Offline AI Music Prototype Lab Protocol: Reproducibility and Licence Records
An offline sound-prototype lab needs more than a promising render: it needs a record another teammate can rerun, inspect and stop before a prototype leaks into a commercial handoff. This is a working protocol for that record, using Stable Audio Open and MusicGen as documented examples—not another winner-picking comparison.
Quick take: establish one experiment record before comparing outputs. It should capture the model revision, prompt, any reference audio, generation parameters, dependencies, output hashes and an explicit distribution hold. Stable Audio Open and MusicGen expose different documented interfaces and licensing boundaries; neither turns a lab file into a default commercial music library.
Last verified: 2026-09-21. This guide compares published project documentation, not a matched hands-on benchmark. Score: withheld until the same prompts, hardware, versions, latency measurements and rights review have been tested side by side.
Start with an experiment record
Stable Audio Open’s model card describes variable-length stereo audio up to 47 seconds at 44.1 kHz, with text conditioning and local Python examples. Its stated limitations say it is stronger for sound effects and field recordings than music, does not generate realistic vocals, and may need prompt engineering. That makes it a plausible first branch for a prototype that needs an impact, ambience, transition or short texture—not a promise of a finished song.
MusicGen Small is a 300M text-to-music checkpoint. Its model card documents text and audio-prompt conditioning, local use through Transformers or AudioCraft, and a 32 kHz waveform path. The project therefore gives a lab a more music-specific starting point, but it does not eliminate setup, repeatability or licensing work.
Minimum experiment record
| Check | What the published documentation supports |
|---|---|
| Brief | State duration, intended scene, acceptance test and prohibited references. |
| Model identity | Record repository, revision, checkpoint, licence label and access condition. |
| Run setup | Save command or notebook, dependency lockfile, device, seed if supported, duration and generation settings. |
| Source inputs | List every prompt and any reference audio; keep consent and provenance evidence with the run. |
| Output evidence | Keep file hashes, listening notes, known defects and a link to the immutable lab storage location. |
| Release gate | Mark the output “prototype only” until the current weight/model terms and distribution plan are reviewed. |
A practical offline lab protocol
- Write one brief that names duration, intended scene, prohibited references and acceptance criteria. Do not start with a request to imitate a named recording or artist.
- Run a small fixed prompt set. Save the exact command or notebook, model revision, package lockfile and generated file hashes alongside the listening notes.
- Separate sound-design and musical tests. Stable Audio Open’s own model card flags music and vocal limitations; test those limits rather than treating a good ambience result as a general music result.
- Check the rights path before sharing outside the lab. MusicGen’s published model-weight licence is non-commercial. For Stable Audio Open, read the then-current Stability licence and map it to the organisation and intended distribution.
- Promote only a documented candidate. A cue that sounds promising but has no reproducible run record or cleared usage path remains a prototype.
Who each route suits
Stable Audio Open as a documented example: its card describes short stereo 44.1 kHz output and separately flags limitations around vocals and music. Capture that model version, the access condition and the exact text-conditioning settings if your lab uses it.
MusicGen as a documented example: its card describes text and audio-prompt conditioning with a local Transformers or AudioCraft path. Capture which checkpoint and interface were used, and treat the model-weight licence separately from AudioCraft’s code licence.
The protocol does not replace: a ready-cleared catalog licence, legal advice, consent for reference audio, or an output-quality evaluation. Those are separate approvals, not fields that a notebook can fill automatically.
Risks and limits to keep visible
The licence labels are not interchangeable. AudioCraft’s repository states that its code is MIT while its model weights are CC-BY-NC 4.0; using the code’s permissive label to infer a commercial right in the weights would be incorrect. Stable Audio Open’s model card labels its model licence as the Stability AI Community License and routes commercial use to Stability’s licence page. Terms, revenue thresholds and permitted deployment patterns can change, so legal review is still needed for a real product or client delivery.
Also distinguish a reproducible technical result from a safe release. An output hash, seed and dependency record show how a file was made; they do not establish that the prompt, any reference audio, the generated result or the final distribution is cleared for every territory and platform.
Sources and update trail
- Stable Audio Open 1.0 model card — output format, intended use, limitations and licence route.
- Stable Audio Tools repository — local installation and inference workflow.
- Stability AI licence information — current commercial-use guidance.
- MusicGen Small model card — checkpoint details, local usage and model-weight licence.
- AudioCraft repository — codebase, installation and licence boundary.
- AudioCraft MIT code licence — code-licence text.
For a broader rights checklist before a client handoff, see our AI music commercial-rights guide. For a product-choice discussion, read the existing MusicGen vs Stable Audio Open comparison.