How to prompt Veo 3.1, audio included

Veo 3.1 generates dialogue, sound effects, and ambient noise natively in the same pass as the video, and Google's own prompting documentation is explicit that audio needs its own labeled clause in the prompt, not a vague aside . That's the single thing that matters most here: treat audio intent as a fourth structural element alongside subject, camera, and style, the same discipline a shot list applies to picture and sound separately. Everything below builds from Google's own documented syntax, not a reverse-engineered guess at one.

The prompt anatomy that works

Google's Veo 3.1 prompting guide lays out a five-part formula: cinematography, subject, action, context, and style & ambiance . Audio isn't a separate slot in that formula: it lives inside style & ambiance as its own labeled sentence. Build every prompt from these clauses, in this order:

1. Cinematography: the camera work and shot composition (dolly shot, tracking shot, wide shot, close-up).
2. Subject: the main character or focal point.
3. Action: what the subject does.
4. Context: the environment and background elements.
5. Audio intent: dialogue, SFX, and/or ambient noise, each on its own labeled clause.

A labeled example, built from Google's own documented syntax:

Cinematography: slow dolly in, eye level, medium shot
Subject: a barista behind a espresso machine
Action: pulls a shot, steam rising past their face
Context: a quiet café, morning light through a front window
Dialogue: none
SFX: espresso machine hiss, cup clink against saucer
Ambient noise: faint café chatter, low

Google's own documentation also flags negative prompts as a separate lever: describe what you want excluded as a positive statement rather than a negation, since "a desolate landscape with no buildings or roads" is the documented pattern, not "no man-made structures" . That's a documented syntax rule, not a style preference: Veo's negation handling is reported to work better on restated positives than on literal "no X" phrasing.

Eight worked examples, easy to hard

Each template is something to adapt, not a script to copy verbatim.

1. Product B-roll with ambient sound

Cinematography: static shot, eye level, close-up
Subject: a pair of wireless earbuds in their case
Action: the case lid opens slowly, earbuds lit from inside
Context: a marble countertop, soft daylight from the left
Dialogue: none
SFX: soft mechanical click on open
Ambient noise: quiet room tone, no music

This is the template to start from before adding dialogue: it isolates camera and ambient-sound phrasing as the only two variables in play. No speech means no lip-sync to get wrong, which is why it's example one rather than example five. Test this prompt with Veo 3.1 Fast on Infer →

2. Vertical cut of the same shot

Cinematography: static shot, eye level, close-up, 9:16
Subject: a pair of wireless earbuds in their case
Action: the case lid opens slowly, earbuds lit from inside
Context: a marble countertop, soft daylight from the left
Dialogue: none
SFX: soft mechanical click on open
Ambient noise: quiet room tone, no music

Google's Gemini API docs list 9:16 as one of Veo 3.1's two supported aspect ratios alongside the 16:9 default , and Infer's own documentation notes vertical and square renders run about 10-15% slower than 16:9 at the same per-second price. Budget the extra wait, not extra spend, when a client asks for a TikTok cut of a 16:9 shoot.

3. Dialogue scene

Cinematography: static shot, eye level, medium close-up
Subject: a woman in a kitchen
Action: explains a recipe step to camera, mid-sentence pause before the last word
Context: soft window light, front-left
Dialogue: "...and then you let it simmer," she says, warm and conversational
SFX: none
Ambient noise: quiet kitchen room tone

Google's own dialogue syntax is a full quoted line attributed to a speaker, not a bare instruction to "add speech" . Google's own model page is candid that this is an evolving capability, not a solved one: "creating videos with natural and consistent spoken audio, particularly for shorter speech segments, remains an area of active development" , so treat a single short line as the safer bet over a long exchange.

4. Sound-effect-led action scene

Cinematography: tracking shot, low angle, wide to medium
Subject: a classic convertible
Action: drives along a coastal road, then brakes hard at a curve
Context: waves crashing against rocks below, late afternoon light
Dialogue: none
SFX: tires screeching, engine roaring, brief silence after the stop
Ambient noise: ocean waves, wind

Google's SFX syntax calls for naming the sound and its trigger explicitly — "tires screeching loudly, engine roaring" is the documented example phrasing . Naming the screech at the same beat as the brake ("then brakes hard," "tires screeching") ties the two together instead of leaving Veo to guess the timing.

