🎹 Lyria 3.5 in Flow Music: 3-Minute Vocals, Real Controls

The thing that sends me back to a DAW with any of these models is tempo. You get a chorus that works and a verse sitting four BPM off what you wanted, and the only fix is another generation and another roll of the dice. Lyria 3.5 lets you set it outright. That is a smaller headline than "better vocals" and a bigger change to how the tool fits a session.

Google rolled it out in Flow Music on July 29, replacing Lyria 3 Pro as the default model.

By the numbers

	Lyria 3.5
Released	July 29, 2026, in Google Flow Music
Replaces	Lyria 3 Pro as the default model
Max vocal track length	3 minutes
Direct controls	Tempo, vocals, drums, bass, track length
New input	Image-to-music
Called out for improvement	Musicality, lyrics, vocals, creative control
Notable language gain	Japanese vocal quality

Controls beat quality upgrades

Every model release claims better output. Few of them change your workflow. Stem-level control does.

Being able to pull the drums back or push the bass without regenerating means a track you almost like becomes usable instead of discarded. With Lyria 3, an 80%-right generation was a loss — you re-rolled and got a different 80%. Now it is a starting point.

Tempo and track length being explicit does the same thing for anyone scoring to picture. If your edit is 94 seconds, you can ask for 94 seconds rather than generating three minutes and finding a cut point that does not fall mid-phrase.

That is the gap nothing in the current generator lineup filled well. Suno and Udio both steer tempo through prompt wording and section regeneration rather than a setting, which works until it does not.

Vocals: better, and still detectable

Google's claim is more expression and emotion plus improved pronunciation.

What to listen for before you trust it on a release: sustained notes are where this class of model gives itself away. Vibrato on a held vowel tends to wobble at a regular rate where a real singer's drifts. Consonant attacks at the start of phrases were Lyria 3's worst tell, and pronunciation is the exact area Google says it improved — so that is the pair to A/B first.

The Japanese improvement is the most concrete gain in the release. Pronunciation on non-English vocals has been the weak spot across every generator in this category, and this is a real step.

Treat an exposed solo vocal as the hard test and a vocal sitting under instrumentation as the easy one. Nothing in this category has publicly passed the hard test, and the launch materials do not claim it has.

Image-to-music is narrower than it sounds

You upload an image and get music composed to match. Google frames it as composing to the visual content, and everything about how these systems encode images points to mood, palette and lighting carrying far more weight than subject matter.

Where it should earn its place is scoring a specific shot when you cannot put what you want into words. Where to expect it to fail is anything literal — do not count on a picture of a guitar producing a guitar. Test that before you build a workflow on it.

Lyrics still need a pass

Prompt adherence and structural awareness are both better — the lyrics stay on topic now, which was not reliable before, and verse-chorus structure lands where you would expect.

Line-level writing is still the weakest layer. You get competent, slightly generic couplets that rhyme where they should. For a background bed nobody is parsing, fine. For anything where the words carry the song, write them yourself and feed them in — the same discipline our Suno guide recommends for custom mode.

Where it sits against Suno

Suno remains ahead on catalogue breadth and on the sheer number of genres it handles convincingly. Lyria 3.5's edge is control granularity and Google-ecosystem integration through Flow.

If you need music generated programmatically rather than in an interface, neither is the answer — that is still ElevenLabs Music, whose API we have shipped production work through and documented in the ElevenLabs Music API guide.

The three-minute vocal ceiling puts Lyria in the same length bracket as a real single, which it was not in before. We broke that comparison out properly in Lyria 3.5 vs Suno.

How we tested

This is a release breakdown written from Google's own Lyria 3.5 announcement, cross-checked against Android Central and TechTimes coverage of the same rollout — not a listening review. Lyria 3.5 shipped on July 29 and we have not yet put a full session through it.

The listening guidance above is the checklist we run every new music model against, applied to what Google says changed. Once we have run real briefs through it, this post gets the artifacts and the numbers. Models also get patched quietly, so re-check any single behaviour before building on it.

What didn't make the cut

Benchmark scores. There is no credible blind-listening benchmark for AI music that survives contact with a real mix, and the ones that exist measure preference on isolated clips.

We also left out the licensing comparison against Suno and Udio. Those platforms are in active litigation with the major labels; Google's position is different and the terms are moving. That deserves its own post once it stops changing weekly.

Pros and cons

Worth it for: scoring to picture, anything needing a specific tempo or length, non-English vocals, and producers who want to fix a track instead of re-rolling it.

Skip it for: exposed lead vocals, lyric-driven songs, and any commercial release you have not checked current terms on first.