
Best-in-class fidelity
Clean high frequencies and a controlled low end mean output usually needs less corrective processing before it sits under a voiceover or a video edit.
Google DeepMind
The instrumental generation Google shipped before Lyria 3.
In short
Figures verified 2026-08-10. This field moves quickly — re-check before relying on them.
Lyria 2 is the instrumental-only generation that preceded Lyria 3, and it remains relevant for exactly that reason: a model that does not attempt vocals avoids the failure mode where a decent instrumental is ruined by a synthetic voice on top. For underscore, background beds and sound design, that constraint is a feature. Its distinguishing capability is negative prompting. Most music models ignore instructions about what to avoid, or actively summon the thing you excluded, which is why prompt guides tell you to phrase everything positively. Lyria 2 accepts exclusions directly, so removing drums or brass from a texture takes one clause rather than three regeneration attempts. Fidelity is high — clean top end, controlled low end — and short generations are quick and inexpensive, which makes it practical for producing many variations of a bed and choosing between them.
Strengths

Clean high frequencies and a controlled low end mean output usually needs less corrective processing before it sits under a voiceover or a video edit.

Excluding an instrument takes one clause instead of several regeneration attempts, which is genuinely rare among music models.

Short generations are quick and cheap enough that producing a dozen variations of a bed and picking one is a realistic workflow.
How it compares
Lyria 2 is Google DeepMind's instrumental music model, generating high-fidelity music without vocals from a text description, and notable for supporting negative prompts — telling it what to leave out, which most music models cannot do.
Compare with MusicGenerate
Searched as lyria 2, google lyria 2 and google instrumental ai. It is the instrumental generation before Lyria 3, and negative prompting is the capability that keeps it relevant.
No. It is an instrumental model, which is a feature rather than a gap — a decent instrumental ruined by a synthetic voice on top is a common failure of models that attempt both.
Negative prompting lets you say what to leave out. Most music models ignore exclusions or actively summon the thing you excluded, so removing drums or brass usually takes several regeneration attempts. Lyria 2 accepts the exclusion directly.
Yes. Clean high frequencies and a controlled low end mean output usually needs less corrective processing before it sits under a voiceover or a video edit.
Pricing follows the Google Cloud or product terms you access it under rather than a consumer subscription. Short generations are quick and inexpensive enough that producing many variations of a bed is practical.
Use Lyria 2 when the brief is instrumental and you want negative prompting. Lyria 3 is the broader model and adds vocal generation, at the cost of the tighter instrumental focus.
For instrumental work with negative prompting, yes. Lyria 3 is broader and adds vocals, at the cost of the tighter instrumental focus that makes Lyria 2 predictable for underscore.
Keep exploring
More models
Models and platforms we track, compared on capability and licence.
The instrumental generation Google shipped before Lyria 3.