openai/whisper

Dependencies

Runtime and development dependencies, with concrete usage notes.

Runtime dependencies

Declared in pyproject.toml and requirements.txt.

Package	Where it's used
`more-itertools`	`whisper/normalizers/english.py` — `windowed` for the spelled-out-number sliding window.
`numba`	`whisper/timing.py` — `@numba.jit(nopython=True, parallel=True)` for `dtw_cpu` and `backtrace`.
`numpy`	Universal: array math, mel filters, alignment-head masks, DTW.
`tiktoken`	`whisper/tokenizer.py` — wraps `tiktoken.Encoding` with Whisper's special tokens.
`torch`	The model itself, plus `torch.stft` in `whisper/audio.py`. SDPA used opportunistically.
`tqdm`	`whisper/__init__.py:_download` (download progress) and `whisper/transcribe.py` (frame progress).
`triton>=2`	x86_64 Linux only. `whisper/triton_ops.py` — DTW and median-filter kernels. Imported lazily; if missing, the CPU implementations are used.

There is no networking, web framework, or database driver. The only system dependency is ffmpeg, which is invoked by whisper/audio.py:load_audio via subprocess.run.

Development dependencies

Declared as the dev extra in pyproject.toml.

Package	Used for
`black`	Code formatting. Pre-commit hook + CI gate.
`flake8`	Lint (`--max-line-length 88 --ignore E203,E501,W503,W504`).
`isort`	Import ordering (`--profile black -l 88 --trailing-comma --multi-line 3`).
`pytest`	Test runner.
`scipy`	`tests/test_timing.py` uses `scipy.ndimage.median_filter` to verify the CPU/CUDA median filter matches.

Build / packaging

Package	Where
`setuptools>=61.2`	`[build-system].requires` in `pyproject.toml`. Required for PEP 621 packaging.
`setuptools-rust`	Optional, only when `tiktoken` has no prebuilt wheel for the platform.
`build`	Used by `.github/workflows/python-publish.yml` to produce the sdist.
`twine`	Same workflow, to upload to PyPI.

Pre-commit

Hooks pinned to immutable commit hashes in .pre-commit-config.yaml:

pre-commit-hooks — check-json, end-of-file-fixer, trailing-whitespace, mixed-line-ending, check-added-large-files.
psf/black — formatter.
pycqa/isort — imports.
pycqa/flake8 — lint.

CI dependencies

GitHub Actions versions are pinned by SHA in .github/workflows/test.yml and managed by Dependabot (.github/dependabot.yml, grouped weekly).

The test matrix uses conda-incubator/setup-miniconda to install ffmpeg, then pip install for everything else. PyTorch CPU wheels are pulled from https://download.pytorch.org/whl/cpu.

Why these and not others

No librosa: the mel filterbanks are precomputed and committed as whisper/assets/mel_filters.npz; torch.stft does the spectrogram. This avoids a heavy dependency.
No ffmpeg-python: dropped in v20230918 (#1242) in favor of a direct subprocess.run(["ffmpeg", ...]) call.
No transformers: the original tokenizer used HuggingFace; replaced by tiktoken in v20230314 (#1044).