Dialogue needs a performance brief
A spoken scene is more than a sound effect attached to a moving face. The audience needs to understand who is speaking, what the line means and how the other person reacts. Start with those decisions before describing elaborate lighting or a sequence of camera moves. This guide concentrates on dialogue; the sound overview covers general output-audio behavior.
According to ByteDance’s Seedance overview, audio and video are generated together. That explains why the prompt should describe the visible performance and spoken content as one scene. It does not mean that every requested word, accent or mouth shape will be reproduced perfectly. Listen to the result rather than assuming readable lips prove the line is correct.
SeaDance generates sound with the picture, does not support uploading an audio file, and has no silent-generation switch. A written line is a request for a new performance, not a locked recording. If the project depends on exact recorded speech, plan how that recording will be produced outside SeaDance before spending on visual attempts.
Write a single speaker before adding a conversation
Name the person, specify the spoken language, quote the line, and describe delivery outside the quotation marks. Keep ambient sound in a separate sentence. Avoid placing performance directions inside the spoken line, where they could be interpreted as words to say. Use a short line that leaves room for a natural expression afterward.
Original prompt: “A shopkeeper in a navy apron stands behind a tidy counter, facing the customer just beside the camera. Locked medium close-up. The shopkeeper says in English, warmly, ‘I saved the last loaf for you.’ A brief smile follows the line. Quiet shop room tone underneath. End with the shopkeeper looking toward the customer.”
The example defines a speaker, an eyeline, a line and an ending. It does not demand a pause after every word or an unrelated camera flourish. If the delivery feels rushed, shorten the text or choose a supported longer duration. Adding more emotional adjectives usually does less than removing a clause the scene cannot accommodate.
Make turn-taking visible in a two-person scene
Give each person a stable name and a distinct position. State who listens during each turn. Keep dialogue in sequence and avoid simultaneous speech while testing the setup. An over-shoulder composition or a calm shared frame can make the active speaker easier to inspect than rapid cuts between changing faces.
Original prompt: “Mara in an olive jacket stands on the left; Ellis in a charcoal sweater stands on the right beside a workshop table. Locked shared medium shot. Mara says in English, ‘Did you leave the light on?’ Ellis listens without speaking, then replies in English, ‘Only in the workshop.’ Mara listens during the reply. Quiet room tone. End with both looking toward the open doorway.”
This is a proposed prompt, not a verified generated performance. If the voices swap, simplify the visual arrangement and repeat the speaker’s clothing beside the line. If both mouths move, explicitly keep the listener silent during the active turn. If the timing remains crowded, split the conversation into separately planned shots and review continuity at the cut.
Budget for dialogue tests, not guaranteed delivery
| Model | SeaDance resolution | 5-second test, sound enabled |
|---|---|---|
| Seedance 2.0 Mini | 480p | 4 credits / $0.80 |
| Seedance 2.5 | 480p | 11 credits / $2.21 |
| MiniMax H3 | 768P | 9 credits / $1.81 |
| Kling 3.0 | 720p | 13 credits / $2.61 |
All dollar equivalents allocate prepaid credits at the Basic monthly rate: $29.90 for 149 credits, or approximately $0.2007 per credit. They are not separate checkout prices. See pricing for plans and packs. New accounts start with 0 credits. Each new attempt is charged again, including a completed take you dislike; a failed generation returns its generation credits. That is different from a refund of a subscription payment. See the refund policy.
Inspect words, mouths and meaning separately
First, listen for the exact requested words. Note omissions, substitutions, added speech and an unintended language. Next, watch whether the correct face speaks at the correct moment. Then judge whether mouth movement and vocal rhythm plausibly agree. A scene can pass the first check and fail the second, so avoid calling all audio problems “lip sync.”
Keep the reference images and character descriptions consistent when a conversation continues in another shot. A matching jacket does not guarantee a matching voice, and a repeated name does not create a saved voice identity. Use the same written delivery notes as guidance, but treat voice continuity as something to evaluate rather than a lock you have activated.
Kling 4.0 is available through the model directory; check the live quote in the generator. According to Kuaishou’s launch overview, closer lip synchronization is an audiovisual improvement. That is a manufacturer claim, not a guarantee of an exact line or a promise that SeaDance supports uploaded voice tracks.
Fix the smallest failure first
If the wrong person speaks, clarify the named speaker and the listener before changing the visual style. If speech is rushed, shorten the line before requesting several new pauses. If the line is mispronounced, try simpler wording when the meaning allows it. Preserve exact wording when required and recognize when further generation is an unsuitable way to obtain it.
If music competes with speech, request a restrained background and clear dialogue. This is prompt guidance rather than a separate mixer control. If a mouth becomes obscured, ask for an unobstructed face and reduce head motion. Keep hands, cups or turning gestures away from the mouth during the line unless that blocking is essential.
Record the reason for each retry. A completed take with a changed sentence is still a completed charged generation, even if it cannot be used in your project. Do not confuse creative rejection with a failed job. A smaller follow-up test can tell you whether your revision solved the speaker or timing problem before you rebuild the full scene.
When separate post-production speech is suitable
Use a separate recording workflow when exact legal wording, a particular performer, careful pronunciation or repeatable voice continuity is mandatory. You can generate visual material and handle the recording and assembly elsewhere. SeaDance does not provide a soundtrack-replacement control, so plan the required work in another tool.
Replacing speech later does not automatically repair visible mouth movement. It is easier to use a reaction shot, a listener, a distant speaker or an object detail when the recording must change significantly. If a close-up talking face is essential, approve the vocal and mouth performance together before treating the shot as final.
For review, save the final spoken transcript beside the intended line and note every change. Ask someone fluent in the requested language to check pronunciation when it matters. Attractive facial motion is not evidence of correct speech. The acceptance decision should follow the required words and performance, with the visual style evaluated afterward.
Frequently asked questions
Can I upload my own dialogue recording?
No. SeaDance does not support uploading an audio file. The generator creates sound with the picture; an existing recording needs a separate production workflow.
How do I specify the language?
Name the spoken language beside each speaker and put the intended line in quotation marks. Keep staging and emotion outside the quotation marks, then listen to the actual result.
Will quotation marks guarantee exact speech?
No. They clarify the request, but generated dialogue can still omit or change words. Inspect the transcript and the visible speaker before accepting the take.
Is a later voiceover enough to fix lip sync?
Not necessarily. A different rhythm can make a visible speaking face less convincing. Consider visual coverage that does not require precise mouth matching when the recording must change.
Sources
Official background checked November 5, 2026. SeaDance settings and billing in this article follow its current product configuration, rather than the full range of manufacturer features.