Sound in your video: voices and lip-sync

A clip can get its sound three ways: the model makes it with the picture, the model speaks in a voice you give it, or a face is lip-synced to a voiceover you made. To lay a finished voiceover or song under clips as it is, see Put a voiceover or a song in a video.

Sound the model makes

In the model menu on Video, the models badged Makes sound make dialogue, effects and ambience with the picture: the Seedance models, Kling 3.0, Kling 4.0, Kling 4.0 Flash, Kling O3 (reference), Vidu Q4, Wan 3.0, Wan 3.0 Prime, MiniMax H3, Grok Imagine 1.5, PixVerse v6 and PixVerse v6 Max. Where sound can be turned off, the bar has a Sound switch: switch it to turn sound on or off. Point at it to see whether sound costs more on that model, and the price on Generate follows it. On Wan 3.0, Wan 3.0 Prime and Grok Imagine 1.5 sound costs nothing extra, and PixVerse v6 and PixVerse v6 Max make it only if you switch it on, for about a third more. Kling 4.0 and MiniMax H3 make sound with every clip, so their Sound switch shows on and can't be turned off. Say in the prompt what should be heard, with spoken lines in double quotes.

On the Seedance models a Music switch sits beside Sound, while Sound is on. While it's off the model adds no background music of its own: the clip has only the voices and sounds your prompt asks for. It's off to start with on Video; switch it on to let the model add music. On a board card it starts off when only voices are wired in or the card says a quoted line from its Start frame, as boards always have, and on when a song or a clip's sound is wired in. It never changes the price.

Kling O3 makes no sound with a video in; the switch says so instead.

On a Creative Board, a clip card that opens on a picture, with a line in double quotes in its prompt, is told to say that line, on Seedance, Kling 3.0, Kling 4.0 and Kling 4.0 Flash; the other models that make sound get the line as you wrote it. The card warns you before Run if its model makes no sound, if Sound is off and the line would be mouthed rather than heard, or if the line won't fit the clip at a natural pace, about two and a half words a second.

A voice, a song or a sound for the clip

Models that take Audio refs learn from recordings you give them, and your prompt says what each is for: a voice the clip speaks in, a song or a bed of sound under it, or a sound effect. They make their own soundtrack from them, so a recording isn't played in the clip as it is, and a voice's words aren't repeated: write what is said in double quotes, and the clip says it in that voice. These are the Seedance models, Vidu Q4, Wan 3.0, Wan 3.0 Prime and MiniMax H3; Vidu Q4 and MiniMax H3 take voices. Kling 4.0 takes no audio refs for now.

1. Pick one of those models on Video.
2. On Reference to Video, add the recording as a sound in the references box: a voiceover or song you made, from the Made tab, or an MP3 or WAV up to 15 MB. A voice note from your phone, such as an M4A or a WhatsApp OGG, goes in as an MP3.
3. Name each in the prompt as @Audio1, @Audio2 and so on, in the order you added them, and say what it is for: "she speaks in the voice of @Audio1", "@Audio2 plays softly under the scene", "@Audio3 is the door slamming". Put what is said in double quotes.

The line under the box says how many recordings the model takes and how long each may be, such as "Up to 3 · each 2–15 s · 15 s in all" on Seedance 2.0. A longer one goes in marked, and the scissors on its tile trim it. Audio refs go beside an image or video ref, except on Wan 3.0 and Wan 3.0 Prime, which take a sound on its own. They add nothing to the price.

If you upload a recording, the bar asks whose voice is in it: a real person, named, with their consent confirmed, or no one speaking or singing. Recordings made on Vessero aren't asked about.

On a board, drop a voiceover card, a music card or a clip's sound on a clip card's Audio refs port to do the same. On Seedance and Wan 3.0, a sound the card's prompt names with @ is used just as the prompt says, and the tag on its wire reads Your prompt says how. One the prompt leaves out is used as boards always have: a voice as the voice the clip speaks in, with a reminder to write the words in quotes, and a song as the background song.

The card checks each sound against the model's lengths before you run it. On Vidu Q4 that's 3 to 12 seconds each, 36 seconds in all. A longer one is named on the card with Trim to 12 s, and a run is refused, with nothing charged, until it fits. A recording that isn't an MP3, such as a WAV, goes to Vidu Q4 as an MP3 we make from it, since that's the only type Vidu Q4 accepts.

To have the clip say the exact words of a recording, in the recording itself, lip-sync it instead: see below.

Lip-sync a face to your voiceover

A talking clip is a face from a picture, lip-synced to a voiceover, and as long as the voiceover. It's made on a Creative Board:

1. Pick a talking model in a video card's model menu: Kling Avatar, Kling Avatar Pro or OmniHuman 1.5. Only these show a Voiceover port, since only they can lip-sync one.
2. Make the voiceover on a Voiceover card, or bring one in from your Library with Add to a board.
3. Wire it into the card's Voiceover port, and the picture of the face into its Start frame.
4. Run the voiceover first, if it isn't made yet: the clip's length, and so its price, comes from it.
5. Run the clip.

A video card on any other model has no Voiceover port. A voice wired into its Audio refs lends the clip its voice, and the words come from the prompt.

Kling Avatar and Kling Avatar Pro take voiceovers of up to 15 seconds, and OmniHuman 1.5 up to 59. A talking clip is priced by the started second of its voiceover: 5 credits a second on Kling Avatar, 10 credits on Kling Avatar Pro and 13.5 credits on OmniHuman 1.5.

Only a talking model lip-syncs a voiceover. If the card is set to any other model, it says so and won't run: pick a talking model, or move the voice to the card's Audio refs to use it as the speaker's voice, with the words written in quotes.

Re-voice a clip you have

To match a clip's mouth to new words, put the clip on a board and wire it into a video card's Video refs port. Under With the clip, choose Lip-sync it, or start from the Lip sync card in the board's Add a card menu. Then wire the voiceover, or another clip's sound, into its Audio port.

• Choose how the clip is read: One person, facing camera, which is faster, or A busier scene, which handles cuts and finds who is speaking.
• The result lasts as long as the sound, up to five minutes, and is priced by the sound's started seconds.
• It reads no prompt: what the clip says comes from the sound.

Photos and recordings of real people need their consent. On a board, an uploaded clip or recording is asked about before a card uses it this way: see Real people: faces, voices and consent.