So, Google dropped Lyria 3 recently. It’s their new audio AI from the DeepMind team, and they are plugging it right into Gemini. Honestly, the AI music space has been pretty rough over the last year. Lots of hype, bad results. We decided to mess around with it here at Ainformer to see if it actually works or if it’s just another tech demo.
What It Actually Does (No Boring Loops)
Past tools just made elevator music. You know the type. Stiff and repetitive. This is a totally different animal.
1. It Sings. Like, Real Words.
This is the wild part. You type a random thought, and it doesn’t just give you a beat. It writes the lyrics and sings them back to you. The weirdest thing? It sounds okay. The breathing noises and pitch are surprisingly on point.
2. Feeding it Pictures
Got a photo? Drop it in the chat. The AI looks at the picture, guesses the vibe, and makes a background track. It is a super weird feature, but honestly, it saves so much time when you hate typing out long descriptions.
3. The 30-Second Limit
Right now, the tracks cut off at 30 seconds. But they pack a lot into that half-minute. Oh, and you get custom cover art generated for every single track (they use some built-in tech called Nano Banana for that). Nice touch.
Can’t Find It? You Aren’t Crazy
A lot of people are logging in right now (February 2026) and seeing absolutely nothing new. The rollout is super slow. Here is the deal:
- It heavily favors English right now.
- Google is throttling access so servers don’t melt.
- Your account needs to be 18+. No exceptions.
Quick Ainformer tip: Go to your Google account settings, swap your language to US English, and check the desktop site. That forces the update for a lot of people.
How to Not Get Bad Music
If you type “make a song,” you will get garbage. Be bossy. Be specific. Try something like this:
“Give me a fast, loud jazz track with a crazy sax solo for a morning vlog.”
Tell it the exact instruments and the mood. It listens.
The Safety Stuff
Don’t try to make it sing like Taylor Swift or Eminem. DeepMind hard-blocked all of that. It just throws an error. Also, they stamp a hidden audio watermark called SynthID on every file, so people can always tell it’s fake.
Final Verdict
It’s a fun toy that’s quickly turning into a legitimate tool for creators. We will keep watching it. Stick around, we’ll be posting more weird AI experiments soon.
Redactor
