gaohz
/
FunASR
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
FunASR/fun_text_processing/inverse_text_normalization/fr/README.md

1.0 KiB
Raw Blame History

Note on French spelling

Due to a 1990 orthographic reform, there are currently two conventions for written French numbers:

  1. Reformed All composite words are joined by a hyphen: e.g. 1122 -> mille-cent-vingt-deux

  2. Traditional Hyphenation only occurs (with exception) for numbers from 17 to 99 (inclusive): e.g. 1122 -> mille cent vingt-deux

As available training data for upstream ASR will vary in use of convention, NeMo's French ITN accomodates either style for normalization e.g.

	python inverse_normalize.py "mille-cent-vingt-deux" --language="fr"  --> 1122
	python inverse_normalize.py "mille cent vingt-deux" --language="fr"  --> 1122

As a result, there exists some ambiguity in the case of currency conversions, namely minor denominations of the dollar e.g.

	300 -> "trois-cents" # Reformed spelling
	300 -> "trois cents" # Traditional spelling
	3 ¢ -> "trois cents" # Valid for both

Cardinals take priority in such cases.

python inverse_normalize.py "trois cents" --language="fr" -> 300