Google DeepMind's Lyria 3.5 lands in Flow Music, promising better vocals and lyrics
The new model targets richer melodies, cleaner pronunciation, and more control over tempo and track length.
Key points
- Google DeepMind released Lyria 3.5, its newest AI music generation model, inside Flow Music on launch day.
- The update targets four areas: musicality, lyrics, vocals, and creative controls including tempo and duration.
- Vocals are designed to sound more expressive and pronounce words more accurately than the previous version.
- Flow Music is Google's consumer tool for making full songs from written prompts.
- No pricing changes or new safety details were announced alongside the model.
Google DeepMind has rolled out Lyria 3.5, the latest version of its AI music generation model, the software that turns a written prompt into a full song. It's live now inside Flow Music, Google's tool for making tracks from a description.
Four things changed, according to Google. Melodies are meant to feel more natural and less repetitive. Lyrics should follow your prompt more closely and carry a clearer verse-and-chorus shape. Synthetic vocals carry more emotion and pronounce words more accurately. And you can now set tempo and track length with less friction.
What actually changed?
| Area | What Lyria 3.5 promises |
|---|---|
| Musicality | Richer, more complex melodies |
| Lyrics | Better prompt adherence and song structure |
| Vocals | More expression, cleaner pronunciation |
| Controls | Tempo and duration adjustments |
The vocals are the headline claim. Pronunciation has been the easiest tell with AI singing, the point where a word gets swallowed or a consonant turns to mush. If Google's improvement holds under real use, that tell starts to disappear.
Where can people try it?
Inside Flow Music, Google's consumer music creation product. Google says the update is available now, with no waitlist mentioned.
Flow Music is aimed at hobbyists and creators who want a finished track rather than a raw audio clip.
What this means for listeners and creators
For listeners, more AI-generated songs are about to appear on social feeds, and they'll be harder to spot. Cleaner vocals and tighter lyric structure were exactly the things that gave AI music away.
For working musicians and producers, the practical questions are the ones Google didn't answer: how the model was trained, what data it learned from, and what rights users hold over the songs they generate. Music generation tools have faced sustained legal scrutiny from record labels and rights holders, and Lyria 3.5's launch post says nothing on that front.
For everyone else, the gap between "obviously a robot" and "could be a real singer" is closing fast. That's not a metaphor; it's a production reality arriving in consumer software this week.
Should you worry about how AI music gets labelled?
Yes, and here's the specific gap worth watching. Google hasn't detailed how Flow Music outputs will be watermarked or tagged when shared to other platforms. If a track sounds fully produced, the label in the corner of the app may be the only signal a listener gets.
Creators uploading AI-assisted songs to streaming services should check each platform's disclosure rules. Spotify and YouTube have both updated their AI-content policies over the past year, and enforcement is tightening.
Listen for pronunciation as Lyria 3.5 spreads. Older AI vocal models slurred consonants or mangled unusual words. That's been the fastest way to identify a synthetic track. If Google's claim holds, that shortcut won't be reliable for much longer.
Credit to the DeepMind team for shipping this update with specific claims attached rather than vague promises. The proof will be in what people actually make with it.
