How do you make a video with sound using Veo 3?

Until this month, AI video was silent. You generated the picture, then added music, effects and voices in an editor. Veo 3, which Google announced at its I/O conference on May 20, generates the audio with the video: street noise behind a city scene, birds in a park, and even characters speaking, with their lips moving to the words. That changes how you write a prompt. You are no longer only describing a shot; you are describing a scene, with a soundtrack.

Where to use it

There are two places for most people:

• The Gemini app (on the web, Android and iOS). Choose "Video" in the prompt bar, type a description, and wait for the clip. This is the simplest way in.
• Flow, Google's new AI filmmaking tool, which brings Veo, the Imagen image model and Gemini together. In Flow you can keep the "ingredients" of a story, such as characters, locations, objects and styles, in one place and reuse them from shot to shot.

Businesses can also reach Veo 3 through Vertex AI, Google Cloud's platform for developers.

A worked example: a short scene with dialogue

A good Veo 3 prompt covers five things: the shot, the action, the look, the words spoken and the sounds. Here is one for a short scene at a lakeside dock:

Medium shot, early morning on a calm lake in autumn, mist on the water, red and orange maples on the far shore.
An older woman in a green rain jacket kneels on a wooden dock, tying a canoe to a post. A young man walks up holding two paddles.
He says: "Is it always this quiet?"
She smiles without looking up and says: "Only before the coffee's ready."
Sounds: water lapping against the dock, a loon calling in the distance, footsteps on wet wood. No music.
Natural light, soft colors, steady handheld camera, realistic style.

A few choices in that prompt are deliberate:

1. Dialogue in quotation marks, with a speaker. "He says" and "She says" make it clear who speaks which line. Keep lines short; the clip is only a few seconds long.
2. A separate "Sounds" line. Listing sounds explicitly, rather than hoping the model infers them, gives you more control. "No music" is useful too, if you plan to add your own.
3. One moment, not a story. A single exchange fits a short clip. A whole plot does not.

When the clip comes back, watch it twice: once for the picture, once with your eyes closed for the sound. Common fixes are small. If the voices sound wrong, describe them ("a calm, low voice"). If background noise drowns the dialogue, name fewer sounds. Change one thing per attempt so you can tell what helped.

Using the clips responsibly

Every Veo 3 output carries SynthID, Google's invisible watermark, and Google has launched a SynthID Detector portal to check files for it. From May 30, Google also adds a visible watermark to generated videos, except for clips made in Flow by Google AI Ultra subscribers. Realistic video with real-sounding voices is easy to misuse, so label AI-made clips as such, avoid imitating real people, and don't put words in the mouth of anyone who could be mistaken for a real person.

What are the limits on access and length?

Access depends on your plan and your country. At launch, Veo 3 was only for subscribers to Google AI Ultra in the United States, a new top-tier plan at $249.99 a month that also bundles Flow, higher limits in other Google AI tools, YouTube Premium and 30 TB of storage. On May 30, Ultra arrived in the UK and is now offered in 73 countries, according to 9to5Google. Check Google's own list for your country before subscribing.

Subscribers to the cheaper Google AI Pro plan ($19.99 a month) are getting a trial: 10 Veo 3 videos with sound in the Gemini app, first on the web and now rolling out on Android and iOS. That allowance does not reset; after the tenth video, the app goes back to Veo 2, which makes silent video. Google says the same limited access is coming to Flow for Pro subscribers.

Length is short. The Gemini app describes Veo 3, still labelled a preview, as producing an eight-second video. Longer pieces have to be built from several clips, which is what Flow's scene tools are meant for: keep the same characters and settings as ingredients, generate shot by shot, and assemble them.

Plans and allowances are changing quickly in these first weeks, so treat the numbers here as a snapshot of the end of May and confirm them on Google's plan pages before you pay.