Stability AI drops open-weight audio models that generate 6-minute tracks on licensed data, Google takes Flow Music global with section-level editing at I/O, and LALAL.AI brings offline six-stem separation into your DAW.

Stable Audio 3.0 delivers open-weight music generation trained entirely on licensed data

Stability AI released Stable Audio 3.0 on May 20 — a family of four models ranging from 459M to 2.7B parameters. The medium (1.4B) and large (2.7B) models generate full compositions up to 6 minutes 20 seconds long while maintaining musical structure and melodic coherence, more than doubling the ceiling of Stable Audio 2.0. The small SFX, small, and medium models ship with open weights under the Stability AI Community License. The large model is API-only.

The licensed-data angle is the real story. Every model was trained on material from deals with Warner Music Group and Universal Music Group, sidestepping the copyright minefield that has defined the last two years of AI audio. For developers: the medium model generates audio in under two seconds on an H200 and a few seconds on a MacBook Pro M4. Companies under $1M annual revenue can use and commercialize outputs freely; above that threshold, you need an enterprise license.

Source: TechCrunch, Music Business Worldwide, Digital Music News

Google Flow Music goes global with Lyria 3 and granular song editing

At I/O 2026 on May 19, Google expanded Flow Music worldwide and integrated the Gemini Omni Flash model. The headline feature for music makers: section-level editing. You can now modify specific portions of an AI-generated song — changing lyrics to a different language, swapping the genre of a verse, or adjusting instrumentation — without regenerating the entire track. This is a meaningful step beyond the "generate and pray" loop most AI music tools still rely on.

Lyria 3, the underlying music model, generates tracks with instrumentals, vocals, and lyrics from text or image prompts. A new music video feature lets you direct visuals conversationally through Gemini Omni. The Flow Music app is live on iOS, with Android in beta. If you work in content production, the combination of granular editing plus AI-directed video generation makes this worth testing against your current workflow.

Source: Google Blog, Android Central, 9to5Google

LALAL.AI brings offline six-stem separation into your DAW

LALAL.AI updated its VST3 plugin on May 22 with offline six-stem processing — vocals, instrumentals, bass, drums, acoustic guitar, electric guitar, and piano — running entirely on your machine. No cloud upload, no credits consumed from your account balance. The plugin works in Ableton Live, Reaper, and any VST3-compatible DAW, with AU support in beta. Available on Windows, macOS, and Linux.

The local processing angle matters. Studios handling client stems or unreleased material have real reasons to avoid cloud-based separation tools. LALAL.AI's approach — activate once, then everything stays on-device — removes that friction. Quality is competitive with the top tier of separation engines, though serious comparison testing against Demucs and SpectraLayers will take a few weeks to shake out.

Source: Bedroom Producers Blog, MusicRadar, MusicTech

Suno fights to keep Warner settlement terms sealed as Sony cases approach pivotal ruling

Suno is actively blocking Universal Music Group and Sony Music from accessing the terms of its November 2025 settlement with Warner Music Group, as discovery continues in the remaining copyright lawsuits. The sealed terms likely include royalty structures and data-use provisions that would inform Sony's ongoing litigation against both Suno and Udio — which is exactly why Suno wants them buried.

The bigger picture: Sony is now the last major label in court. Its fair-use cases against Suno (Massachusetts) and Udio (Southern District of New York) are heading toward a summer 2026 ruling that could define whether training AI models on copyrighted music constitutes transformative use. A summary judgment hearing is expected in July. The outcome will determine whether AI music companies must license or can train freely — and every tool in this space is affected.

Source: Music Business Worldwide, Billboard, Digital Music News

ElevenLabs expands music generation across its API

ElevenLabs pushed several updates to its SDK in May, adding song_id tracking in music generation responses and new generation parameters that give developers finer control over output. These are incremental API improvements, not a new model — but they signal that ElevenLabs is serious about building out the music side of its "one API for all audio" strategy, launched in April with the ElevenMusic iOS app.

For developers already using ElevenLabs for voice synthesis or sound effects, the unified API approach is the draw: one subscription, one set of endpoints for speech, music, and SFX. The music generation itself is still catching up to dedicated tools like Suno and Udio in output quality, but the integration convenience is real if you're building products that need all three audio types from a single provider.

Source: ElevenLabs Release Notes
Tool of the week
Stable Audio 3.0 Medium

The medium model (1.4B parameters) is the sweet spot of the Stable Audio 3.0 family: open weights, runs on consumer hardware in seconds, generates up to 6:20 of musically coherent audio, and trained entirely on licensed data so your outputs are commercially clear. You own what you generate and can distribute freely under the community license (enterprise license required above $1M revenue).

Download the weights from Hugging Face, run it locally, fine-tune it on your own material. For sound designers and composers who've been waiting for a self-hostable generation model that doesn't come with a copyright asterisk, this is the most practical option available right now. The small models (459M params) are lightweight enough to run on smartphones for quick prototyping on the go.

Open weights (Community License) · API & self-host · stability.ai/stable-audio
Worth watching
The Sony Music fair-use ruling that could reshape AI audio

With Warner and UMG settled, Sony Music's cases against Suno and Udio are the last major-label lawsuits standing — and likely the ones that will produce binding precedent. The core question: is training a generative music model on copyrighted recordings transformative fair use? A summary judgment hearing is expected in July 2026.

The outcome affects everyone building in this space, not just the defendants. A ruling against fair use would force a licensing-or-nothing model across the industry, which favors well-funded incumbents who've already cut deals. A ruling in favor would open the floodgates for new entrants training on publicly available music. Either way, if you're building tools or products that depend on AI-generated music, this summer's decision sets the ground rules. Plan accordingly.