We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind pushed out Lyria 3.5 today, folding it directly into Flow Music rather than keeping it as a standalone lab demo. That's the telling part. This isn't a research paper flexing benchmark numbers — it's a tool meant for people who want a finished song, not a proof of concept.

The headline changes cluster around four areas: musicality, lyrics, vocals, and control. On musicality, Lyria 3.5 is supposedly better at building melodic structures that don't collapse into repetitive loops the moment a track passes the two-minute mark, which has been a chronic weakness of generative music models since the genre started. Lyrics get a boost too, with tighter prompt adherence — meaning if you ask for a break-up song about missing a bus, you're more likely to get exactly that instead of a vague approximation.

Vocals are where Google seems to be leaning hardest. The company says the model now produces more emotionally nuanced singing, with pronunciation that holds up better across phrasing and genre. Anyone who's played with earlier AI vocal generators knows the usual failure mode: technically in tune, emotionally dead, and prone to mangling consonants at the worst moments. If Lyria 3.5 genuinely narrows that gap, it matters more for usability than any melodic upgrade, since a flat vocal can sink an otherwise decent track instantly.

The last piece, creative control, sounds modest but is arguably the most practical addition. Users can now set tempo and duration directly instead of hoping the model lands somewhere close to what they wanted. That's the kind of unglamorous feature that turns a novelty generator into something people actually use for real projects — a jingle that needs to hit exactly 30 seconds, a loop that has to match a video cut precisely.

Google didn't publish benchmark comparisons against Lyria 2 or rival tools like Suno and Udio, so for now this is a claims-based rollout rather than a measured one. The real test starts the moment users inside Flow Music start typing prompts and judging the output against their own ears, which tend to be far less forgiving than a press release.