Google launches Veo 3 with native audio

Background

Generative media models have mostly been organised by modality: one model for images, one for video, one for music. When a generated clip needed sound, the usual route was to make the picture first, produce the audio separately, and match the two afterwards. On 20 May 2025, at its I/O developer conference, Google announced a group of generative media models and tools that take a different approach, including a video model that produces sound as part of the same generation.^[1]

What was announced

The centrepiece was Veo 3, which generates video together with audio.^[1] Two further models were introduced in the same announcement: Imagen 4 for images and Lyria 2 for music.^[1] Alongside the models, Google introduced Flow, an AI filmmaking tool.^[1] Taken together, the announcement covered pictures, sound, and the stage at which generated material is put together into a piece of work.^[1]

How people could use it

Veo 3 was first available to US subscribers of Google AI Ultra, through the Gemini app and through Flow.^[1] Enterprise users could reach it on Vertex AI.^[1] Imagen 4 was available the same day in the Gemini app, in Whisk, on Vertex AI, and across Workspace in Slides, Vids, Docs and other applications.^[1] Generated content was watermarked with SynthID.^[1] SynthID is Google's watermarking system, which embeds a marker in generated media so that it can later be identified as machine-made.

Why it matters

Putting picture and sound into a single generation changes where a production starts. The output is not a silent clip waiting for a soundtrack but footage that already carries audio, so the later work becomes selecting and revising rather than assembling sound from scratch. For people making short films, advertising or social video, that shortens the distance between an idea and a watchable clip.

Flow suggests Google was not only shipping models but also a way of working: a tool aimed at filmmaking, with the video model placed inside a production workflow.^[1] Imagen 4 spread along a different path. On the same day it appeared in the Gemini app, Whisk, Vertex AI and several Workspace applications, which puts image generation inside everyday tasks such as writing documents, building presentations and editing video.^[1]

The watermarking is worth noting too. Generated content was marked with SynthID,^[1] leaving a trace that can be checked later when the question is whether a piece of media came from a model. As generated video and audio become more convincing, that kind of provenance marking gives platforms and viewers a way to judge where material came from.

For Google, the release put video, image and music models plus a production tool out at the same moment,^[1] distributed through a consumer subscription, Google AI Ultra, and an enterprise platform, Vertex AI.^[1] Veo 3's native audio is the part that changes how the work is done; the rest of the package shows how the models were meant to be reached.