Grapheme to Phoneme models for Norwegian Bokmål

This repo contains code to run G2P models for Norwegian bokmål¹, which produce phonemic transcriptions for close-to-spoken pronunciations (such as in spontaneous conversations: spoken) and close-to-written pronunciations (such as when reading text aloud: written) for 5 different dialect areas:

1. East Norwegian (e)
2. South West Norwegian (sw)
3. West Norwegian (w)
4. Central Norwegian (Trøndersk) (t)
5. North Norwegian (n)

Setup

pip install nb_g2p

Usage

>>> import nb_g2p
>>> list(nb_g2p.transcribe("hei på deg!"))
[('hei', 'H AEJ1'), ('på', 'P OAH0'), ('deg', 'D AEJ1')]

Transcription standard

The G2P models have been trained on the NoFAbet transcription standard which is easier to read by humans than X-SAMPA. NoFAbet is in part based on 2-letter ARPAbet and is made by Nate Young for the National Library of Norway in connection with the development of NoFA, a forced aligner for Norwegian. The equivalence table below contains X-SAMPA, IPA and NoFAbet notatations.

X-SAMPA-IPA-NoFAbet equivalence table

Unstressed syllables are marked with a 0 after the vowel or syllabic consonant. The nucleus is marked with a 1 for tone 1 and a 2 for tone 2. Secondary stress is marked with 3.

License

These models are shared with a Creative_Commons-ZERO (CC-ZERO) license, and so are the lexica they are trained on. The models can be used for any purpose, as long as it is compliant with Phonetisaurus' license.