FlowtandBriefAI · Software · ResearchMeet Flowtand

Google Releases Lyria 3 Pro With Longer Tracks and Structure Control

Announcement

Google DeepMind

Browse Models topic page →

Google announced Lyria 3 Pro, a music generation model that produces tracks up to three minutes long and responds to prompts for structural elements such as intros, verses, choruses and bridges. Google says it is available in Vertex AI public preview, AI Studio, the Gemini API, Vids, the Gemini app and ProducerAI, with watermarked outputs and filters intended to prevent imitation of named artists.

What Google announced

Google announced Lyria 3 Pro, described as its most advanced music generation model, which creates tracks up to three minutes long. According to the announcement, the model has improved awareness of song structure, so a user can prompt for specific parts such as intros, verses, choruses and bridges. Google says it is distributing the model across several surfaces: Vertex AI, where it is in public preview for businesses needing audio at scale; Google AI Studio and the Gemini API for developers; Google Vids for video creation; the Gemini app for paid subscribers; and ProducerAI, which Google describes as available globally to free and paid subscribers. The post also states that outputs carry the SynthID watermark, that filters check outputs against existing content, and that the model does not mimic named artists, treating such names as broad inspiration. These are Google's claims about its own product; the excerpt provides no independent test results.

Scenario for a corporate video producer

A video producer at a mid-sized company could try Lyria 3 Pro in Google Vids when a training or marketing video needs a short instrumental bed or a song with a defined verse-and-chorus shape. The structural prompting described by Google would matter most for clips long enough that a single loop sounds repetitive. A producer could draft one or two candidate tracks in Vids, place them under narration, and check whether the commissioned structure holds across the edit. This is a conditional experiment, not an established workflow; Google's post notes Vids access is rolling out to Workspace customers and AI Pro and Ultra subscribers, so availability should be confirmed before planning around it.

A trial plan with checks

A team could design a small comparison: take several existing video projects, generate candidate tracks with Lyria 3 Pro using explicit section prompts, and compare against the music those projects currently use on clarity under voiceover, fit to runtime and how much manual trimming is needed. Because the source gives no quality metrics, the team should define its own pass criteria in advance, such as whether a track reaches the target length without awkward transitions and whether editors accept it without heavy rework. Failure checks should include listening for unintended resemblance to known recordings, verifying SynthID where detection matters, and confirming that a human editor or music supervisor reviews every track before publication. Access tier and rollout timing should be verified first.

Trade-offs and when to hold back

Longer generations and section-level control reduce some editing work, but the source does not describe audio fidelity, latency, pricing, licensing terms or commercial usage rights, so teams should not assume any of these. Structural prompting may still produce sections that need cutting or rearranging. The watermarking and anti-mimicry measures are protective but do not remove the need for rights review, and Google's statement that training used materials it has rights to use does not settle questions about output ownership. Organizations with strict brand-sound guidelines, or projects where an original human composition is central, should treat this as an optional drafting aid rather than a replacement, and should wait for legal review of the applicable terms before commercial release.

Context and limitations

  • Based on a short source excerpt; the full article was not analyzed.
  • Availability timing and tiers differ by product; the excerpt notes rolling rollouts and paid-subscriber access for some surfaces.
  • All capability claims come from Google's announcement; no independent evaluation of output quality is included.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.