musegauge

Needs a GPU, about 28 GB free disk, Linux x86_64.

musegauge scores a folder of generated music with several existing metrics in one command: Frechet Audio Distance (fadtk), Kernel Audio Distance (kadtk), a CLAP text-audio score (laion-clap) and Audiobox Aesthetics. Each metric runs in its own Python environment, which the tool builds by itself with uv, so tools that need different torch versions can be used together. One run writes a results.json and a short report card for a paper appendix.

Version 0.1.0. Source: https://github.com/aroy1990-dev/musegauge. Not on PyPI; no container image is published. Linux x86_64 is supported; macOS is best effort; Windows is not supported.

Quick start

pip (Python 3.10 or newer):

python3 -m venv .venv && source .venv/bin/activate
pip install musegauge
musegauge doctor
musegauge run --generated ./my_audio --prompts prompts.csv --out ./out

If python3 -m venv fails because ensurepip is missing (a Debian or Ubuntu system Python without the python3-venv package), use another Python, or uv venv.

uvx (no install; needs uv):

uvx musegauge run --generated ./my_audio --out ./out
uvx --from . musegauge --version                         # from a clone

Docker (slim image; not built or tested by the builder, see docker/README.md):

uv build && docker build -f docker/Dockerfile --target slim -t musegauge:slim .
docker run --rm --gpus all -v "$PWD/my_audio:/data/audio:ro" -v musegauge-cache:/cache \
  -v "$PWD/out:/out" musegauge:slim run --generated /data/audio --out /out

The first run builds the plugin environments and downloads the model weights into the cache folder (~/.cache/musegauge, or /cache in the container): about 24 GB of environments and 3.7 GB of weights for all four plugins. musegauge setup --all --fetch-weights does this ahead of time. All install paths: docs/INSTALL.md.

Try it on synthetic audio

From a fresh clone, in a new shell:

git clone https://github.com/aroy1990-dev/musegauge.git && cd musegauge
python3 -m venv .venv && source .venv/bin/activate
pip install .
python tests/make_fixtures.py fixtures
musegauge run --generated fixtures/gen_small --metrics aesthetics.audiobox@1 --out out
cat out/report.md

tests/make_fixtures.py writes synthetic sine-and-noise clips (no real music). The first run builds the Audiobox environment (about 5 GB) and downloads its weights (about 400 MB). Scores on synthetic audio mean nothing about quality.

Input and output

• Input: a folder of .wav, .flac, .ogg or .mp3 files (--generated DIR, with optional --prompts FILE), or a clips file in JSON lines (--manifest FILE).
• Metrics: a suite (--suite t2m-basic, the default, or t2m-full) or --metrics ID,ID. musegauge list shows them; musegauge info METRIC_ID shows a definition and its licences.
• Reference: --reference DIR or --reference bundled:fma_pop (FAD only).
• Output folder: results.json, report.md, per_clip/, logs/, requests/, responses/.

One example output, from docs/examples/report.md (synthetic audio; the numbers mean nothing):

| Metric | Score | Mean ± std | 95% CI | n | Status |
| --- | --- | --- | --- | --- | --- |
| fad.vggish@1 | fad | 25.369 | — | — | ok |
| fad.clap-laion-music@1 | fad | 1.274 | — | — | ok |
| clapscore.laion-music@1 | clap_cosine | 0.232 ± 0.059 | [0.200, 0.262] | 12 | ok |
| aesthetics.audiobox@1 | CE | 2.405 ± 0.557 | [2.115, 2.705] | 12 | ok |
| aesthetics.audiobox@1 | PQ | 6.228 ± 0.827 | [5.757, 6.655] | 12 | ok |

Things to know

• Offline use. --no-fetch blocks downloads from libraries that honour the proxy settings, and sets the Hugging Face offline variables. It is best effort, not a guarantee; for a hard guarantee cut the network outside the tool (for example a container with no network). It was tested only with the proxy method, not with a truly blocked network. Run musegauge setup --fetch-weights first; run --no-fetch never builds an environment.
• Network without --no-fetch. FAD and KAD with VGGish contact github.com on every run (torch.hub). An HTTP error there is reported as NETWORK_ERROR.
• Licences. Every metric's commercial_ok is unknown, so every run warns UNKNOWN_LICENCE. --commercial refuses such metrics. Details: docs/LICENCES.md.
• Comparability. Numbers are comparable only with the same metric ids, reference set, clip length, device and thread setting. fad_inf is random from run to run. See docs/METRICS.md.
• Shared machines. --threads N limits CPU threads in each plugin (docs/INSTALL.md).

Documentation

• docs/INSTALL.md: install paths, offline use, disk space, GPU driver, threads
• docs/METRICS.md: what each metric is and where numbers can differ
• docs/LICENCES.md: code, weights and data licence status per metric
• docs/plugins/: one page per plugin (commands, files, downloads, differences from upstream)
• docs/PLUGIN_GUIDE.md: how to write a plugin
• docs/UPSTREAM_NOTES.md: problems found in the upstream tools
• docs/TESTING_REAL_PLUGINS.md: golden tests with the real tools
• docs/DECISIONS.md: choices made, and every change from the build spec
• docs/VERIFIED_FACTS.md, docs/PROGRESS.md, docs/BUILD_REPORT.md: checks and build record

Licence of this code

Apache License 2.0 (LICENSE; decision D2). The metric packages and model weights that the plugins download have their own licences: docs/LICENCES.md.