Google’s prompt guide gives audio one technique section and dialogue one numbered element. That is the whole official foundation for putting sound into a Veo 3.1 prompt, and most guides write straight past it.
Everything below comes from three pages: the prompt guide, the Veo model page and the Flow announcement. Where a rule you have read elsewhere is not on them, this page says so instead of repeating it.
What Google actually publishes about audio
The three pages say different sizes of thing. The prompt guide carries the method: a section called “Fuse visuals with sound design”, and dialogue as one of seven elements. The model page carries the capability claim and the honest limit. The Flow announcement carries where audio now reaches — “For the first time, we’re also bringing audio to existing capabilities like ‘Ingredients to Video,’ ‘Frames to Video’ and ‘Extend.’”
The capability sentence sits on the model page: Veo adds sound effects, ambient noise and even dialogue, “generating all audio natively”. No menu, no parameters, no switches. The prompt guide covers the structure those three capabilities are written into.
Dialogue is an element, audio is a technique
The guide names seven elements in order: shot framing and motion, style, lighting, character descriptions, location, action, dialogue. Audio is not an eighth. It gets its own technique section instead, which tells you something about how the model was tuned: speech is an instruction you place, ambience is a quality you describe.
The dialogue element itself is two sentences, and both matter:
“Veo 3 can generate dialogue. You can either pick a topic for characters to talk about, or give characters specific things to say.”
A topic is shorter to write and leaves the wording to the model. Exact words give you control and also fix how much screen time the line needs. On an eight-second clip, that second effect is usually the one that decides.
The character element carries a quieter rule. Write “specific and detailed descriptions about each character’s appearance, voice, action, and dialogue” — voice belongs in the character description, not in a separate audio block at the end.
The one official audio sample, read closely
There is exactly one audio example on the guide page, and its shape is worth copying.
A keyboard whose keys are made of different types of candy. Typing makes sweet,
crunchy sounds. Audio: Crunchy, sugary typing sounds, delighted giggles.Read what it does. The visual sentence comes first and contains the sound-producing action — typing. Then a separate audio section names two textures and one human reaction. It does not describe a mix. It defines the sounds the shot should contain.
Google’s instruction around it is one line: define the sounds you want to hear explicitly, so the audio matches the visuals.
Worked examples: two prompts, one with dialogue and one with sound only
Both cases run eight seconds, because the model page prints one length: “Veo videos are 8 seconds long.”
Example one: two lines in one eight-second shot
Framing and motion: medium shot, static camera, slight stop-motion jerk
Style: stop-motion animation, hand-built set, visible clay seams
Lighting: warm single desk lamp from above
Character: a hiker in a red rain shell, mid-thirties, visible stop-motion gait
Location: a forest clearing at dusk, felt trees, cotton-wool mist
Action: the hiker plants a small flag in the moss and turns to camera
Dialogue: Hiker: "This is my spot now." Ranger: "The bear disagrees."
Audio: quiet forest room tone, a flagpole pushed into soft ground, one distant owlBoth speakers are named before their lines, which is the structure Google’s own stop-motion example uses. The constraint is the second line. The model page states that natural and consistent spoken audio, particularly for shorter speech segments, remains an area of active development, so a third exchange is where this prompt starts losing words rather than gaining them.
Example two: sound effects only, no dialogue
Framing and motion: close-up, static camera, slight handheld drift
Style: documentary product footage, 50mm
Lighting: hard key from the right, deep shadow behind
Character: two hands only, no face in frame
Location: a stainless steel worktop
Action: a jar lid turns, releases, and the lid is set down on the steel
Dialogue: none
Audio: a rubber seal breaking once, a metal lid touching steel, quiet kitchen behind bothThe explicit “Dialogue: none” is doing real work. The guide’s instruction is to define the sounds you want, so the absence of speech is worth writing down rather than leaving to chance.
We have not had these two clips checked on a studio monitor, and neither case comes with a claimed result. What is stated above is what each line is targeting in the official documentation.
Inline or its own section: both are documented
The guide settles this in one sentence: you can integrate audio cues into your prompt, or include them in a separate section. The candy keyboard example uses the separate section. The guide’s long rally example does the opposite, describing engine roar and gravel spray inline in the middle of a prose prompt.
The layout below is assembled from the guide’s own element names. It is a structure to write into, not a template Google publishes:
Framing and motion: ...
Style: ... / Lighting: ...
Character: ...
Location: ...
Action: ...
Dialogue: ...
Audio: ...One continuity fact is published, and it changes how you write the audio on a sequence. Extend generates each video from the final second of the previous clip. A loud transient in that last second biases what comes next, so the audio continuity of a scene is decided by how the previous clip ends.
When the audio comes back wrong
Three shapes cover most of it. Silence usually means the prompt described only the picture, which is what the sound-design section exists to fix. A line that arrives clipped or garbled runs into the published limitation on shorter speech segments, and switching from exact words to a topic lowers the demand. Effects that land late or on the wrong object usually come from describing the sound without anchoring it to the visible event, which is the one thing both official examples do.
The syntax third-party guides teach that Google does not publish
Bracketed tags such as a square-bracket SFX label, a 48 kHz sample rate, a word ceiling per eight seconds, millisecond lip-sync offsets, volume values and negative audio prompts all circulate as if they were specification. None of them appears on the prompt guide, the model page or the Flow announcement. They may be a workable personal habit, and this page does not test them either way. The point is narrower: they are not published by Google.
Music deserves the same caution. The three official pages name sound effects, ambient noise and dialogue. Music is not one of the seven elements and is not named as a capability, so any guide that teaches a music field is teaching its author’s shorthand.
What Google has not published about audio
No sample rate or channel layout. No cap on words or lines per clip. No dialogue language list. No negative audio syntax. No volume or level control. No music element. No documented switch for turning audio off.
Against that, one limitation is stated plainly, and it is the sentence most audio guides skip: creating videos with natural and consistent spoken audio, particularly for shorter speech segments, remains an area of active development. The Runway Gen-4 comparison covers what the other side of this decision publishes, including a per-second price for calling Veo 3.1 through another vendor’s API.
Frequently asked questions
Does Veo 3.1 generate dialogue? Yes. The guide says the model can generate it, and that you can give characters either a topic or the exact words.
Where does the audio go in the prompt? Either place is documented: audio cues can sit inside the prompt, or in a separate section. The official example uses the separate section.
Can I ask for music or a score? Google’s three audio pages name sound effects, ambient noise and dialogue. Music is not among them, so treat score prompting as unpublished.
How many lines fit in one clip? Google publishes no word or line count. The only published length is eight seconds, and short-segment speech is flagged as under development. The getting-started guide takes a first clip end to end, and the API and pricing guide covers the route where audio is billed per second.
Write the picture first, give speech its own line, and name the sounds you actually want. The elements are published; the rest is your taste.