Lyria 3 is a beta feature in the Gemini app that turns a text description or an image into a 30-second music track with generated cover art. It writes its own lyrics, exposes controls for style, vocals and tempo, and stamps every output with a SynthID watermark. Here is what it does well and where the limits sit.
How Lyria 3 works
Lyria 3 is built on Google DeepMind's music generation research. You give it a prompt — text, an image, or both — and it returns a finished 30-second track. Cover art is generated alongside the audio by Nano Banana, Google's image model, so a single prompt produces both the sound and the artwork that goes with it.
Three things sit on top of the prompt:
- Automatic lyrics. The model drafts words from your description rather than asking you to write a verse first.
- Creative control. Style, vocals and tempo can be steered instead of being left entirely to the model's reading of your prompt.
- Higher output complexity. Google positions this release as producing more varied and realistic arrangements than earlier Lyria models.
Access is restricted to users aged 18 and over. The launch language list covers English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese, with more described as coming.
What the 30-second cap actually decides
Clip length is the most important constraint in the product, and it is worth being blunt about what it rules in and out.
Thirty seconds covers a social post, a channel ident, a loop under a product demo, a mood reference you hand to a composer, or a placeholder while a real score is commissioned. It does not cover a song. Anything that needs verse-chorus-verse structure, a bridge, or a deliberate emotional arc has to be assembled elsewhere, and the Gemini app is not a digital audio workstation.
That makes Lyria 3 a sketching tool rather than a production tool. Sketching tools are judged on iteration speed and on how faithfully they translate intent, not on final master quality, and that is the right lens to bring to it.
SynthID answers one question, not two
Every track generated in the app carries SynthID, an imperceptible watermark identifying Google AI-generated content. It is genuinely useful, and it is routinely misread.
SynthID answers "was this produced by Google AI?" It does not answer "am I allowed to use this?" Provenance and licensing are separate questions with separate owners, and a watermark grants no rights. Before a Lyria output goes into an ad, a client deliverable, or anything monetised, read the terms attached to your Gemini tier. Those terms, not the watermark, decide commercial use.
How we would evaluate it
If you are deciding whether Lyria 3 belongs in a workflow, these are the questions worth answering with your own prompts rather than from a launch post:
- Does control survive regeneration? Set a tempo and a vocal style, run the same prompt several times, and check whether those settings hold or quietly drift.
- How much variation does one prompt give? A sketching tool is only useful if repeated runs produce genuinely different options rather than near-duplicates.
- Does it take direction, or only vibes? Compare a vague prompt against a specific one. The gap between them tells you how much prompt effort is worth spending.
- What actually comes out of the app? Confirm the export format and whether you get a single mixed track or anything more granular, because that determines whether the output can be edited downstream at all.
- Do the lyrics need a pass? Auto-generated lyrics are a starting point. Budget review time if anything is going public under your name.
Where it fits
Lyria 3 belongs to the same push as the rest of Google's generative media work: fast, broadly available, aimed at the first draft rather than the last. If you are mapping that landscape more widely, our overview of AI video generation in 2026 covers the moving-image side, and our piece on Gemini audio and live translation covers what the same stack does with speech.
The honest summary: Lyria 3 makes a 30-second idea cheap to produce and easy to share. That is a real capability, and it is a narrower one than "make music with AI" tends to suggest.
You can try it at gemini.google.com.
