a quarter through, halfway, and near the end. At each marker, note whether the suspect audio is early or late relative to the reference. Use one unit throughout, such as milliseconds. Define positive error as late audio and write that convention beside the measurements.
In the illustrative interview, suppose an opening event has zero error, a middle event is 120 milliseconds late, and the closing event is 240 milliseconds late. That suggests a roughly consistent rate difference. It does not establish the cause. Add another marker between those events before deciding that the entire file needs one stretch. A handwritten table is sufficient; precision comes from identifiable events, not elaborate reporting.
Follow the shape of the error
If every marker is similarly late, treat the problem as offset: move the clip while preserving its duration. If the error grows progressively after a correctly aligned opening, investigate drift. Sound Devices distinguishes these head-alignment and rate-related problems in its 970 guide. If the error changes abruptly and then stays steady, look for an edit, missing material, or a discontinuity instead.
A combination is possible. An initial offset can sit on top of a rate difference, so first align one reliable early anchor and measure the remaining pattern. Do not average away a sudden jump. A single global speed adjustment can make the first and last events agree while leaving a wrong middle section, which is why two-point alignment alone is insufficient.
Investigate before choosing a correction
Check the original recording settings, source duration, project interpretation, and any speed changes already applied. Two devices showing the same nominal sample rate are not proof that their clocks ran identically. Sound Devices explains that independently running clocks can diverge and that appropriate shared clocking prevents that particular problem during recording.
Also compare the source with its proxy or intermediate. If the source holds sync but a derivative does not, repair that conversion path before stretching good audio. If the problem appears only during preview, test a short export through a direct playback path. Monitoring delay is a different hypothesis from an error baked into the file. Record what each comparison rules out instead of changing several settings simultaneously.
Correct a consistent rate difference carefully
For a stable drift pattern, measure the interval between the same two events in the reference and suspect audio. The required duration multiplier is reference interval divided by suspect interval. In a hypothetical example, 600.000 seconds divided by 600.240 seconds is about 0.999600. The suspect interval needs to become slightly shorter. Software may ask for duration percentage or speed percentage; those are reciprocals, so check the control's meaning.
Anchor the first event, apply the correction to a copy, and verify interior markers. A pitch-preserving duration tool can change length without deliberately changing pitch, as Audacity's Change Tempo documentation explains. Listen for altered consonants, transient softness, and unnatural room sound. A mathematically correct endpoint does not guarantee an acceptable recording.
Illustrative decision
The interview's middle and late anchors improve together after a small duration correction. Keep that version for comparison, then listen through transitions rather than assuming the calculation proves the whole take.
Repair discontinuities as separate sections
When timing jumps at one point, inspect that neighborhood closely. Compare waveforms before and after the jump, look for a cut, and check whether the reference itself changes. If an interruption divides the recording into two internally consistent regions, align those regions independently instead of stretching the uninterrupted first region to compensate for the second.
Place any necessary audio edit where it preserves the performance and room continuity. A pause may be useful, but a silent-looking waveform can still contain room tone or a breath. Listen across the edit without watching it. If no clean boundary exists, a replacement phrase or another microphone may be preferable to repeated micro-corrections that change speaking rhythm. Preserve editorial choices in a short repair note.
Verify the exported deliverable
Render the intended delivery version, then check that file rather than only the editing timeline. Revisit every diagnostic marker and add spot checks between them. Watch mouth movements at normal playback speed; waveform alignment is helpful evidence, but the audience experiences speech and picture together.
Check the beginning after repairing the end, and the end after any later cut. Confirm that a new render has not reused an old cached file or a similarly named export. In the illustrative interview, the acceptance record should state which file was checked, which anchors were compared, and whether any sections remain uncertain. Avoid claiming frame-perfect alignment when the visible action cannot support that level of precision.
Put it into practice
- Check opening, interior, and closing anchors in the final file.
- Listen for processing artifacts and repaired seams.
- Retain original files, correction values, and export identity.
Make the next recording easier to diagnose
Record an identifiable sync event and, where practical, another near the end of a long take. Test the actual camera-recorder combination for the expected recording duration before an important session. A short test may establish offset while missing a slow accumulation of error.
Where the equipment supports it, follow the manufacturer's synchronization instructions and distinguish timecode identification from the clock reference that controls recording rate. Sound Devices' 788T guidance specifically discusses locking word-clock sources to prevent drift. That is equipment-specific practice, not a promise that any timecode cable solves every setup. Even a well-configured rig benefits from a final reference check after media transfer.
Choose the repair from the observed pattern
| Observed pattern | First action | Failure sign |
|---|---|---|
| Similar error everywhere | Shift the clip without changing length | One end improves while the other worsens |
| Smoothly growing error | Check interpretation, then test proportional retiming | Interior anchors remain wrong |
| Sudden step in error | Inspect discontinuity and align regions | Global stretch damages previously correct regions |
| Preview only | Compare an export and playback path | Changing media to compensate for monitoring delay |
Questions worth asking
Why does audio align at the beginning but not the end?
The files may run at slightly different effective rates, or a later discontinuity may create a new offset. Measure multiple points before deciding which explanation fits.
Will changing the sample-rate label fix drift?
Can I fix drift by cutting and moving every sentence?
What if only the final export is wrong?
Sources & scope
Cases are illustrative editorial examples, not measured performance results. The workflow is tool-agnostic.