musegauge
Score a folder of generated music with several existing metrics in one command.
Needs: a GPU, about 28 GB of free disk, Linux x86_64. Tested on Linux x86_64 only. Other systems are not tested.
musegauge runs these existing metrics:
| Metric | Tool |
|---|---|
| Frechet Audio Distance | fadtk |
| Kernel Audio Distance | kadtk |
| CLAP text-audio score | laion-clap |
| Audiobox Aesthetics | Audiobox Aesthetics |
Each metric runs in its own Python environment. The tool builds these by itself with uv, so
metrics 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 0.1.4. On PyPI. Source: https://github.com/AMAAI-Lab/musegauge.
Container image: ghcr.io/aroy1990-dev/musegauge.
Authors
musegauge was created and is maintained by Abhinaba Roy (@aroy1990-dev), AMAAI Lab, Singapore University of Technology and Design (SUTD). Contact: abhinabaroy1990@gmail.com
Quick start
Docker (published image)
No install is needed beyond Docker. Python, uv and musegauge are inside the image.
docker pull ghcr.io/aroy1990-dev/musegauge:0.1.2
docker run --rm --gpus all -v "$PWD/my_audio:/data/audio:ro" -v musegauge-cache:/cache \
-v "$PWD/out:/out" ghcr.io/aroy1990-dev/musegauge:0.1.2 \
run --generated /data/audio --out /out
The image 0.1.2 was built and tested on 2026-10-06 and 2026-10-07 on a Linux x86_64 machine
with Docker 27.3.1. doctor, an offline run (--network none --no-fetch) and GPU use all
passed. The tests found a cache-volume bug, which is fixed in 0.1.2. See docker/README.md
and docs/PROGRESS.md (M7).
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 (build it yourself)
From a clone. docker/README.md has the full commands and the published-image alternative.
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
First run
The first run builds the plugin environments and downloads the model weights into the cache
folder (~/.cache/musegauge, or /cache in the container). For all four plugins this is
about 24 GB of environments and 3.7 GB of weights. musegauge setup --all --fetch-weights
does this ahead of time. All install paths are in docs/INSTALL.md.
Try it on synthetic audio
From a fresh clone, in a new shell:
git clone https://github.com/AMAAI-Lab/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,.oggor.mp3files (--generated DIR, with optional--prompts FILE), or a clips file in JSON lines (--manifest FILE). - Metrics: a suite (
--suite t2m-basic, the default, ort2m-full) or--metrics ID,ID.musegauge listshows them.musegauge info METRIC_IDshows a definition and its licences. - Reference:
--reference DIRor--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, so 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-fetchblocks 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. Runmusegauge setup --fetch-weightsfirst.run --no-fetchnever 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 asNETWORK_ERROR. - Licences of the metrics. Every metric's
commercial_okisunknown, so every run warnsUNKNOWN_LICENCE.--commercialrefuses such metrics. Details:docs/LICENCES.md. - Comparing numbers. Numbers are comparable only with the same metric ids, reference set,
clip length, device and thread setting.
fad_infis random from run to run. Seedocs/METRICS.md. - Shared machines.
--threads Nlimits CPU threads in each plugin (docs/INSTALL.md).
Documentation
| File | Contents |
|---|---|
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 |
Citation
If you use musegauge in your work, please cite it:
@software{musegauge,
author = {Roy, Abhinaba},
title = {musegauge: scoring generated music with existing metrics in one command},
year = {2026},
version = {0.1.4},
url = {https://github.com/AMAAI-Lab/musegauge}
}
Licence
Copyright 2026 Abhinaba Roy.
Licensed under the Apache License, Version 2.0. See LICENSE for the full text
and NOTICE for the attribution notice. You may use, change and share this code
under those terms, but you must keep the copyright and NOTICE text.
The metric packages and model weights that the plugins download have their own licences. See
docs/LICENCES.md.
Metadata
Release files for musegauge 0.1.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| musegauge-0.1.4.tar.gz | 520.3 kB | Without provenance | Details for musegauge-0.1.4.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| musegauge-0.1.4-py3-none-any.whl | Python 3 | none | any | Details for musegauge-0.1.4-py3-none-any.whl |
Total release size: 856.7 kB
Release files / musegauge-0.1.4.tar.gz
| Download URL | musegauge-0.1.4.tar.gz |
|---|---|
| Size | 520.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
be4d3f73d8c8bff1f7905d8ec379714c93f534579e54e154db8d8b246d5baf64
|
|
BLAKE2b-256 checksum How to use checksums |
63a48c28c13ced3f0922ed937c2206f713066a6b6b1cc1625698fca901bec3f4
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.9.4
|
Release files / musegauge-0.1.4-py3-none-any.whl
| Download URL | musegauge-0.1.4-py3-none-any.whl |
|---|---|
| Size | 336.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
2c93115ca248524ebe91372e2617c9ef5a40ab37ec5df862ed6cd0585d793497
|
|
BLAKE2b-256 checksum How to use checksums |
e9f037aa872f8f2e1ec196604c6f8952eb6d299423e04463a91d2bad17d01f4d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.9.4
|
Security for musegauge 0.1.4
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.