Google Flow Music Lyria 3.5

The first thing I noticed about Flow Music's Lyria 3.5 is that the vocals finally stopped sounding like a cheap vocoder. Previous versions had that telltale warbly shimmer, the kind you immediately recognize as AI singing. With 3.5, the phrasing actually breathes. Consonants land crisper, vibrato doesn't feel random, and—this matters more than any spec sheet—the voice can hold a note without sliding off pitch halfway through. I spent a weekend doing side-by-side generations with the same lyric prompts, and the difference is not subtle.

What stands out to me is that the musicality improvement isn't just about "more pleasing chords." The model seems to have a stronger sense of structure. It understands verse-to-chorus tension now. You can give it a loose prompt like "melancholic synth-pop bridge that resolves into a bright major chorus" and it actually delivers a functional transition, not just a key change. That's huge for anyone using this as a songwriting sketchpad rather than a final-take machine.

Lyrics are better in a subtler way. 3.5 doesn't write more grammatically correct sentences—it writes more singable lines. The syllable stress matches the melody more naturally. Before, I'd get awkward prosody, like forcing the word "photograph" onto three eighth notes with the accent on "graph" when it shouldn't be there. Now the accent patterns feel human. Not perfect, but human enough to use as a starting point without rewriting everything.

Vocals are the headline feature, and for once that's not hype. You can specify vocal texture better: husky, airy, nasal, belt. I tested "breathy female voice, close-mic," and got something that genuinely sounded recorded in a small room, not synthesized in a void. The background vocal generation is also improved—backing harmonies are coherent, not just the same line doubled and pitched up.

That said, there are still tells. Long sustained notes can develop a metallic edge after about four seconds, and anything faster than 150 BPM makes the lyrics slur. But for a v3.5 release, this feels like a real leap rather than an incremental polish. If you're doing practical work with AI music tools, this is worth a hands-on test. The prompt engineering workflow is basically unchanged: describe the vibe, enter lyrics, pick the vocal character. Google just made the model listen better.