Quick answer: If the audio is early or late by the same amount throughout an AI talking-head video, move the audio track by the measured offset. If the error grows toward the end, inspect clip duration, speed changes and frame-rate handling before regenerating. If the sound is correctly timed but the generated mouth shapes are wrong, sliding audio cannot repair the animation.
This guide covers AI avatars, talking photos and edited AI dialogue clips, including workflows where a separate voice recording is added after video generation. Keep the original exported file and work on a duplicate of your editing project.
Three kinds of sync timing
The timing examples below distinguish correct synchronization, a fixed delay, and progressively growing drift. We also prepared a separate schematic 12-second timing animation. That diagram is not shown in the YouTube clip below and is not a test of a lip-sync model or proof of a repair.
Real AI talking-head example — inspect the ending: We generated this eight-second clip with Runway Seedance 2.0 Mini at 720p using its generated speech and video. The speaker says, “The sound should match my lips. Watch the timing carefully.” After the spoken line ends, the mouth appears to make small additional movements during the remaining silence. Watch that final moment closely: it illustrates a possible generated mouth-animation issue rather than a simple fixed audio-track offset. This is an unmodified output, not a before-and-after repair. We have not measured a precise delay or isolated its technical cause.
First, identify the type of error
| What you see | Likely issue to investigate | First action |
|---|---|---|
| Same delay at start and end | Fixed track offset or trimmed silence | Shift audio or video on the timeline |
| Start matches, end does not | Timing drift, clip speed, timestamps or frame-rate interpretation | Compare source and export durations; inspect frame-rate handling |
| Mismatch starts after a cut | Misaligned edit or independently trimmed track | Check that cut and nearby clips |
| Words are timed correctly but lips form the wrong sounds | Generated mouth animation error | Re-run the affected short dialogue segment if supported |
Measure the problem before spending credits
- Test the downloaded file in a second player, preferably with wired headphones or speakers instead of Bluetooth audio. If the issue appears only in one editor or playback setup, diagnose that first.
- Find two or three clear cues near the beginning, middle and end. A visible lip closure for “p” or “b” can help, but the sound onset is not always exactly the same instant as the visible gesture.
- Zoom to individual frames and inspect the audio waveform. Write down the approximate timing difference for each cue.
- Classify the error as fixed, growing, sudden after a cut, or incorrect generated mouth shapes.
- Change one thing on a duplicate timeline, then export a short preview and replay the exported file.
Frame conversion: time in milliseconds = (number of frames ÷ timeline frames per second) × 1,000. At 30 fps, three frames are 100 ms; at 24 fps, three frames are 125 ms. Do not assume every clip uses 30 fps.
Fix 1: a constant audio offset
Suppose a speaker’s visible mouth action happens at frame 60, while the matching audio cue is at frame 63 on a 30 fps timeline. The sound is approximately 100 ms late. If the same gap occurs near the end, move the audio three frames earlier, preview, and check again. If the sound comes before the mouth, move it later instead.
Keep the original track untouched until you confirm the export. If your editor offers a frame-nudge command, use it instead of dragging by eye. Avoid shifting only one section when the same delay affects the whole clip.
Fix 2: timing drift that grows
For example, if the first and last spoken phrases are both 100 ms late, shifting the entire audio track earlier by about 100 ms may help. If the first phrase aligns but the last is 400 ms late, do not shift the entire track: investigate durations, speed changes, cuts and frame-rate handling. If the first line is aligned but the last line is not, a single shift cannot solve both. Compare the original video and audio durations with the exported version. Look for independently stretched audio, accidental speed changes, variable-frame-rate sources, incorrect interpretation of footage, or timestamp problems. A mismatch between the project frame rate and the source frame rate is worth investigating, but it does not by itself prove that it caused the drift.
Where variable-frame-rate footage causes editing problems, make a constant-frame-rate working copy from the original source and retest. CFR conversion can make frame timing more predictable for editing, but it does not automatically repair every sync problem. Preserve the original audio, avoid repeated conversions, and export a short sample before changing the whole project. If drift begins only after a cut, examine that edit rather than stretching the full soundtrack.
