Stable Audio Open

Stability AI’s release of Stable Audio Open, a text-to-audio generative model trained on Creative Commons datasets like FreeSound and Free Music Archive, is prompting both praise and skepticism. Commenters welcome a model built on licensed “commons” data and see it as an ethical step forward compared to scraping proprietary content, but many criticize the non-commercial license as misleadingly labeled “open source” and unusable for serious commercial work. The thread broadens into debates over copyright, fair use, and whether training on artists’ work without consent is acceptable, alongside comparisons to competing music generators and speculation about future creative tools built on this kind of technology.

Licensing and “Open Source” Debate

  • Model released under Stability’s non-commercial research license; commercial use requires a paid membership.
  • Several commenters argue this is misleadingly branded as “open source” and is part of a broader pattern in AI of abusing the term.
  • Some push back that open source has become overly corporate-friendly and that non-commercial or strict copyleft licenses can be desirable.
  • One commenter suggests the license may not be enforceable on model weights, claiming weights aren’t currently copyrightable, but others don’t endorse this and note legal risk.
  • Frustration expressed that many AI enthusiasts ignore OSS licenses (e.g., GPL) when building projects.

Training Data, Creator Rights, and CC Licenses

  • Model is trained on FreeSound and Free Music Archive. This is praised as “commons in, commons out” and as an ethical alternative to scraping.
  • Others note that Creative Commons licenses have conditions and are not “free of copyright issues.”
  • Clarification: the model’s Hugging Face page says only CC0, CC BY, and CC Sampling+ audio were used, with attribution lists; this alleviates concerns about more restrictive FMA tracks.
  • Some doubt that many original FMA contributors would have anticipated or welcomed this ML use, despite the license terms.

Ethics, Copyright, and “AI = Theft?”

  • One line of argument: training on copyrighted data is not theft; the core legal issue is models outputting copyrighted text/audio verbatim, which can be mitigated with filters.
  • Others counter that copyright law did not anticipate machines that can memorize and regenerate large volumes of content; they see training on unlicensed work as unethical or potentially infringing.
  • NYT v. OpenAI is cited: critics say GPT stores large archives; defenders respond that evidence shows only short, overfit snippets are reproduced, not whole articles.
  • Comparisons made between models and compression algorithms; disagreement on whether “ability to reproduce” means a copy is legally “stored” in the model.
  • Multiple commenters suggest current copyright and fair use doctrines are inadequate for ML and will need rethinking.

Model Capabilities and Comparisons

  • Stable Audio Open is generally seen as targeting sound effects, loops, and textures rather than full songs or vocals.
  • Some report decent quality but complain about harsh high frequencies and lack of speech/singing.
  • Udio and ElevenLabs music demos are repeatedly cited as higher quality; others dismiss current AI music as bland, structurally shallow, and easy to spot if you listen closely.
  • Disagreement on detectability: some claim listeners couldn’t reliably distinguish AI vs. human tracks; others are confident they could.

Use Cases and Creative Ideas

  • Proposed “AI 8-track” app: hum a melody on multiple tracks, convert each line into instruments by text prompt, then lightly mix for rapid song sketching.
  • Commenters note that Google MusicLM, Suno, and Meta’s MusicGen already do variants of “hum-to-style,” but a polished workflow app is still missing.
  • One user experiments with “promptmusic” where terse textual prompts generate entire unusual tracks, observing that most song information resides in the model, not the prompt.
  • Several wish for strong audio-to-audio features: e.g., giving a drum pattern and having AI compose around it; current workarounds involve chaining separate tools and research models.

Broader Reflections and Miscellaneous

  • Some see this as an “Ethereum-merge-style” ethical inflection point: proof that high-quality models can be trained on consented/commons data.
  • Others dispute the crypto analogy and dive into a long subthread debating Proof of Work vs. Proof of Stake, environmental impact, and whether PoS inherently enriches the already wealthy.
  • A few commenters express cautious optimism and are glad Stability is still shipping models despite reports of company troubles.
  • Questions are raised about reusing such models for audio restoration/denoising; no clear answer or recommended open tools emerges.
  • A hosted demo and landing page are noted for trying the model online.