techiaith-g2p
Uned Technolegau Iaith, Prifysgol Bangor / Language Technologies Unit, Bangor University

The code in this package is dedicated to the public domain under CC0 1.0 Universal (see
LICENSE) by Prifysgol Bangor University.

CC0 waives the Language Technologies Unit's own rights; it cannot waive anyone else's. The
bundled pronunciation data is third-party and keeps its own terms, so the attributions below
still apply to anything you redistribute:

* Geiriadur Ynganu Bangor (the Bangor Pronouncing Dictionary), BSD-2-Clause, Copyright (C)
  2005-2025 Prifysgol Bangor University --
  techiaith/g2p/data/geiriadur-ynganu-bangor/ (see its LICENSE and README), including
  bangordict.dict, bangordict.en.dict, bangordict.xx.dict, cmudict.dict, and phoneset.md.
  https://github.com/techiaith/geiriadur-ynganu-bangor

  Its README asks: "If you make use of or redistribute this material we request that you
  acknowledge its origin in your descriptions." This NOTICE is that acknowledgement: the
  origin of every file in this directory is the School of Linguistics and the Language
  Technologies Unit, Bangor University.

  One file in that directory, cmudict.dict, has a second origin worth calling out on its
  own: it is 119,305 English words and pronunciations, e.g.
  "'cause (foreign,en,cmu) k @ z /kəz/", sourced from CMUdict and re-transcribed into the
  Bangor phoneset. Attribution is therefore dual -- Bangor for the file and its BSD-2-Clause
  distribution, Carnegie Mellon University for the underlying English word list and
  pronunciations (see the CMUdict entry below).

* CMUdict, 2-clause BSD, Copyright (C) 1993-2015 Carnegie Mellon University --
  techiaith/g2p/data/cmudict/LICENSE, and the derived
  techiaith/g2p/data/english/cmudict_native.dict.

* Unicode CLDR and the Unicode emoji data files -- the Welsh emoji names in
  techiaith/g2p/data/emoji_cy.tsv (and the generated techiaith/g2p/c/cy_emoji_data.h) are
  derived from:

    - Unicode CLDR, common/annotations/cy.xml and common/annotationsDerived/cy.xml,
      release-48-2. SPDX-License-Identifier: Unicode-3.0.
      Copyright (C) 1991-2026 Unicode, Inc.
    - Unicode emoji-test.txt, version 17.0, which determines WHICH sequences are emoji.
      Copyright (C) 2025 Unicode, Inc. https://www.unicode.org/terms_of_use.html

  Modifications: restricted to the sequences Unicode marks fully-qualified or component;
  zero-width-joiner sequences excluded; CLDR's ": " qualifier separator collapsed to a
  space so it is not read aloud. Both licences permit this. Regenerate with
  scripts/gen_emoji_cy.py -- see techiaith/g2p/data/upstream/README.md.

  This table replaced one extracted from espeak-ng's cy voice, which is GPLv3 and therefore
  incompatible with shipping only permissive components. No espeak-ng data remains.

No voice model is bundled in THIS package. The models carry their own licences:
https://huggingface.co/techiaith/cy_en_GB-bu_tts

Note for anyone packaging a model alongside this G2P: the shipped 3-speaker
cy_en_GB-bu_tts model is warm-started from the CSTR VCTK Corpus and is therefore
**CC BY 4.0**, not public domain. It carries an attribution obligation that must appear in
your artifact's own NOTICE. See LICENSING.md.

## Part-of-speech tagger training data (`techiaith/g2p/data/pos/pos_{cy,en}.bin`)

The POS models that select heteronym readings were trained on:

- **MASC: the Manually Annotated Sub-Corpus of American English** (version 3.0.0), Open American
  National Corpus, https://anc.org/data/masc/ — licensed under the Creative Commons
  Attribution 3.0 United States License (CC BY 3.0 US). Ide, N., Baker, C., Fellbaum, C., Fillmore,
  C. and Passonneau, R. (2008), "MASC: the Manually Annotated Sub-Corpus of American English",
  LREC 2008. Used as training data for `pos_en.bin` and, as word frequencies, for the language
  prior `data/lang/lang_prior.tsv`; no MASC text is redistributed. Both are derived works and are
  distributed with this attribution.
- **brawddegau-tagiedig** (Welsh tagged sentences), Uned Technolegau Iaith, Prifysgol Bangor,
  https://github.com/techiaith/brawddegau-tagiedig — CC0-1.0. Training data for `pos_cy.bin` and,
  as word frequencies, for `data/lang/lang_prior.tsv`.

No accent or pronunciation data in this package is derived from either corpus; they contribute
part-of-speech tags only.
