Datasets:

facebook

/

multilingual_librispeech

like 50

The French train set has a lot of misspellings in transcriptions and cannot be used for supervised training.

Noticed this because a student tried to fine-tune an ASR model on this training set and WER of their model increased greatly.

E.g. first utt in training set where ~40% of words are misspelled: "le waikato sort de ce lac apr s l'avoir travers dans toute sa largeur or du confluent au lac le fleuve se dveloppe sur un parcours de cent vingt milles environ paganel s'adressant en franais john mangles pour ne pas atre compris des sauvages le pria d'estimer la vitesse du canot john la porta trois milles peu pr s par heure"

The text seems to have been corrupted around characters with diacritics
e.g. "après" -> "apr s". "français" -> "franais"