Skip to content

Jabberjay v0.1.3

Latest

Choose a tag to compare

@github-actions github-actions released this 05 Nov 18:57
Immutable release. Only release title and notes can be modified.
54f200b

Added

  • DF Arena 1B model (Model.DFArena1B) β€” Idiap's
    Speech-Arena-2025/DF_Arena_1B_V_1
    (RAPTOR, arXiv:2603.06164): a
    Wav2Vec2-XLS-R-1B encoder with attention-gated layer fusion and a
    Conformer head, the strongest open-weights system on the
    Speech DF Arena
    leaderboard (pooled EER 9.52%; 0.91% on In-the-Wild). Implemented natively
    rather than via trust_remote_code: the checkpoint is pinned to a
    reviewed revision, loaded with weights_only=True and strict=True, and
    built on the meta device so the 1B parameters aren't randomly initialised
    first. The weights are licensed for non-commercial use only; ~4.6 GB
    download
  • download_pretrained_model(..., revision=) β€” optional commit pin
  • preprocess(..., preemphasis=) in Models/Spectra/shared.py β€” DF
    Arena shares Spectra's resample + 64600-sample pad/crop but not its
    preemphasis
  • Opt-in integration tests (pytest -m integration) β€” run DF Arena 1B
    against its real pinned checkpoint on the bonafide/spoof reference clips,
    proving the native port loads with strict=True and reads the logits in
    the right order. Deselected by default and skipped unless the weights are
    already cached, so a plain pytest never triggers the ~4.6 GB download;
    the weekly integration workflow runs them after the model sweep with
    JABBERJAY_REQUIRE_WEIGHTS=1, which turns a missing checkpoint into a
    failure rather than a silent skip
  • Python 3.15 support β€” added to the CI test matrix, the PyPI
    classifiers, and black's target-version. Verified locally on 3.15.2: the
    unit suite passes and the full model sweep gives results identical to 3.14.
    PyYAML has no 3.15 wheels yet, so it builds from source on install (pure
    Python, without the optional libyaml C extension). 3.15 is also now the
    default development version: .python-version, ty's target, the lint
    and docs CI jobs, the weekly integration sweep, and the uploaded coverage
    report all use it. 3.12–3.14 remain fully supported and tested
  • just test-integration β€” runs the opt-in integration tests

Fixed

  • README links on PyPI β€” the social preview image, the example-script
    links, and the CONTRIBUTING.md link were relative paths, which break on
    the PyPI project page; they now point at GitHub
  • Stale model lists β€” the CITATION.cff and .zenodo.json abstracts
    listed only the original seven model families (now including the Spectra
    family and DF Arena 1B), and the README / getting-started / CLI examples
    and examples/choosing_a_model.py were missing DF Arena 1B
  • Docs accuracy β€” the score breakdown was described as "transformer
    models" only (it covers every model except Classical and RawNet2), the
    README's sample CLI output didn't match the real output, and the
    getting-started comparison table truncated longer model names

Security

  • msgpack floor added at >=1.2.4 β€” closes
    CVE-2026-114613
    (high), an unbounded memory leak in the msgpack C extension on
    strict_map_key validation failure. msgpack is a transitive
    dependency (via librosa), so it is now declared directly: the lockfile
    moving to 1.2.4 alone left existing environments free to keep the
    vulnerable 1.2.2

Changed

  • Dependency floors raised to current releases β€” huggingface-hub>=2.4.0
    (previously >=1.31.0, still admitting the untested 1.x line),
    matplotlib>=3.11.2, numpy>=2.5.4, pillow>=13.0.0,
    scikit-learn>=1.9.1, torch>=2.15.0, torchaudio>=2.11.1,
    transformers>=5.22.1, so every install resolves to the versions the
    full model sweep was verified against
  • Model licence policy β€” models no longer have to be Apache 2.0 or MIT:
    other licences, including non-commercial, are accepted when the licence
    is flagged in the README and docs (as DF Arena 1B's is).
    CONTRIBUTING.md and the model-request issue template are updated to
    match, and the contributor guide now covers safe weight loading and every
    place a new model must be listed
  • ty now type-checks the whole repository (tests, examples, scripts),
    not just src/, in CI, pre-commit, and just type-check; the six test
    diagnostics this surfaced are fixed
  • Smaller sdist β€” social_preview.png is no longer packaged (580 KB β†’
    436 KB)

Dependencies

  • torch 2.14.1 β†’ 2.15.0, torchaudio 2.11.0 β†’ 2.11.1, transformers
    5.20.1 β†’ 5.22.1, huggingface-hub 2.2.0 β†’ 2.4.0 (lockfile; verified with
    the full model sweep β€” every result is identical to v0.1.2), plus the
    CUDA 13.0 β†’ 13.2 runtime wheels and triton 3.9.0 that torch 2.15
    pulls in on Linux
  • Lockfile refreshed to the latest compatible release of every package
    (~50 transitive updates, including python-dateutil 2.9 β†’ 4.0 via
    matplotlib, numba 0.68.1, contourpy 1.4.0)
  • Dev: ruff 0.17.3, ty 0.0.97, pytest 9.2.0, mkdocs-material 9.8.0,
    mkdocstrings 1.0.8 (pymdown-extensions 12.3, watchdog 7.1.0);
    astral-sh/setup-uv 10.3.3. Dev dependency floors raised to these
    versions and the ruff-pre-commit hook rev bumped to v0.17.3 to match