syamjithnk/arnum-tts
Viewer • Updated • 180 • 300
Arabic text & speech quality
Arabic that looks right but is broken: numerals in TTS and OCR, PDF extraction, rendering, presentation forms in public corpora.
Note 45 utterances × 3 numeral forms. fish: 11/15 Western vs 1/15 Arabic-Indic digits; ArTST 0/30 on digits.
arabic-lint
🔤Paste Arabic, watch a real tokenizer turn it into [UNK]
Note Paste Arabic: spot stored presentation forms and watch AraBERT turn them into [UNK]. In-browser.
syamjithnk/arabic-corpus-audit
Viewer • Updated • 341 • 85Note 341 Hub datasets scanned for stored Arabic presentation forms; re-sample finds them in HPLT/DocHPLT and FinePDFs-edu.
syamjithnk/arpdf
Viewer • Updated • 60 • 116 • 1Note PDF round-trip benchmark: does Arabic survive generate → extract?
syamjithnk/arshape
Viewer • Updated • 60 • 54Note Rendering benchmark: 3 fonts × 5 strings × 4 rendering paths.
syamjithnk/ocr-numeral-fidelity
Viewer • Updated • 7k • 75Note OCR numeral fidelity: 7,000 trials, five models. Value accuracy and digit-script preservation are different outcomes.