Setup Whisper

Tue, 15 Jul 2025 08:24:28 +0100

I wanted to generate chapter markers from a devlog audio recording using OpenAI’s Whisper, and figured I’d run it locally. Whisper is Python-based, and I’m on Arch. What could go wrong?

Turns out… not much, but it still took a few hops.

Choosing the Right Setup

I already had Python installed, but rather than littering system Python or managing a bunch of ad hoc virtualenvs, I decided to do it properly — with Poetry.

```bash sudo pacman -S poetry poetry new whisper-transcriber cd whisper-transcriber ```

So far so good.

PyTorch + CUDA: the PyPy Pitfall

My first attempt to install torch, torchvision, and torchaudio failed in a confusing way — no versions found at all. The clue was in the command: I’d accidentally run it with pip-pypy3. PyTorch doesn’t build wheels for PyPy. CPython only.

Sorting Out Python Versions

My system Python was 3.13. PyTorch had just released 3.13 wheels for torch, but not yet for torchaudio — version mismatch. I used pyenv to install 3.12 instead:

```bash pyenv install 3.12.3 ```

Updated pyproject.toml:

```toml python = ">=3.12,<3.14" ```

And re-pointed Poetry:

```bash poetry env use $(pyenv prefix 3.12.3)/bin/python ```

Poetry ignored me the first time because 3.13 was still hardcoded. After recreating the environment and verifying the version, I was ready.

PEP 668: the “Externally Managed” False Alarm

Even inside the Poetry shell, Arch’s patched Python threw a --break-system-packages error. This check is meant to protect system Python — but it was firing inside a fully isolated Poetry environment. Safe to ignore. I added the flag:

```bash poetry run pip install torch torchvision torchaudio \ --index-url https://download.pytorch.org/whl/cu121 \ --break-system-packages ```

Worked perfectly.

The Result

```bash poetry run whisper output000.mp3 --model base --output_format json ```

Transcribed 2.5 hours of audio, timestamped segments ready for chapter generation. All local, GPU-accelerated, isolated from system Python, and repeatable.

In Summary

If you’re on Arch and want Whisper with CUDA:

Use poetry + pyenv
Set Python to 3.12 (not 3.13)
Install torch with --break-system-packages and the cu121 index
Whisper just works

A bit of fiddling up front, but now it’s a solid local tool — one less cloud dependency to think about.