Answer: Kling 4.0 generates an audio track alongside the picture, and the prompt is the only mixing desk you get. Describe ambience, the specific events that belong to the action, and — most usefully — the sound the shot ends on. A silent prompt does not produce silence; it produces whatever the model associates with the scene.
That last sentence is the whole argument for writing audio. If the prompt says "a maintenance robot runs through a rainy station", the model will pick some ambience. If the prompt says "begin with rain, quick mechanical footsteps and the low hum of the train, end on the small click of the lunchbox latch", it picks the arc you asked for.
The three-part structure that works
- Bed. The continuous layer: room tone, wind, rain, traffic, a server room hum. One clause.
- Events. The sounds tied to specific actions, in order: footsteps, a splash, a door, a splash of water. Name them as they happen.
- Landing. The last thing the audience hears. This is where a clip stops sounding generated and starts sounding cut.
A template:
Begin with [ambience]. At [beat] you hear [event]; [second event] as [action].
End on [the sound the cut lands on]. No music, no dialogue.Three published examples, dissected
The lunchbox. "Begin with rain, quick mechanical footsteps and the low hum of the train. End on the small click of the lunchbox latch." Three sentences, one arc: weather, character, payoff. The click is a small sound, and naming it small is what makes the model place it instead of burying it.
The dance. "The drum and landing hit together." That is not an ambience choice — it is a sync instruction, and it tells the model that the audio and the picture share a beat.
The paper city. "Layer the rush of the paper river, the creak of folded card and the three tiny sets of footsteps at the end." Three layers, three scales, and the last one is a detail most models would skip: tiny footsteps.
What the prompt cannot do
- Fine mixing. No level control, no stems. If the ambience is too loud, the answer is a re-render with different wording.
- Exact dialogue sync. You can ask for a line, but frame-accurate lip sync is not something the prompt buys; place dialogue in post if the mouth matters.
- Removing audio afterwards. Once rendered, the track is in the file. Generate a second take with "room tone only" instead.
Directing with absences
Negative audio instructions are surprisingly effective because they remove the model's default impulses:
| Instruction | Effect |
|---|---|
| "no music" | stops a score appearing under a scene that does not want one |
| "no dialogue" | removes invented voices, which is the most common unwanted element |
| "room tone only" | the closest thing to a silent plate |
| "no lettering or titles" | also removes any on-screen text the model would invent |
Pairing a positive bed with one or two absences is more reliable than a long list of prohibitions.
When to render silent and dub later
- The shot needs licensed music. Generate with "no music" and add it in the edit.
- Dialogue must be exact. Write "no dialogue", render, then record and place.
- You are iterating on the picture. A silent draft is cheaper to judge: you are looking at blocking, not at a soundtrack you will replace.
FAQ
Can I hear the audio before paying? No — the track is generated with the render; the quote covers the whole job.
Does audio cost extra? No. The per-second price covers picture and sound.
Can I ask for a specific song? No. Genres, instrumentation and tempo survive; named copyrighted tracks do not.
Do other Kling models generate sound? Audio support differs by model and mode; check the controls on the model page before budgeting a sound-led piece.
The six motion studies on Callirra's Kling 4.0 page each end with a sound instruction — reading three of them is faster than reading this post twice.