Skip to content

pip and uv

This page changed after phonometry 4.0.0, the latest release on PyPI, so it may describe code that release does not have yet. What is new since then is listed under Unreleased in the changelog.

phonometry installs from PyPI with pip or uv on Python 3.12, 3.13 and 3.14, as one wheel with no compiled code of its own. Both tools land on the same packages; the difference is how they keep a project’s dependencies, and how each is told to take torch from PyTorch’s CPU index.

Checked on 17 November 2026 with phonometry 4.0.0 from PyPI, in the python:3.12-slim, python:3.13-slim and python:3.14-slim images, on Python 3.12.16, 3.13.18 and 3.14.10: pip 26.3 with NumPy 2.5.5 and SciPy 1.18.1, uv 0.13.9 on Python 3.13.18, and torch 2.15.0 from PyTorch’s CPU index.

Terminal window
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install "phonometry[full]"

On Windows the environment is activated with .venv\Scripts\activate.bat in cmd or .venv\Scripts\Activate.ps1 in PowerShell in place of source .venv/bin/activate, here and in the block of a clone below. PowerShell on a Windows client refuses scripts under its default execution policy, so Activate.ps1 runs there once Set-ExecutionPolicy -Scope CurrentUser RemoteSigned has allowed local ones; cmd needs no such step.

python -m pip install phonometry alone brings NumPy and SciPy and nothing else; that is enough for every metric of the library. python -m pip check reports no broken requirement afterwards, with or without extras. Check what the environment imports:

import phonometry
print(phonometry.__version__) # the release you installed
Extra Brings So that
[plot] matplotlib .plot() and the filter response plots draw
[report] reportlab, svglib .report() writes a PDF fiche; its figure panel also needs [plot]
[audio] python-soundfile, whose wheel carries libsndfile except on the platforms named below phonometry.io reads FLAC, AIFF, Ogg/Opus, MP3 and compressed WAV, and writes FLAC
[perf] numba the impulse time weighting runs compiled
[full] all four above every feature but the torch backend
[torch] torch, array-api-compat the FDTD engines step on PyTorch

On a bare python:3.13-slim image for x86-64 or ARM64 no extra needs a system package: every one of them installs and runs from its wheels there, [audio] included. The image’s six other architectures have no wheel of NumPy, SciPy, matplotlib or Pillow, so pip builds them from source there, and [audio] loads the system’s libsndfile. numba supports none of its three 32-bit ones, so [perf] and [full] are not supported there, and torch publishes no wheel for any of the six and no source, so [torch] does not install there from PyPI; Docker and servers says what the rest takes. Without numba the impulse time weighting runs as plain Python and says nothing about it. A feature whose extra is missing says which one to install, with its pip and its conda spelling:

ImportError: Plotting requires matplotlib. Install it with: pip install "phonometry[plot]" (conda: conda install -c conda-forge matplotlib-base)

The oldest release of each dependency the library is tested with, on each Python, is listed in Getting Started. On Windows on ARM, on Linux with musl, on macOS on Intel and on 32-bit Windows, some of those releases have no wheel, and pip builds them from source there, as even a plain install does with SciPy on 32-bit Windows, and as [plot] does with matplotlib on Linux on ARM with musl and on 32-bit Windows; that section names which. numba no longer supports 32-bit Windows, nor 32-bit Linux, so neither [perf] nor [full], which holds it, is supported on either, where python -m pip install "phonometry[plot,report,audio]" installs the other three extras. torch publishes no source to build, so it installs on none of the four from PyPI, and on Windows on ARM from PyTorch’s CPU index, as the next section does. python-soundfile publishes no wheel that carries libsndfile for Linux with musl, nor at its floor for Linux on ARM or Windows on ARM, so [audio] loads the system’s libsndfile there: on Alpine, apk add libsndfile-dev.

[full] leaves [torch] out, because the torch wheels PyPI serves on Linux bring gigabytes of CUDA libraries. On a Linux machine without an NVIDIA GPU, install torch from PyTorch’s CPU index first, then the extra, which keeps the torch it finds, in the virtual environment made above. On Windows on ARM, where PyPI has no torch, the same two commands bring [torch] on Python 3.12 and 3.13, from torch 2.7.0 and 2.8.0:

Terminal window
python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
python -m pip install "phonometry[torch]"

The environment then holds the CPU build of torch and no NVIDIA package, and an FDTD run on Compute(device="torch:cpu") gives the same pressures as the NumPy run. Where a simulation runs covers the other devices.

Terminal window
git clone https://github.com/jmrplens/phonometry.git
cd phonometry
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install .

pip builds the wheel with setuptools and no compiler, and the build also passes under PYTHONWARNINGS=error.

uv pip takes the same requirements as pip, in an environment uv creates:

Terminal window
uv venv
uv pip install "phonometry[full]"

For torch, --torch-backend cpu tells uv to take torch from PyTorch’s CPU index, with no NVIDIA package:

Terminal window
uv pip install "phonometry[torch]" --torch-backend cpu

In a uv project, uv add writes the requirement into pyproject.toml and the lock file, and uv run runs in the project’s environment:

Terminal window
uv init acoustics
cd acoustics
uv add "phonometry[full]"
uv run python -c "import phonometry; print(phonometry.__version__)"

uv add resolves the same way in a project for Python 3.12, 3.13 or 3.14. On 32-bit Windows and 32-bit Linux, where [full] is not supported, name [plot,report,audio] in its place, with uv pip install as with uv add; uv builds from source there what pip does, SciPy, matplotlib and kiwisolver on 32-bit Windows, and on 32-bit Linux the packages Docker and servers names.

Name the index in pyproject.toml, mark it explicit so that only torch is taken from it, and route torch to it:

[tool.uv.sources]
torch = [{ index = "pytorch-cpu" }]
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

Then add torch by name beside the extra:

Terminal window
uv add "phonometry[torch]" torch

The name matters. A source applies to a dependency the project names itself, so uv add "phonometry[torch]" alone, with the same pyproject.toml, locks the torch of PyPI’s default index together with its CUDA and NVIDIA packages, and says nothing about it.