Last month I opened a two-camera episode and the guest's mouth finished moving about four frames before her voice arrived. Not at the start. At minute 41. The first forty minutes were fine.
That is the part nobody warns you about. Audio desync is rarely a wall you hit at 00:00:00. It creeps. It starts as a feeling that something is slightly wrong, you scrub back, you tell yourself it is fine, and then twenty minutes later the lips and the words have visibly divorced. By then you have already cut half the episode against a bad reference.
I have fixed this enough times to have opinions. Here is what actually causes it, what to do in the moment, and the boring habits that stop it happening again.
First, decide which kind of desync you have
There are only two, and they need completely different fixes.
Constant offset. The audio is late by the same amount for the whole file. Usually because someone hit record on the recorder before or after the camera, or the file got dropped into the timeline at the wrong point. This is the easy one. You slide the audio until it matches and you are done forever.
Drift. The gap grows over time. Perfect at the top, one frame off at twenty minutes, five frames off at the end. This one is a clock problem, not a placement problem, and sliding will never fix it. If you slide to fix the end, you break the start.
The test takes thirty seconds. Sync a hard transient near the top of the episode, a clap or a laptop lid or the first plosive. Then jump to the last five minutes and look again. Still matched? Constant offset, already solved. Slipped? Drift.
Ninety percent of the time people spend fighting sync is spent applying the constant-offset fix to a drift problem, over and over, wondering why it will not stick.
Why drift happens, in plain terms
Your camera and your audio recorder each have their own internal clock, and those clocks are never exactly identical. One runs at 47,999 samples per second where it thinks it is running 48,000. Over five minutes nobody notices. Over ninety minutes that tiny error piles up into something you can see.
The usual suspects, in the order I actually find them:
- Two devices recording separately with no shared clock or timecode. Standard for double-ender setups.
- A camera recording 29.97 while the project is 30, or the reverse. This is the classic one, and it produces almost exactly one second of slip per thousand seconds.
- Remote recording where one side's browser resampled the audio on the way through.
- Variable frame rate footage from a phone or a screen recorder, which is not really a frame rate so much as a suggestion.
- A guest whose laptop dropped packets for ten seconds, which is not drift at all but looks like it until you inspect the waveform.
That last one matters. Dropped audio from a bad connection creates a hard step, not a slope. If you see the gap jump instantly at one point and stay constant after, you have a dropout, and the fix is a cut, not a stretch.
The fix that works for constant offset
Line up on a transient, not on a word. Words have soft attacks and your eye lies to you. Find a clap, a table knock, a chair, anything with a vertical spike in the waveform. Zoom the timeline in until one frame is wide on your screen. Nudge frame by frame until the spike sits on the same frame in both tracks.
Then check the mouth. Waveform match and lip match are not always the same thing, especially if the mic was three metres away and you are now looking at a real speed-of-sound delay. Trust the lips. The audience watches lips.
If you record double-ender, and you should, the whole workflow gets easier because you are always doing this deliberately instead of discovering it later. I wrote up how we handle that in our notes on double-ender recording, including the one clap habit that has saved more episodes than any plugin.
The fix that works for drift
You cannot slide your way out of drift. You have to change the length of something.
The clean version is a speed change on the audio, applied to the whole file, at a ratio that is almost never a round number. Measure the offset at the very start and at the very end, work out the difference across the known duration, and stretch the audio by exactly that percentage. If the audio ends 3 frames late over a 60 minute episode at 30fps, that is 0.1 seconds over 3600 seconds, so you stretch by roughly 1.0000278. Do it once, to the master audio file, before you cut anything.
The 29.97 versus 30 case has a known ratio, 1.001, and you should just apply it directly rather than measuring. Same for 23.976 against 24. If your numbers land near 0.1 percent, that is what you are looking at.
The ugly version, which I use more often than I would like to admit, is segmented sync. Cut the audio into chunks at natural gaps, usually the pauses between questions, and re-sync each chunk. Nobody hears a two frame slide inside a silent gap. This is unglamorous and it works, and on an episode with a phone recording in the mix it may be the only option.
The thing I will not do is nudge sync at every cut and hope. That produces an episode that is subtly wrong everywhere, which is worse than an episode that is obviously wrong in one place.
How far off is actually visible
Useful thresholds, from watching a lot of people watch a lot of episodes:
- Under 1 frame: invisible. Stop fiddling.
- 1 to 2 frames: nobody consciously notices, but the episode feels slightly cheap. Fix it if it is cheap to fix.
- 3 to 4 frames: viewers notice something, most cannot name it. This is where comments start saying the audio sounds off when the audio is perfect.
- 5 frames and beyond: it is a dubbed film. Everyone sees it.
Audio arriving early is roughly twice as noticeable as audio arriving late, because late audio matches how sound behaves in the real world. If you have to leave a small error, leave it late.
Where sync breaks that people forget
Sync survives the edit and then dies in three places.
The first is clips. Vertical exports get re-encoded, re-timed and sometimes re-framed by whatever tool made them, and a clip cut from a perfectly synced master can still ship one frame out. Always check a clip on its own before it goes out, especially if captions are riding on top of it. We build clips from the finished master for exactly this reason, and the checking is part of the job rather than an afterthought. If you are shipping them on a schedule, the rhythm in posting cadence for podcast clips assumes the clips are actually correct when they go out.
The second is the intro and any music bed that was laid before the picture was locked. Move a cut, and everything downstream that was manually placed goes with it or does not.
The third is file naming. If your camera A file and your audio file for episode 12 are named the same thing as episode 11, someone will eventually sync the wrong pair and the offset will look like drift when it is actually the wrong take. That sounds absurd until it happens. The habits in naming conventions that save hours exist because of days like that.
The prevention list I would actually give you
None of this is exotic. All of it removes work later.
- Clap once, on camera, in frame, at the start of every recording session. Clap again after any break where a device was stopped and restarted.
- Set every device to the same frame rate and the same sample rate before you start, and check it on the day, not from memory.
- Record a safety audio track on the camera itself, even if it is a bad mic. It is your reference, not your final sound.
- Never let a recorder run past its file split boundary without knowing where the split happens.
- Keep the raw files. If sync goes wrong in the edit, the answer is almost always upstream.
A single-camera show with a lav going into the camera has almost none of these problems, which is one of the quiet arguments for keeping the setup small until the show is actually working. More cameras is more sync surface. If you want the honest version of that trade, the video podcast equipment guide lays out what earns its place and what does not.
Why we do this by hand
Every episode we cut is watched end to end by a human editor. Nothing is auto-cut by silence detection, and that is not a philosophical stance, it is a practical one. Silence detection has no idea whether the audio drifted. It will happily cut on a silence that has moved four frames away from where the picture thinks it is, and it will do that six hundred times in ninety minutes.
A person watching the whole thing catches drift at minute 41 because they are looking at minute 41. That is most of the value.
If the packaging side matters to you too, our sister studio ClickTheory handles thumbnails, and the same principle applies there. Someone looks at the actual thing.
Do the two-point test. Top of the episode, end of the episode. Constant offset, slide it. Drift, stretch it once and only once, before you cut. Dropout, cut it out and move on.
And if you would rather hand the whole mess to someone else, that is what we do. First cut in 48 hours, revisions the next working day, and you can see the pricing without talking to anyone. Either way, please clap at the top of the next one. It costs you one second and it saves me an afternoon.