Why your AI presenter looks slightly off — and what actually fixes it

Generated mouth movement that looks wrong is the most common complaint about presenter video, and it is almost never the thing people assume. When lip-sync reads as off, the cause is usually one of four things, and three of them are fixable in minutes.

Work through them in this order.

1. The source clip (most common)

If you registered your own presenter, the source footage dominates the result. The requirements are narrower than most people expect:

• Frontal. The face square to the camera. A three-quarter view is the single most common cause of mouth movement that reads as wrong, because the model is inferring a mouth shape it cannot fully see.
• Evenly lit. Mixed lighting — a window on one side, overhead on the other — creates shadows that move with the face and get baked into the synthesis.
• Not occluded. Hair across the face, a microphone, hands, glasses with strong reflections.
• Steady. A moving head or a walking shot adds motion the model has to reproduce.
• Roughly 8–30 seconds. Too short and there is not enough to work with; too long adds nothing.

A clip shot at a desk, facing a window, is the classic bad input: three-quarter angle plus directional light plus a slight head tilt.

2. Speaking rate

Synthesis quality degrades as speech gets faster. Very fast narration, or audio with heavy emphasis and rapid pitch changes, produces mouth shapes that cannot keep up and read as sloppy.

If the script is dense, the fix is usually not to slow the voice artificially — it is that the slide is carrying too much and should be split. See how long each slide should be; a slide that needs three minutes is two slides.

3. Compression

If the mouth looks fine in the source render and wrong after upload, you are looking at a compression artefact. Video re-encoded aggressively, or squeezed into a small player, loses exactly the fine detail around the mouth that makes the movement read as natural.

Check the output at full size before concluding anything is wrong with the synthesis.

4. Uncanny valley, genuinely

Some of it is real. A presenter that is almost right is harder to watch than one that is clearly synthetic, because viewers spend effort working out what is off. If everything above checks out and it still reads wrong, the options are:

• Use a different presenter. Different presenters suit different content and audiences; a mismatch between a very casual presenter and formal compliance material is a real effect.
• Reduce the presenter’s screen size. A smaller presenter window draws less attention to the mouth and puts it on the slide, which is where the information is anyway.
• Drop the presenter entirely for that deck. Voiceover over slides is a legitimate output, and for dense reference material it is often better.

A five-minute diagnosis

1. Check the source clip against the list above. Fix the angle and the light first — this resolves most cases.
2. Listen to the speed. If it is faster than comfortable conversational pace, slow it or split the slide.
3. View the output at full resolution. If it looks fine there, it was compression.
4. Swap the presenter and re-render one page. If a different presenter looks right, it was a mismatch.
5. Only then conclude it is a model limitation and choose one of the options above.

Worth saying plainly

Perfect lip-sync is not achievable, and any tool claiming otherwise is overselling. What is achievable is mouth movement good enough that viewers stop noticing it and look at the slide — which is the actual goal. If viewers are watching the mouth, something upstream went wrong, and it is usually the lighting.