If an H3 Max character looks as though they are saying the wrong thing, listen before rewriting the prompt. Did the voice say the wrong words, or did the mouth fail to match words that were correct? Those are different problems. Moving the audio earlier will not repair a missing word.
The useful first step is to identify what you would reject in the clip—not to add “perfect lip sync” to an already crowded prompt. Here is a way to review a dialogue take and choose a smaller, more specific next attempt.
This guide concerns the H3 Max workflow on AI Live Generator. We use the Turbo text-to-video and image-to-video routes. The review process and prompt below are recommendations, not findings from a measured lip-sync test.
Before you judge the mouth, check what you can hear
Play your generated clip with its player volume turned up. Our looping example videos start muted; that does not mean a generated result has no audio. The result player has its own controls. Also check whether your browser tab or device is muted.
If playback is silent or seems delayed, download the result and play the same file in another player. If the issue disappears there, do not spend credits regenerating a file that already plays correctly. If it remains, you have narrowed the problem to something beyond that first player; you have not yet proved the model was the cause.
When you can hear the voice, review the clip three ways:
- Listen without looking. Write down the words you actually hear, including extra syllables and unfinished phrases. Compare that with the requested line.
- Watch with the sound off. Notice when speaking starts and stops. Does the face remain visible? Does the mouth keep moving through what should be a pause?
- Watch and listen together at normal speed. Check the beginning, middle, and end of the line. Is the whole voice apparently late, or does the match break only on particular words?
The separate passes stop an attractive face from distracting you from a wrong sentence. They also stop a wrong sentence from being mislabeled as a timing error.
Choose the next change from the failure you found
The line is wrong, extra words appear, or the ending is cut off
Treat this as a dialogue-content problem first. Put one short line in quotation marks and assign it to one visible speaker. Remove competing narration and background conversations for the next attempt.
Read the line aloud at the pace you want. Include the pause before it and the reaction after it. If that performance does not fit the selected duration, shorten the sentence or choose a longer clip. More seconds provide room; they do not guarantee the model will deliver the script exactly.
Do not replace a missed line with a paragraph of timing instructions. A simple instruction you can evaluate is more useful here than a complicated one you cannot.
The words are present, but the voice is hard to understand
Listen for music, effects, or another voice covering the line. For the next take, remove those competing sounds rather than simply asking for every sound to be “clearer.” Keep a little room tone and make the speech the only foreground sound.
Turning up the playback volume raises the whole mix, not just the dialogue. Our generator does not provide separate speech and music tracks to rebalance. The synchronized-audio guide explains how to give the sound layers different priorities in the prompt.
The voice seems early or late by the same amount
First repeat the same-file playback check above. If a consistent offset survives it, an external video editor may let you shift the audio track against the picture. That is an editing option, not an automatic repair available in our generator.
A single shift only helps when the mismatch really is consistent. If the start improves but the ending gets worse, or the mouth forms the wrong shapes throughout, sliding the entire track is not a complete solution.
Timing varies, or the mouth continues after the line
Make the speaking moment easier to see and judge: one face, a front or slight three-quarter view, no cut during the sentence, and no hand covering the mouth. Ask for a brief closed-mouth pause after the line.
These choices simplify the brief; they are not a lip-sync lock. Raising the output resolution gives you a different visual setting, not a promise of corrected dialogue timing. A successful-looking single retry also does not establish that one wording change fixed the problem reliably.
Give the next take one small job
Imagine a watch repairer explaining that a repair is finished. The scene does not need a long speech, a moving camera, music, and a close-up of tiny gears all at once. Start with the face and one short line.
Here is an original, untested text-to-video prompt to adapt:
Five-second medium close-up of an adult watch repairer seated at a quiet workbench. The camera stays at eye level, with the repairer's face and mouth clearly visible. After a brief still moment, the repairer says once, in a calm natural voice: "It works. Listen." The lips close after the sentence, followed by a small smile. Keep the hands below the face and hold the same framing throughout. Audio: the one speaker close and clear, quiet room tone, then a soft watch tick after the speech. No music, narration, other voices, or cut.The watch tick belongs after the words so it does not compete with them. The closed-mouth finish gives you something specific to inspect. Neither instruction guarantees compliance; both make it easier to say what the result did or did not deliver.
Set the duration control to five seconds as well—writing it in the prompt does not set the control. If you use image-to-video instead, adapt the scene to your uploaded image and choose a starting frame where the mouth is visible. Do not expect an obstructed profile to become a reliable speaking close-up just because the prompt asks for one. See the image-to-video guide for preparing that first frame.
Keep Balanced or Quality prompt expansion unchanged while reviewing a wording change. On this site those are prompt-expansion choices, not speech-accuracy levels. Changing the sentence, framing, duration, resolution, and expansion mode together makes the next result harder to interpret.
Know when another generation is the wrong tool
fal describes H3 Max's audio as generated alongside the picture. That capability is not a reason to accept a line that says something different from your script. The official model overview demonstrates audiovisual generation; it is not a guarantee about your particular sentence.
Our current interface has no audio-upload field, phoneme timeline, voice-cloning control, or lip-sync correction slider. The Turbo text-to-video input schema lists prompt and generation settings, not an editable dialogue timeline. Other H3 products and routes can have different controls.
If an exact spoken product name, approved claim, or recorded performance is non-negotiable, decide how you will secure that audio before spending on repeated takes. One alternative is an externally edited voiceover over a shot without a visible speaking mouth. That changes the creative idea, but removes the need to match that mouth. AI Live Generator does not record or assemble that voiceover for you.
At publication, a five-second generation here uses 15 credits at 480P or 25 credits at 768P. Two new five-second 480P attempts use 30 credits; they are new generations, not edits to the downloaded clip. Check the button's current cost and Pricing before submitting.
Keep the original file. Name the one thing the next take must do better. If you cannot name it, review the clip again before opening another generation.