Fix 3: the mouth animation itself is wrong
A generator may animate the mouth inaccurately even when audio and video tracks are aligned. If the wrong shapes appear on particular syllables, a global offset may make other words worse. Use the actual voice file that will appear in the finished video, make the mouth large enough to see, avoid occlusion and extreme head turns, and try a shorter line if your chosen tool supports segment-based generation. None of these inputs guarantees perfect lip sync. Fast speech, overlapping dialogue, extreme head turns, partially hidden mouths and difficult tooth or lip details can still cause visible errors.
Practical prompts and audio preparation
For an AI talking-head generation step, start with this scene instruction and adapt it to the tool’s supported controls:
One adult speaker in a stable medium close-up, facing camera. Keep the mouth clearly visible throughout one short spoken line. Minimal head rotation, no cuts, consistent lighting and natural facial motion. Use the supplied dialogue audio at its original timing. Do not add extra spoken words.
For separately generated narration (for example, from ElevenLabs), export the final spoken line with its intended pauses. Keep an untouched copy. Do not trim leading silence from only one track after lip-sync generation. If a model accepts an audio reference, follow its actual input specifications; prompt text alone may not control exact speech timing.
Runway or another AI generator: what to check
Runway’s dedicated Lip Sync workflow can animate a selected face using uploaded dialogue audio; standard text-to-video generation does not automatically offer the same audio-driven control. Check the selected workflow and its supported inputs before spending credits. When using a tool that generates speech and video together, test the untouched output before importing it into an editor. When combining silent video with an external voice, synchronization must be verified separately.
When regeneration makes more sense
Editing can correct track timing, but it cannot invent accurate mouth shapes. Consider regenerating a short affected segment when the mouth is obscured, the face turns sharply away, the generated articulation is wrong on particular words, or multiple speakers overlap. If the source audio is noisy or clipped, clean or replace it first, then test a short passage. Recheck the exported result rather than assuming a new generation is better.
Export checklist
- Keep the original source files and the edited version separately.
- Check the first, middle and final spoken phrase in the exported MP4, not only the editing preview.
- Verify that the intended audio is present, untrimmed and at the expected playback speed.
- Review any speed changes, frame-rate conversions and cuts made after lip-sync generation.
- If a specific word remains wrong while the rest aligns, consider regenerating only that passage rather than the entire video.
What we actually tested
We prepared two different assets, neither of which is evidence of a successful lip-sync repair. Timing diagram: On November 14, 2026, we created a 12-second schematic animation with three four-second sections. Mouth events occur at 0.6, 1.6 and 2.6 seconds in each section. Clicks in the second section are delayed by 300 ms; those in the third section are delayed by 150, 300 and 450 ms. It was rendered at 24 fps, 960 × 540, H.264 MP4 with AAC click audio. These timings are defined by the animation design, not measured from a real speaker; no Runway credits were used for the diagram. Talking-head example: Separately, we generated the eight-second, 720p Seedance 2.0 Mini clip embedded above, with generated speech and video. It used Runway generation credits. The clip has a visible possible failure worth inspecting: small mouth movements appear to continue after the spoken sentence ends. This observation is not a controlled measurement of delay, a demonstrated correction, or a benchmark of model accuracy.
Frequently asked questions
Why is my AI lip sync delayed?
A constant delay may come from track alignment, trimmed silence or an export/editing step. Compare a downloaded file with the editor preview before deciding whether the model caused it.
Can audio drift be fixed without regenerating?
Sometimes. Check source and export durations, unintended speed changes, cuts and how the editor handles timestamps or variable frame rates. A single track shift will not fix a gap that grows over time.
Does changing FPS automatically fix lip sync?
No. A constant-frame-rate working copy may help an editor handle timing consistently, but conversion is not a universal fix. Preserve the originals and test a short export.
How much lip-sync delay is noticeable?
It depends on the footage, playback conditions and whether audio leads or lags. Rather than relying on one universal threshold, compare visible speech cues at the start, middle and end, then test the exported clip at normal speed.
Related reading and sources
For more technical background on synchronizing speech and video across challenging conditions, see Disney Research’s DiVAS paper. This article is a practical diagnostic guide, not a benchmark of competing lip-sync generators.
The practical takeaway
Measure the error before spending credits. If the delay stays constant, test a precise track shift. If it grows, investigate durations and editing changes. If the generated mouth shapes are wrong, fix the source conditions or regenerate only the affected segment where possible. Keep the original files, change one variable at a time, and judge the exported result.