5. Chained segment: continuing the story

[Segment 1, 8s]
Cinematography: wide establishing shot, static
Subject: a hiker
Action: reaches the top of a ridge at sunset
Context: warm backlight, long shadows
Ambient noise: wind, distant birdsong

[Extension, +7s]
Cinematography: slow push-in to medium
Subject: same hiker
Action: sits down, looks out at the view
Context: same ridge, light fading slightly
Ambient noise: wind continuing, footsteps settling on rock

Veo 3.1's documented extension path adds roughly 7 seconds to a prior Veo-generated clip per call, repeatable up to 20 times . Each extension call takes the previous segment as its input, so continuity phrasing ("same hiker," "same ridge") in the new segment's prompt matters more here than in a standalone 8-second clip: there's nothing else telling the model this is the same scene.

6. Chained toward the ~148-second ceiling

Eight seconds from the initial generation plus 7 seconds per extension, 20 extensions deep, lands at 8 + (7 × 20) = 148 seconds, the exact math behind the "chainable to ~148s" figure on Infer's own model page. Structure a long-form sequence as a table before you generate a single clip:

Plan the beats before the first call, since each extension only sees the clip immediately before it, not the whole sequence's script.

7. Image-to-video from a starting frame

Input image: [product hero shot, earbuds case on marble, provided]
Cinematography: static shot, close-up
Subject: the case from the input image
Action: steam drifts through frame from an unseen cup, off to the side
Context: match the input image's lighting and composition
Dialogue: none
SFX: none
Ambient noise: faint café murmur, low

Google documents Veo 3.1's image-to-video mode as using the supplied image as the first frame and animating forward from it . Everything after "context" still applies to what happens past frame one; only the opening frame is anchored to the reference.

8. Reference images ("Ingredients to Video")

Reference images: [1: product, 2: model's outfit, 3: background style]
Cinematography: medium, eye-level shot
Subject: the person from reference 2, wearing the outfit shown
Action: walks toward camera through the setting in reference 3
Context: consistent with reference 3's palette and lighting
Dialogue: none
SFX: footsteps
Ambient noise: light outdoor wind

Google calls this its "Ingredients to Video" workflow: reference images of a scene, character, object, or style, carried into the generation to keep the same subject consistent across shots, and Google states the feature "now includes audio generation" . The Gemini API docs cap this at three reference images per call , fewer than the five some rival models allow, so give each of the three a distinct job rather than splitting one concept across all three.

Try Veo 3.1 Fast in the Infer playground →

What doesn't work

Long or multi-person dialogue is the least reliable audio type. Google's own DeepMind page states plainly that "creating videos with natural and consistent spoken audio, particularly for shorter speech segments, remains an area of active development" . The fix isn't a cleverer prompt: keep dialogue to one short line per 8-second segment, and treat a full back-and-forth conversation as something to storyboard across multiple chained segments rather than one generation.

Negating with "no X" phrasing is documented to underperform. Google's own guide recommends restating the exclusion as a positive description instead . If a generation keeps including something you don't want, rewrite the context clause around what should be there rather than stacking negations onto it.

Vertical and square renders take longer, and that's not a prompt problem. Infer's documentation notes 9:16 and 1:1 renders run about 10-15% slower than the 16:9 default at the same price. No amount of rephrasing closes that gap; it's a rendering-time property of the aspect ratio, not the prompt.

Extending past the documented cap doesn't work. The extension feature tops out at 20 calls past the initial 8-second clip, landing at roughly 148 seconds total. There's no documented way to extend a clip that's already at that ceiling, or one that wasn't itself generated by Veo: extension inputs have to be prior Veo output.

Settings that matter

Resolution and duration are the two levers with real cost consequences on Infer. Veo 3.1 Fast tops out at 1080p at $0.09-$0.10 per second, with no separate audio surcharge: you're billed the same rate whether the clip is silent or carries dialogue, SFX, and ambient noise together. An 8-second clip, the native generation length, runs $0.72-$0.80; chaining toward the ~148-second ceiling runs $13.32-$14.80 at that same per-second rate. Full worked math across every provider, including where Google's own direct API and fal.ai price the same clip higher, is in the Veo 3.1 pricing guide.

Aspect ratio is the other setting worth planning around before you generate: 16:9 is the default and the fastest to render, 9:16 costs the same per second but takes about 10-15% longer. If a batch of vertical clips is on a deadline, generate the 16:9 version first to confirm the prompt works, then switch aspect ratio for the final render rather than debugging in the slower format.

Infer hosts every model on this leaderboard, Veo 3.1 Fast included, and has no reason to undersell a documented limit just to make the sales pitch cleaner; the dialogue caveat above is Google's own words, not ours.

Run your next prompt through Veo 3.1 Fast on Infer →

Related reading: the guides hub, how to prompt Seedance 2.0, Kling 3.0 Pro vs Veo 3.1 Fast, Seedance 2.0 Pro vs Veo 3.1 Fast, Sora 2 vs Veo 3.1 Fast, and best AI video models with native audio if audio is the reason you're comparing models at all.