FlowtandBriefAI · Software · ResearchMeet Flowtand

Google adds Lyria 3 music generation, with SynthID watermarks, to the Gemini app

Announcement

Google DeepMind

Browse Models topic page →

Google announced that the Gemini app now offers beta music generation powered by DeepMind's Lyria 3. Users 18+ in supported languages can describe an idea or upload a photo or video to produce a 30-second track, with lyrics and custom cover art generated automatically. Google says every track made in the app carries a SynthID watermark, and the app's verification tools now cover audio as well as image and video. Subscribers get higher usage limits, and Lyria 3 is also rolling out on YouTube's Dream Track.

What Google announced

Google said the Gemini app now includes beta music generation using Lyria 3, which it describes as DeepMind's latest generative music model. Users aged 18 and over, in English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese, can type a description or upload a photo or video, and the app composes a 30-second track with generated lyrics and cover art made with Nano Banana. Google says Lyria 3 writes its own lyrics, offers more control over style, vocals and tempo, and yields more realistic, musically complex tracks than earlier Lyria models. Every track made in the app carries a SynthID watermark, and the app's verification feature now checks audio alongside image and video. Google frames the feature as a way to express yourself rather than to produce finished music, and says training was done with attention to copyright and partner agreements.

A scenario for a social media manager

Consider a social media manager who needs a short, on-brand audio sting for a recurring series of Stories or short videos. Under Google's description, they could prompt Gemini with a genre, mood and inside reference, or attach a reference image, and receive a 30-second clip with lyrics and cover art they can download or share by link. The watermark would let colleagues check later whether a file was made with Google AI. Because Google presents the output as a quick expression tool, not a finished soundtrack, a team could try it for fast, low-stakes drafts and keep final production decisions with a human editor.

A bounded trial

A content team could run a two-week pilot: write five briefs the way Google describes (a genre, mood and memory, plus one photo-driven request), generate tracks for each, and log how many needed rewriting or were unusable. Compare against a baseline of stock or licensed library tracks chosen for the same briefs, and measure time to a usable clip and editor acceptance, not just generation speed. As a failure check, test whether prompts naming a specific artist still produce usable, non-imitative results and whether team members can verify a known AI-made file using the app's SynthID check. A human should review anything published.

Trade-offs and caution

The 30-second cap, beta status and higher limits for AI Plus, Pro and Ultra subscribers shape where this fits: short-form and internal use, not full productions. Google says its filters and artist handling 'might not be foolproof', so teams with brand or rights exposure should treat verification and reporting as part of the workflow. The announcement does not cover licensing terms, commercial-use details or audio quality beyond Google's own claims. If a team needs longer pieces, tight stylistic control or independently validated provenance, this feature is not the right starting point.

Context and limitations

  • Based on a short source excerpt from Google's announcement; the full article was not analyzed.
  • Capabilities, language coverage and limits are Google's stated claims; no independent testing or output quality was assessed.
  • Google itself says its filtering approach might not be foolproof, and notes generative AI is experimental.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.