Arabic-Indic numerals silently break TTS: 7% vs 73% on the same sentences
The same Arabic sentence written with Western digits (2026) versus Arabic-Indic digits (٢٠٢٦).
Both are correct Arabic and both appear constantly in real copy.
On one engine: 73% of numbers recoverable with Western digits, 7% with Arabic-Indic. Not
mispronunciation - ٢:٤٥ came back as انفافس اسم مفاعس. A second engine handles both at 80%,
so it is a missing normalisation step, not a hard problem.
Nobody notices because TTS demos are prose, and prose has no numbers. It surfaces in finished
work, where a date or a price is spoken as noise.
The fix is a one-line mapping (٠١٢٣٤٥٦٧٨٩ -> 0123456789) before synthesis. The problem is
knowing you need it.
Set, scorer and per-item results are in this repo, CC BY 4.0. The harness takes a new engine in
a few lines - if you work on Arabic speech, I would like to know which engines get ٢٠٢٦ right.
Caveat stated up front: the absolute percentages understate both engines. The scorer counts a
number as recovered only on an exact match after normalisation, so near-misses score zero. The
gap between the two numeral systems is the finding, not the absolute values.