Last updated 08-09-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Harmonai.org
Harmonai is a Stability AI lab that ships open-source generative audio models and notebooks for musicians. Its first public release, Dance Diffusion, lives in the sample-generator repo and covers unconditional sample generation, single-file style transfer, and interpolation between two clips. The org also maintains oobleck audio codecs and an audio-diffusion PyTorch fork on GitHub.
Unlike closed DAW plugins or hosted beat makers, Harmonai publishes training code and pretrained checkpoints you can run locally or in Colab. The Dance Diffusion notebook lists six public models, including maestro-150k at 16 kHz and glitch-440k at 48 kHz, plus datasets ranging from piano MAESTRO clips to field recordings from xeno-canto.
Producers, researchers, and hobbyists who want to experiment with diffusion audio rather than rent a black-box generator are the core audience. The site routes newcomers to a Discord community, and Stability AI's stable-audio-tools repo still documents the Dance Diffusion U-Net architecture Harmonai introduced.
Dance Diffusion notebook supports unconditional generation, style transfer, and two-clip interpolation
sample-generator repo lists six pretrained models with sample rates of 16 kHz or 48 kHz
GitHub org hosts 3 public repos, including oobleck VAE codecs for neural audio synthesis
Pretrained checkpoints cover piano MAESTRO clips, Song-A-Day vocals, and glitch.cool samples
Discord community link on harmonai.org for contributors and weekly office hours
Open training code lets you fine-tune diffusion models on your own audio libraries
Publishes full training notebooks instead of locking generation behind a hosted app.
Ships multiple pretrained checkpoints at 16 kHz and 48 kHz for different sound types.
Backed by Stability AI with active Discord support for audio researchers.
Dance Diffusion architecture still documented in Stability AI's stable-audio-tools repo.
Homepage is a landing page with little documentation beyond GitHub links.
Requires technical setup in Colab or a local GPU environment.
Newer Stability AI audio products moved to separate repos, so Harmonai feels maintenance-focused.
What is Harmonai.org?
Harmonai.org is a Stability AI lab focused on open-source generative audio tools. The site describes the mission as making music production more accessible by releasing models musicians can run and train themselves.
What is Dance Diffusion on Harmonai?
Dance Diffusion is Harmonai's first generative audio release in the sample-generator repository. The Colab notebook covers random sample generation, style transfer from one audio file, and interpolation between two clips.
Is Harmonai free to use?
Harmonai publishes open-source code and notebooks on GitHub with no listed subscription fee. You run the models locally or in Colab and supply your own compute.
Which pretrained models does Harmonai provide?
Harmonai's Dance Diffusion notebook documents six checkpoints, including maestro-150k for piano at 16 kHz, glitch-440k at 48 kHz, and honk-140k trained on Canada Goose recordings from xeno-canto.
How do I join the Harmonai community?
Harmonai.org links to a Discord server for people building with its audio models. The homepage also points to the Harmonai-org GitHub organization for code and issues.
How does Harmonai relate to Stable Audio?
Harmonai introduced Dance Diffusion, and Stability AI's stable-audio-tools repo still documents the Dance Diffusion U-Net backbone. Newer Stable Audio releases live under Stability AI, while Harmonai keeps the original open training stack.