
Highest raw audio fidelity in the category
SynthID watermarking is applied at generation, which is increasingly what distributors and platforms want to see on AI-assisted audio.
Google DeepMind
DeepMind's song model, distributed through Google rather than sold directly.
In short
Figures verified 2026-08-10. This field moves quickly — re-check before relying on them.
Lyria's significance is less about beating Suno on any single axis and more about who is shipping it. A model distributed inside Google's stack arrives with the licensing conversations already had, the watermarking already applied, and a distribution surface no startup can match. In practice that makes Lyria 3 the option that enterprise and platform teams can actually get approved, which is a different competition from the one consumer tools are fighting. Musically it is strong on instrumental coherence and restrained arrangement, and Google has been notably more conservative than its competitors about vocal generation and artist imitation. The trade-off is access: you are working within Google's product surfaces and API terms rather than an open creative tool, and the model is not available to run yourself. For most individual creators that makes it something they encounter inside another product rather than something they choose.
Strengths

SynthID watermarking is applied at generation, which is increasingly what distributors and platforms want to see on AI-assisted audio.

Instrumental arrangement holds together over longer durations rather than drifting, which suits underscore and background work.

Access through Google Cloud means procurement, data handling and compliance questions have answers a legal team will accept.
How it compares
Lyria 3 is Google DeepMind's text-to-music model, generating full instrumental and vocal music from a description. It reaches users through Google's own products and cloud APIs rather than as a standalone subscription, and output carries SynthID watermarking.
Compare with MusicGenerate
Searched as lyria, lyria 3, google lyria, deepmind music ai and google ai music generator. Most people meet it inside another Google product rather than choosing it, which is why the access question comes up more often than the quality one.
Through Google's own products and cloud APIs rather than as a standalone subscription. For most individual creators that means encountering it inside another product rather than choosing it directly.
Yes, output carries SynthID watermarking applied at generation. That is increasingly what distributors and platforms want to see on AI-assisted audio, so it is an advantage for commercial work rather than a restriction.
It generates both instrumental and vocal music, though Google has been notably more conservative than its competitors about vocal generation and anything resembling artist imitation.
Access is mediated by whichever Google surface you reach it through. Cloud API access is aimed at teams; consumer surfaces expose a simpler prompt box with far less control.
Whichever Google Cloud or product agreement you are using, rather than a single music licence. That suits organisations with procurement requirements and is more friction than a hobbyist wants.
On raw vocal expressiveness, no. On distribution, compliance and watermarking, it is in a different category — which is why it tends to win enterprise decisions and lose consumer ones.
Keep exploring
More models
Models and platforms we track, compared on capability and licence.
DeepMind's song model, distributed through Google rather than sold directly.