Google's Lyria 3.5 music model enables individual song section editing and vocal/instrument tuning

Google quietly leveled up its music generation game with Lyria 3.5, and the headline feature here isn't the audio quality — it's the control. The new model, available now through Google Flow Music, introduces something called Selective Section Painting, which lets you grab a chorus, a bridge, or a stray 15-second hook and rework it in isolation. No more regenerating an entire three-minute track because the outro felt off.

That kind of surgical editing is a meaningful shift for AI music tools, most of which still treat a song as one indivisible blob you either accept or discard. Lyria 3.5 also lets users dial in tempo and duration separately for vocals, drums, bass, and other layers, and it can stretch a fragment of melody into a full arrangement instead of forcing you to start from a blank prompt. Tracks can now run anywhere from 30 seconds to a full three minutes, and Google says the model handles pronunciation and vocal realism noticeably better than its predecessor.

Google is framing all of this as a creative-tool upgrade rather than a party trick, which tracks with where the company positioned Lyria 3 a few months back. But the training data question remains stubbornly unanswered. When Lyria 3 launched, Google said only that it used material YouTube and Google were licensed to use under terms of service, partner deals, and applicable law — vague language that told rights holders almost nothing. Asked specifically about what fed Lyria 3.5, Google didn't respond.

That silence matters more now than it did with earlier models, precisely because the tool is more capable. A system good enough to isolate and rebuild individual song sections is also a system that's clearly learned fine-grained structure from something — and artists, labels, and publishers are going to keep asking what that something was, especially as Google leans harder into music as a Flow product rather than a lab demo.