Stable Audio 2.0 can generate full 3-minute tracks

Most AI music tools have historically felt like digital scratchpads: fun for generating a ten-second loop, but entirely incapable of producing a real song with an actual arc.

Stability AI launched Stable Audio 2.0, setting a new benchmark for generative sound by producing full tracks up to three minutes long in 44.1kHz stereo from a single prompt.

Swapping U-Net for DiT

The biggest leap here is musical coherence over time. Stable Audio 2.0 can generate structured compositions that actually follow standard musical formats, complete with intros, melodic developments, and outros.

Under the hood, Stability overhauled the entire latent diffusion pipeline. It replaced the previous U-Net backbone with a Diffusion Transformer, similar to the architecture used in Stable Diffusion 3, combined with a highly compressed autoencoder to master long sequence modeling.

Beyond text-to-audio, the 2.0 release adds an audio-to-audio workflow. Users can now upload raw sound bites or voice recordings and use natural language to transform them into polished stems, custom sound effects, or entire style-transferred backing tracks.

To keep things legally clean, the model was trained exclusively on a licensed catalog from AudioSparx comprising over 800,000 tracks and sound effects, honoring creator opt-outs. For user uploads, Stability partnered with Audible Magic to run real-time content recognition and block copyrighted material. The tool is free on the web today, with API access and a 24/7 YouTube radio stream rolling out alongside it.

Why it matters

We think this release moves AI audio from novelty territory straight into functional production pipelines.

For indie game devs, video editors, and podcasters, having a tool that generates a full, three-minute, coherent background track at commercial fidelity without licensing landmines is huge. The audio-to-audio feature also means musicians can hum a melody on their phone and immediately test how it sounds as a synthesized brass section or orchestral swell.

Stability's cautious approach to copyright stands out as well. At a moment when major record labels are gearing up for legal battles over AI training data, leaning into licensed datasets and real-time audio filtering provides an enterprise-friendly shelter that rival experimental tools lack.

The lingering question is whether these three-minute tracks will sound genuinely soulful or simply feel like highly competent elevator music. But as a baseline utility for media creation, the bar for generative audio just jumped significantly higher.