Google DeepMind has cranked up the ambition of its AI music engine. Rolling out today in Google Flow Music, Lyria 3.5 is pitched as a model that doesn’t just stitch notes together but tries to write music with genuine emotional weight — richer melodies, better lyrics, and vocals that sound less like a synthesizer pretending to sing and more like an actual performer.
The headline improvement is musicality. According to Google, Lyria 3.5 generates more complex melodic structures, which in practice should mean fewer of the flat, looping motifs that plague earlier generative audio and more of the tension-and-release that makes a track worth replaying.
Lyrics get a serious upgrade too. The model now shows stronger prompt adherence and, crucially, structural awareness — so when you ask for a verse-chorus-verse arrangement, it’s meant to actually respect that shape rather than dumping words wherever they fit. Pair that with the reworked vocal engine, which promises more realistic and emotionally nuanced singing plus improved pronunciation, and the gap between AI output and a demo recorded in someone’s bedroom narrows considerably.
For anyone who has wrestled with AI music tools that treat you like a passenger, the new creative controls are the practical win. Lyria 3.5 lets you dial in tempo and duration, giving you a firmer grip on how a piece is paced and how long it runs — the kind of hands-on parameters producers actually care about.
Here’s what’s new at a glance:
- Improved musicality — richer, more complex melodic structures
- Enhanced lyrics — better prompt adherence and structural awareness
- Improved vocals — more realistic, emotionally nuanced singing with cleaner pronunciation
- Creative control — adjustable tempo and duration
Access comes through Google Flow Music on two tiers. The AI Pro plan runs $19.99/month and includes 1,000 Flow credits, aimed at hobbyists and creators dabbling in generated tracks. The AI Ultra plan sits at $200/month with a hefty 25,000 credits, clearly targeting power users and anyone treating AI music as part of a real production pipeline.
The direction here is telling. Instead of chasing raw novelty, Google is refining the parts that separate a passable jingle from something with feeling — phrasing, vocal expression, and structure. Whether Lyria 3.5 truly delivers on the “emotional” promise is something ears will judge, but the feature list suggests DeepMind is done treating AI music as a party trick and is starting to build a proper instrument.
