Old Occitan handwriting

The model was trained with Old Occitan texts, mainly from the 13th and 14th centuries from the Languedoc or Provence. Overall, 7 texts were used for this purpose with a total of about 830 pages and about 190,000 tokens.

The following texts were used:

Las Leys d’amors (= Bibliothèque municipale de Toulouse, cote 2883/-4),
Le roman de Flamenca (=Bibliothèque municipale de Carcassonne, cote 34),

and from the Bibliothèque nationale de France (BnF):
La vida de sant Enimia (= Arsenal 6355), NAF 11180, NAF 1050, Latin 1139 and Français 846.

The model makes independent word segmentations and is capable of resolving the common abbreviations. Punctuation is left as it is in the manuscript (i.e. elided vowels, e.g. in the article, are not replaced by apostrophes). The CER on the training set is 2.6%, on the validation set 3.51%.