A developer has released a free collection of 15,834 AI-generated audio samples built with Stability AI's Stable Audio 3 Medium, covering 24 drum categories and 28 pitched instrument types. Both packs are available for immediate download with no GPU required.

What Happened

Developer shiehn published sas-sample-generator on GitHub today: an open-source batch pipeline that uses Stable Audio 3 Medium to produce large libraries of audio samples in two packs:

  • Drums pack: 10,359 one-shots across 24 categories including kicks, snares, hi-hats, and percussion
  • Instruments pack: 5,475 multi-zone pitched samples across 28 instrument categories

The pre-built packs are hosted publicly and free to download. Running the pipeline to generate entirely new packs requires a GPU with 48GB VRAM and a HuggingFace account with model access.

Why It Matters

Stable Audio 3 Medium is a 2-billion-parameter latent diffusion model released by Stability AI earlier this month. It generates variable-length audio from text prompts in under 2 seconds on high-end hardware and supports audio inpainting for targeted editing of existing recordings.

Most AI audio tools focus on full-length music tracks or individual sound design. This pipeline targets the sample library market instead: generating thousands of consistent, quality-gated one-shots that producers can drop directly into a DAW. Every WAV file includes its generation prompt embedded in the metadata, making it straightforward to reproduce or tweak results.

Key Details

  • Pre-built packs: Drums pack (10,359 one-shots, approximately 1.4 GB) and Instruments pack (5,475 samples, approximately 26.6 GB), free direct download
  • Model license: Stability AI Community License allows non-commercial use; commercial projects require a separate license from Stability AI
  • Quality gating: Automated filtering removes clipped, silent, and off-pitch samples before packaging
  • LUFS normalization: All drum samples normalized to -16 LUFS for consistent perceived volume across DAW contexts
  • Pitch accuracy: Instruments generated at 2-4 real source pitches to reduce pitch-shifting artifacts in samplers
  • Metadata: Each WAV file includes its generation prompt in the file metadata and a paired sidecar text file

What to Do Next

Download either or both packs directly from the links in the Key Details section above. No HuggingFace account or GPU is needed to use the pre-built samples in your existing workflow.

To generate fully custom packs using your own text prompts, clone the sas-sample-generator repository and follow the RunPod GPU pod setup guide in the README. You will need approved access to the Stable Audio 3 Medium model on HuggingFace and a cloud GPU with at least 48GB VRAM (RunPod RTX A6000 at approximately $0.49 per hour).