Google Lyria 3 Pro

Lyria 3 generates a complete, high-fidelity track — vocals and instrumentation together — from a text prompt or lyric sheet, spanning genres from pop to funk to Motown-style arrangements per Google's own model description.

On AutorunX it's exposed as a prompt/lyrics-to-song engine inside Song Builder: describe the style or paste your lyrics, and Lyria 3 returns a produced track built to that description.

Lyria 3 uses latent diffusion applied to temporal audio latents, trained on audio data annotated with text captions at multiple levels of detail. Instead of predicting discrete speech-style tokens one at a time, the model iteratively refines a full audio latent toward the target described by the prompt — a fundamentally different generation process than autoregressive, token-based music models.

On AutorunX, Lyria 3 is reached indirectly through OpenRouter rather than a direct provider integration, which adds a routing layer relative to engines AutorunX connects to natively — one reason it's presented as a selectable alternative in Song Builder rather than the primary path.

It takes the same input as AutorunX's other song engines — a style prompt or a pasted lyric sheet — and returns a finished track with composition and vocals handled together, so switching to Lyria 3 for a given song idea is simply a different choice in the engine picker.