Veo & FlowPrompting basics

Veo 3 Prompting: Dialogue, Sound Effects and Ambience

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

Veo 3 generates audio with the video, so your prompt needs a sound layer as well as a picture layer. Describe the shot first, then add a sentence for ambience, one for sound effects tied to visible actions, and, if someone speaks, the exact line in quotation marks with who says it and how. Keep lines short and say "no music" when you plan to add your own.

The thing that sets Veo 3 apart from most video models is that it hears the scene as well as sees it. Footsteps land when feet touch the ground, rain sounds like rain on the surface it hits, and a character can say a line with lip movement to match. That only happens reliably when you ask for it clearly. This guide covers how to lay out a Veo 3 prompt in Google Flow so that picture and sound are both specified, how to write dialogue the model can perform, and what to do when the audio goes wrong. Flow’s interface and model options change often, so where you see a different label, trust what is on screen.

Two layers in one prompt

Think of a Veo 3 prompt as a picture layer followed by a sound layer. The picture layer is the familiar shot description: subject, action, setting, camera, light. The sound layer is new, and it has three parts, which you can include or leave out as needed.

  1. Ambience. The continuous background of the place: "quiet office hum, a distant printer, muffled traffic through glass".
  2. Sound effects. Sounds tied to actions you have described in the picture layer: "the mug clinks as she sets it down". If the action is not in the picture, the sound has nothing to attach to.
  3. Dialogue. A short line, in quotation marks, with the speaker and delivery: "She says, calmly: 'We should have left an hour ago.'"

Put the sound layer at the end of the prompt, after the picture is fully described. Starting with sound tends to produce clips where the camera and action are vague.

Writing dialogue Veo 3 can perform

Dialogue is the most impressive thing Veo 3 does and the easiest to get wrong. These rules come from the kind of errors that cost credits.

  • Keep it short. A clip of 8 seconds (Veo 3.1's longest; Gemini Omni Flash goes to 10) has room for one line of roughly six to twelve words. Two characters exchanging lines is possible but less reliable than one.
  • Quote the exact words. "A man says something about being late" produces improvised, mumbled speech. "He says: 'We’re late again, aren’t we?'" produces that line.
  • Name the speaker by description. "The woman in the red coat says" rather than "she says" when more than one person is in frame.
  • Give a delivery note. "whispering", "cheerfully", "out of breath", "with a slight laugh". One note, not three.
  • Specify language and accent only if it matters, and in plain words: "speaking British English". Avoid imitations of real people.
  • Frame the speaker’s face. Lip sync is only visible in medium shots and close-ups. In a wide shot, the voice can read as off-screen, which is sometimes exactly what you want.

If you need a sequence of lines, generate one line per clip and join them in Scenebuilder or an editor, keeping the character description identical each time. The Ingredients and Scenebuilder guide covers keeping the speaker’s face consistent across those clips.

Sound effects and ambience that fit the picture

Veo 3 is good at matching sound to visible action when the action is specific. "Footsteps" is vague; "boots crunching on gravel" tells the model both the surface and the shoe. A few patterns that work well:

  • Describe the surface and material: rain on a tin roof, a wooden door creaking, a glass bottle rolling on concrete.
  • Describe distance: "a train passing far away" versus "a train roaring past close by".
  • Describe the room: echoey stairwell, carpeted bedroom, open field. Reverb follows.
  • Say what is absent: "no music", "no voices". This is one of the few places where a negative instruction is useful, because it stops the model adding a score you will have to remove.

Music is the weak point. Asking for a specific genre or a named piece gives unpredictable results and may be declined. If a soundtrack matters, ask for "no music" and add your own track afterwards.

Settings at a glance

Sound is mostly controlled in the prompt, but a few interface settings affect it.

SettingWhat it doesStart with
ModelVeo 3.1 (Lite, Fast, Quality) or Gemini Omni Flash; Omni Flash adds voice references and custom voices for clips made with ingredientsCheck the model menu and credit cost before generating dialogue
ModeText to Video gives the most control over who speaks; Frames to Video needs the speaker’s face in the start frameText to Video while learning dialogue
Aspect ratioFrame shape; vertical frames suit single-speaker pieces for social16:9 for scenes, 9:16 for a talking-head feed post
Outputs per promptNumber of clips generated; dialogue delivery varies between them2 for dialogue so you can pick the better performance
ExtendContinues an 8-second Veo 3.1 clip (generated with Veo 3.1 Lite); the next line goes in the extension promptOne new line per extension

Common sound problems and fixes

  • The character speaks the wrong words. Shorten the line and put it in quotation marks on its own sentence. Remove any other spoken text from the prompt, including signs or names.
  • Unwanted music. Add "no music" at the end of the sound layer.
  • Dialogue but no lip movement. Move to a medium shot or close-up and say the person is "speaking to camera" or "speaking to the man opposite".
  • Subtitles or captions appear. Add "no on-screen text". Veo sometimes adds captions when it detects dialogue.
  • Ambience is generic. Name two or three specific sounds instead of a mood word like "peaceful".
  • Two voices blur together. Give each speaker a distinct description and delivery, and consider one line per clip.

Try it

A woman in her sixties in a green apron stands behind the counter of a small bakery in early morning light, dusting flour from her hands. Medium close-up, camera static, warm window light from the left. She looks up and says, warmly: "You're early. The sourdough needs ten more minutes." Sound: a soft oven fan, a tray being set down on a wooden counter, faint street noise through the door, no music.

Change only the line of dialogue and the delivery note on your second run, keeping the picture and ambience fixed. Then try the same prompt as a wide shot to hear how the voice shifts from on-camera to room sound.

Frequently asked questions

Can Veo 3 generate dialogue with lip sync?
Yes. Veo 3 generates speech synchronised to the picture when you write the exact line in quotation marks, name who says it and frame the speaker’s face in a medium shot or close-up. Keep each line to roughly a dozen words for the best results.
How do I stop Veo 3 adding music to my clip?
End your prompt’s sound description with the words "no music". It is one of the few negative instructions that works consistently, and it leaves you free to add your own licensed track in an editor afterwards.
Why does Veo 3 add subtitles to my video?
The model sometimes renders captions when it detects dialogue in the prompt. Add "no on-screen text" to the prompt and regenerate; if captions persist, try a different output from the same prompt.

Nothing you makeis lost.

Free · No card · Read-only · 2-minute setup

DreamdriveFree plan · Chrome
Add to Chrome — it’s free