THIRD_PARTY_NOTICES.md

# Third-party notices and redistribution scope

This is a multi-component release, not a claim that every component shares one
license. The Hub card retains `license: other`; consult the component sources.

* GOPT-derived scorer and preprocessing source retain the original project's
  BSD-3-Clause notice (`LICENSE`, copyright Yuan Gong, 2022). The published
  scorer is the publisher's frozen StreamPA checkpoint, not the original GOPT
  model and not the legacy v6 release. No training data or optimizer is bundled.
* The publisher's fine-tuned Whisper Base weights and their actual tokenizer
  and processor are included. OpenAI's Whisper project publishes code and
  model weights under MIT (`licenses/Whisper-MIT.txt`); the Transformers Hub
  [Whisper Base card](https://huggingface.co/openai/whisper-base) labels its
  converted distribution Apache-2.0. Both upstream notices are retained.
  The included fine-tune is not interchangeable with generic Whisper Base.
* Charsiu frame-classifier code is from `lingjzhu/charsiu` commit
  `13a69f2a22ca0c0962b75cc693399b0ae23a12c9`, under its MIT notice
  (`licenses/Charsiu-MIT.txt`). The Charsiu weight repository does not expose
  sufficiently explicit weight-redistribution terms for this release to
  assign it a license. Its weights/tokenizer are **not redistributed here**.
  `scripts/setup_assets.py` retrieves their official fixed revisions listed
  in `assets/runtime_sources.json`; users must check the upstream terms.
* The pronunciation lexicon contains only isolated normalized word-to-phone
  entries required by the frozen frontend. It is derived from
  [speechocean762](https://www.openslr.org/101/), released by Beijing
  Speechocean and Microsoft under CC BY 4.0. It omits utterance IDs, ordered
  transcripts, recordings, boundaries and human scores. The normalization
  and deduplication are adaptations. Attribution: Zhang et al.,
  *speechocean762: An Open-Source Non-native English Speech Corpus For
  Pronunciation Assessment* (2021), https://arxiv.org/abs/2104.01378 .
  CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ .
* g2p-en 2.1.0 is installed from its official package (Apache-2.0 notice in
  `licenses/G2P-Apache-2.0.txt`). Its runtime uses NLTK CMUdict and English
  POS tagger resources. They are fetched from a fixed `nltk/nltk_data`
  revision, with individual SHA256s, rather than redistributed as corpus
  archives here. Refer to their bundled notices and
  https://github.com/nltk/nltk_data/blob/gh-pages/DATASET-LICENSES.md .
* PyTorch, Transformers, NumPy, librosa, SoundFile, NLTK and safetensors are
  separately installed packages, each subject to its own upstream license.

Synthetic verification audio stays outside the public bundle. This release
does not distribute private recordings, human assessment labels, experiment
predictions, proprietary corpus archives or third-party optimizer states.