AI Music Models

The vocal benchmark every other song model gets measured against.

The song model that lets you keep editing after the first take.

DeepMind's song model, distributed through Google rather than sold directly.

The song model that lets you specify the structure before it writes.

Full songs at a fraction of the premium tier's cost per generation.

Style conditioning from a reference clip rather than a text description.

The draft model — fast and cheap enough to think out loud with.

API-first song generation with an unusually fast release cadence.

The spectrogram experiment that turned into a song platform.

The instrumental generation Google shipped before Lyria 3.

The long-form option for people who need to run the model themselves.

Editing inside existing audio, not just generating from nothing.

The open song model with a licence you can actually build a business on.

The open model that actually sings the words you gave it.

Full-length songs generated in a fraction of the usual time.

The open model with a training set it can actually name.

Hum an idea and get it arranged — the original melody-conditioned model.

Meta's fast model — the same quality without the sequential wait.

The separation model everything else gets benchmarked against.

The separation tool that made stem splitting ordinary.

A vocal-transformation platform rather than a song generator.

Template-driven tracks for creators who want predictable, not surprising.

Generation plus a library, for when generation alone misses.

Edit the arrangement section by section instead of regenerating.

The shortest path from nothing to a track on a streaming service.

Continuous, never-repeating music built to be embedded.

The one that hands you MIDI instead of a finished file.

Stem separation for people who do not want to install anything.