Veo 3.1 Native Audio: How to Prompt Sound and Speech

You write a prompt for a barista handing over a coffee and saying "Enjoy your day." The clip looks great. The audio is a vague café murmur, the line is missing, and a cheerful music bed you never asked for is playing underneath. So you regenerate, pay again, and get a different wrong result.

That's the usual first run with Veo 3.1. The model makes sound and picture together, but it only has your words to work from. If the prompt describes the picture and says nothing about sound, the model fills the gap with guesses. This post covers how to write the audio part so the guesses get rarer.

What "native audio" means

Native audio means the sound is generated in the same pass as the video. You don't record a voice-over and sync it later. Google describes Veo 3.1 as producing "richer native audio, from natural conversations to synchronized sound effects". Google's own announcement lists Veo 3.1 and Veo 3.1 Fast, plus features like up to three reference images, first and last frame control, and scene extension.

Two practical consequences:

• Audio is part of the prompt. There's no separate audio step to fix things later.
• Audio and picture can fail together. A mouth that doesn't match the line, or a sound with no source on screen, is a generation problem. You solve it by re-prompting, not by editing a track.

One availability note, as of October 23, 2026: Google's Gemini API documentation says the Veo 3.1 preview there has been shut down, and points developers to a different Google model. That affects people calling Google's API directly. It doesn't change how the prompting below works, but model availability moves, so check the model list in whichever tool you use.

The audio syntax Google recommends

Google Cloud's Ultimate prompting guide for Veo 3.1 gives a simple convention. Each kind of sound gets its own label:

• Dialogue: put speech in quotation marks, attributed to a speaker. Example from the guide: A woman says, "We have to leave now."
• Sound effects: use an SFX: label. Example: SFX: thunder cracks in the distance
• Ambient noise: use an Ambient noise: label. Example: Ambient noise: the quiet hum of a starship bridge

The same guide suggests a visual formula: cinematography, subject, action, context, then style and ambiance. Audio slots in after that. A full prompt then reads like this:

Medium shot, a barista in a small sunlit café slides a paper cup across the counter to a customer and smiles. The barista says, "Enjoy your day." SFX: ceramic clink, milk steamer hiss. Ambient noise: low café chatter, no music. Warm, natural light, shallow depth of field.

Notice what's in there: who speaks, the exact words, two specific sounds, and a background. That's about 50 words and every one of them is doing a job.

A repeatable prompt template

Use this order and fill every slot, even if the answer is "none."

1. Shot and subject: Close-up of a street vendor flipping skewers over charcoal.
2. Action: one main action, not five.
3. Setting and light: Night market, string lights, light drizzle.
4. Dialogue: The vendor says, "Two for five, fresh off the grill."
5. SFX: SFX: sizzling fat, tongs clacking.
6. Ambient noise: Ambient noise: distant crowd, rain on a tarp.
7. Music: say no music or name a mood. Silence on this line invites the model to add some.

If you want a more cinematic result, borrow vocabulary from our guide to camera movement prompts and keep the motion description concrete, as covered in describing the motion, not just the scene.

Writing dialogue that actually lands

Speech is where most people lose generations. A few habits help.

Keep lines short

Clips are short, so a long speech has to be rushed or cut. I'd aim for one sentence, or two brief ones, per speaker. I haven't found a published word limit for Veo 3.1, so treat that as a rule of thumb from experience, not a spec. Test your line and see.

Attribute every line

The man says, "..." is better than a bare quote. When two people are in frame, name them by something visible: The woman in the red coat says, "...", then The man with the umbrella replies, "...". Unattributed quotes are how you get the wrong mouth moving.

Spell things the way they should sound

Brand names, acronyms and unusual words may be pronounced oddly. Write the name phonetically if it matters, and listen before you use the clip. Don't assume a product name will come out right.

Don't ask for on-screen text and speech to match

Generated text inside video is unreliable. If you need a readable caption, add it afterwards. For stills, see how to get readable text inside AI images.

Sound effects and ambience: be specific, be visible

Vague audio words ("realistic sounds") give you nothing to work with. Name the object and the action:

• Weak: SFX: city sounds
• Better: SFX: a bus door hissing open, footsteps on wet pavement

Tie sounds to things the viewer can see. A sound with no source in frame is more likely to drift or get ignored. And separate foreground from background: SFX are the events, ambient noise is the bed under them. Mixing them in one sentence makes both muddier.

Control the shot so the audio has something to match

Audio quality depends on the picture being clear. Some tips:

• Start from a good frame. If you animate a still, pick one with a visible face and a clear action. Our post on picking a start frame that animates well covers this.
• Plan transitions. For a controlled opening and ending, see first and last frame video. Google's guide notes that Veo 3.1 supports timestamp prompting, such as [00:00-00:02] Medium shot..., to assign actions to timed segments.
• Match the frame shape to where it will run. Vertical for Reels and Shorts, wide for YouTube. See aspect ratio for social media.
• Keep the same person across shots. Voice and face drift between separate generations. Our guide to keeping a consistent AI character explains the options.

Mistakes that waste credits

• Writing audio as an afterthought. One tacked-on word like "with sound" does little. Use the labels.
• Cramming. Three speakers, five SFX and a music cue in a few seconds will likely collapse into noise. Pick the one or two sounds that matter.
• Contradicting yourself. Silent empty street plus Ambient noise: crowd chatter is a coin flip.
• Using negative phrasing for sound. Google's guide does show describing exclusions clearly, such as "no buildings or roads", but a plain positive instruction like no music is easier to follow than a list of banned sounds. How well any model obeys negatives varies; see negative prompts: when they help and who ignores them.
• Judging on one run. Results vary from run to run. Generate two or three takes of the same prompt before you change the wording.

Limits worth knowing

Be realistic about what this can do.

• It isn't a mixing desk. You can't set levels, pan sounds or isolate the voice from the music. If you need clean separate stems, record or generate voice and music separately.
• Run-to-run variation is real. The same prompt can give a perfect line once and a mumble next time.
• Some Google editing features skip audio. Google's guide states that its add/remove object feature "does not generate audio" and that these features are "experimental and actively improving."
• Exact pronunciation and accents aren't guaranteed. Always listen with headphones before publishing.
• I couldn't verify clip length or resolution from the pages I read for this post, so check the limits shown in your tool before you plan a script around them.

A cheap way to work

Audio-capable video generations cost more than a still, so test cheap first. In Mediafy Studio, pick Video, choose a model such as Veo 3.1 or Veo 3.1 Fast, and write your prompt. The Enhance button can expand a short prompt, and you can add the SFX: and Ambient noise: lines yourself afterwards. The credit cost shows on the Generate button before you run it, and failed generations are refunded automatically. Credits are pay-as-you-go; see how Mediafy credits work for details.

A sensible workflow:

1. Draft the idea with a faster, cheaper model or a still image to check the framing.
2. Run the full prompt on Veo 3.1 Fast to check the dialogue and sound.
3. Rerun the best prompt on Veo 3.1 only if the Fast result is close and you want to push quality.

If you want a structured path through all of this, the AI Creator Lab course goes deeper on video workflows.

Want to try the barista prompt above? Sign in with Google at Mediafy Studio and run it.

Sources

• Google Developers Blog, "Introducing Veo 3.1 and new creative capabilities in the Gemini API": https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/
• Google Cloud Blog, "Ultimate prompting guide for Veo 3.1": https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
• Google AI for Developers, "Video Generation in the Gemini API": https://ai.google.dev/gemini-api/docs/video