Lexicon explorer

The code-switching sentences Hebrish trains on are written around a list of Hebrew words and Israeli names: arnona, kupat holim, a Jerusalem street, a politician's surname. That list is the lexicon. Each entry has one canonical spelling in Latin letters, following the Hebrew-Latin-Token-Classifier romanization, and every transcript uses that spelling, so the model learns to write the word one way.

Being in the lexicon means sentences were written for a word. It does not mean the model recognises it. The recorded count is how many recordings in the public dataset (train and test) have a transcript containing the word, found by text match. Many words have no recording yet, and every recording is one speaker's voice.