Development

The test suite runs every recipe end to end on a tiny random model, so it needs no download and no GPU and finishes in about a minute. CI runs the linter, the type checker in strict mode and the tests on every pull request; the same three commands run locally.

Rules a change should follow:

• A stage must be safe to run again in a directory that holds the files of an interrupted run: reuse or replace them, never fail on them.
• Anything that decides a stage's result goes into its fingerprint; anything that only says how to run (the tracker, the budget, the output directory) stays out.
• Bump a version when the code of a method, backbone, head or converter changes what the same inputs produce; the experiments that use it rebuild, every other cache stays.
• Documentation is generated from the README files and the code (this site included). Edit those, not the generated pages.

uv run pytest tests/core tests/modeling  # runner, progress, validation gate, API, every recipe on a tiny random model (wakeword: a fake front end)
uv run pytest tests/data                # data recipes with fake teachers, a fake Piper and a fake Whisper
uv run ruff check . && uv run mypy      # what CI runs on every pull request, besides the tests

Licence: MIT. Models and datasets a config names keep their own licences.

Source: README.md. This page is generated; edit the source and run npm run sync.