Arabic-Indic numerals collapse to 7% while Western digits hold at 73%
The same Arabic sentence, written two ways, is not equally speakable.
Arabic uses two numeral systems. Both are correct, both appear constantly in real copy:
في عام 2026 ... Western digits
في عام ٢٠٢٦ ... Arabic-Indic digits
15 sentences x 3 numeral forms, 45 utterances per engine, scored on one question: could a
listener recover the figure?
Western 2026 73% (fish-audio s2.1-pro) 80% (Apple Majed)
spelled ألفين وستة وعشرين 73% 53%
Arabic-Indic ٢٠٢٦ 7% 80%
The failures are not near-misses. ٢:٤٥ came back as انفافس اسم مفاعس. ٢٨ in a date
came back as تخاناتر.
Apple scores identically on both digit forms because it normalises the numeral before
synthesis, so this is a missing normalisation step rather than a hard problem.
Why it goes unnoticed: TTS demos are prose, and prose has no numbers. The failure only
appears in a finished deliverable with a date, a price or a phone number in it.
Practical rule: normalise every numeral to Western digits before synthesis.
Caveat stated plainly - the absolute percentages understate both engines. The chain is
TTS -> transcription -> a number parser, and errors anywhere are charged to the voice.
The 7% vs 80% gap is the robust part; transcription noise cannot manufacture that.
The set, the scorer and the per-utterance results are here, CC BY 4.0. It is
engine-agnostic - adding another takes a few lines. I would like to know which other
engines get ٢٠٢٦ right.
Correction, 20 August 2026 — the spelled row in the post above is wrong. Revised figures below.
Four gaps in my own scorer (arnum.py) were marking correct Arabic speech as a failure. They were published as known limitations in the dataset card before they were fixed; they are now fixed, so the affected numbers have to be revised in public rather than quietly.
What was wrong:
- Fractions of an hour were absent.
النصف/الربعwere in no dictionary, so6:30spoken correctly asالساعة السادسة والنصفparsed as{6}and scored LOST. Now read as 30 and 15 — only when the sentence containsساعة, becauseنصفmeans "half" in every other context and mapping it unconditionally would be a new bug, not a fix. - Tanween left a trailing alef. The diacritic was stripped but not the alef under it, so
سبعونًاmissed the tens table and674parsed as604. - A two-word ordinal could not match.
الثانية عشرة(twelfth) can only live in a dictionary pre-joined; the pair is now joined before lookup. Same fix stopsالحادية عشرةreading as 11 + 10 = 21. - The group never closed after a clock hour.
الثانية عشرة وخمسة وأربعينaccumulated 12+5+40 into a single57instead of{12, 45}.
Old → new:
| cell | before | after |
|---|---|---|
fish spelled |
11/15 (73%) | 13/15 (87%) |
Apple spelled |
8/15 (53%) | 9/15 (60%) |
| fish overall | 23/45 (51%) | 25/45 (56%) |
| Apple overall | 32/45 (71%) | 33/45 (73%) |
ArTST spelled |
4/15 (27%) | 4/15 (27%) — unchanged |
| every digit-form cell | — | unchanged |
Three rows flipped in total, all spelled, all LOST → OK. No row went the other way.
Two things worth stating about the method, because they are what make this a correction rather than a new run:
- Nothing was re-synthesised. Every score was re-derived from the transcripts already stored in the results files, so the audio underneath the change is identical and the only variable is the parser.
rescore.py(now in the repo) reproduces it, and refuses to write if a digit-form score moves — a spelled-form scorer fix must not be able to touch them. - The headline finding is untouched. It lives entirely in the digit forms — 7% on Arabic-Indic vs 73% Western on fish, 80%/80% on Apple — and not one of those six cells moved. Neither did the ranking or any conclusion: fish is still stronger than Apple on the spelled form, Apple is still the only form-agnostic engine, ArTST still scores 0/30 on digits.
test_arnum.py went from 19 tests to 34; the original 19 still pass, which is the actual safety net here.
Full old → new per cell, and the reasoning for the time gate: https://huggingface.co/datasets/syamjithnk/arnum-tts#changelog. The write-up at https://syamjithnk.com/arabic-tts-numerals carries the same correction.