On a Tuesday morning that signals Google's latest push into an increasingly crowded creative frontier, Google DeepMind switched on Lyria 3 — its most advanced music generation model to date — inside the Gemini app, placing AI composition tools in the hands of anyone with a Google account and a passing idea.
The rollout, beginning on desktop today and expanding to mobile within days, is a significant step in the AI music race — though not yet a decisive one. Lyria 3 brings generative music from enterprise infrastructure into everyday consumer life, but it enters a market where rivals have already set a formidable bar.
What the Machine Can Do
Lyria 3 generates 30-second tracks from text prompts or uploaded images and videos, automatically producing lyrics, harmonics, and custom cover art through an integrated tool called Nano Banana. Where its predecessors required users to supply their own words and accept relatively rudimentary arrangements, Lyria 3 writes the lyrics itself — and does so across eight languages: English, German, Spanish, French, Hindi, Japanese, Korean, and Portuguese.
The leap in technical capability is measurable. Earlier Lyria versions offered basic style selection. Lyria 3 lets users define vocal character, tempo, genre, and acoustic texture with granular precision. Its output, according to internal documentation, produces "more realistic and musically complex arrangements" than any prior iteration — and the model's range spans pop, funk, Motown, global folk forms, and ambient soundscape work.
The Copyright Tightrope
No story about AI music generation can avoid the central anxiety of the industry it now threatens to reshape: intellectual property.
Google has moved carefully. All audio produced by Lyria 3 is imperceptibly watermarked using SynthID, DeepMind's proprietary detection technology, allowing any generated or AI-edited track to be identified. Users can upload audio files to Gemini to verify their AI origins. On YouTube, where Lyria 3 integrates with Dream Track for Shorts, creators may invoke artist names only as "broad inspiration" — style filters prevent direct vocal imitation, and reporting mechanisms exist for IP disputes.
The guardrails reflect years of tension between AI developers and the music industry, and they signal that Google is aware this technology does not arrive in a vacuum. Whether those guardrails will prove sufficient is a question that courts, not engineers, may ultimately answer.
Good Progress, but not the Best
Industry watchers at CTOL Digital Solutions, a digital consultancy that conducted house evaluations of the model, offered an assessment that was both appreciative and unsentimental.
On the positive side, they found Lyria 3's audio quality competitive with leading rival Suno V5— noting fewer artifacts, cleaner output, and notably good handling of distorted instruments. The UI was praised for its ease and playfulness; testers turned a shopping list into a punk rock track. SynthID watermarking was recognized as a meaningful trust and safety feature.
The criticisms, however, were pointed. Compositional creativity, evaluators wrote, "feels quite boring and lacks creativity and originality" — Lyria 3 has made genuine progress, but has not yet mastered the artistry that distinguishes it from average musicianship, let alone the benchmark set by Suno V5. Access issues were also reported in the United States, Japan, and Greece during early testing.
CTOL's conclusion was blunt: "It’s a notable achievement for Google, but it doesn’t move the needle for the music industry overall—and Suno V5 still leads across almost all the dimensions."
The race for the soul of music's future, it seems, is still very much underway.
This article is based on materials published by Google DeepMind and evaluations conducted by CTOL Digital Solutions.