techiaith-g2p

The permissive Bangor Welsh grapheme-to-phoneme (G2P) engine: turns Welsh (and code-switched English) text into the phoneme-id sequence consumed by Bangor University's retrained Piper/VITS voices. Dictionary-based, permissively licensed data only, and dependency-free -- it imports nothing beyond the Python standard library, so it drops into an inference server, a screen-reader add-on, or a CLI without pulling in a model runtime.

Install

pip install techiaith-g2p

Use

from techiaith.g2p import BangorG2P

g2p = BangorG2P(english_mode="native")
ids = g2p.text_to_ids("Bore da. Sut wyt ti'n teimlo heddiw?")
print(ids)

english_mode and the package's data_version() must match the target voice's .onnx.json. For the full text-to-audio example, see the model card: https://huggingface.co/techiaith/cy_en_GB-bu_tts

Do not feed this voice to stock piper. Its phoneme_type: "text" path treats phonemes as individual codepoints, one at a time -- it does not understand the multi-codepoint phoneme ids this G2P produces, and the output will be mispronounced. Use the ids from text_to_ids directly against the model's input layer, as the model card describes.

Links

• Voice model: https://huggingface.co/techiaith/cy_en_GB-bu_tts
• Pronunciation dictionary: https://github.com/techiaith/geiriadur-ynganu-bangor
• Source: https://github.com/techiaith/techiaith-g2p
• License: CC0-1.0 — public domain dedication (see LICENSE; the bundled pronunciation data keeps its own BSD-2-Clause terms, see NOTICE)