danielrosehill/Gefilte-Hebrew-Sentences

Gefilte Hebrew Sentences

Modern, everyday Hebrew sentences with nikud and an English translation, graded 1 to 5 by how hard they are to read aloud for a non-native speaker, plus recordings of them.

Why it exists. Most Hebrew speech data comes from native speakers. This set is for a personal model of one immigrant's Hebrew: accented, with imperfect pronunciation. That speech is part of everyday Hebrew too. The nikud is there so the reader can get the vowels right, and the grades let him work up from the easy sentences. The model it feeds is Gefilte.

Tabs

Tab	Rows
sentences	60,696 sentences, with takes = 0 until read.
takes	One row per recording, joined on sentence_id. Empty until the first push.

Columns (sentences)

id, text (Hebrew without nikud), text_nikud, english, difficulty (1 easiest to 5), difficulty_score, syllables, split, takes, tatoeba_id, tatoeba_author, english_tatoeba_id, english_author.

How it was made

• Sentences and translations: Tatoeba, Hebrew sentences with an English link. Long sentences, unusual characters and the repetitive "Tom" sentences were dropped.
• Nikud: machine-made with dicta-il/dictabert-large-char-menaked (CC BY 4.0), then a string fix that moves holam and shuruq onto the vav. It is not checked by a human, so expect some wrong vowels.
• Grades: a script, not a person: length, syllables, rare words and sounds that are hard from an English start. The five grades are equal fifths of this set, so they rank sentences against each other.

Licence and attribution

Tatoeba sentences are CC BY 2.0 FR. Each row names its Tatoeba id and author. A row's author may be \\N where Tatoeba lists none. Recordings are by Daniel Rosehill.