ArNum-TTS — does the number survive?

Author: Syamjith NK License: CC BY 4.0 (data) · MIT (code)

A small evaluation set for one narrow, practical question: when an Arabic text-to-speech system reads a sentence containing a number, can a listener recover the number?

Not naturalness. Not accent. Not prosody. A voice can be beautiful and still be unusable for anything containing a date, a price or a percentage — and that failure is invisible in every TTS demo, because demos do not contain numbers.

Write-up: One of these reads ٢٠٢٦ as a year. The other reads it as noise.

Why it exists

Arabic is written with two numeral systems. The same year is 2026 or ٢٠٢٦, and both are correct Arabic. Nothing in the literature says whether TTS systems handle them equally, and every Arabic deliverable — a government awareness film, a bank's IVR, a news read — is full of them.

Design

15 sentences × 3 numeral forms = 45 utterances. Each sentence is identical across the three forms except for how the number is written:

form	example
western	في عام 2026
arabic_indic	في عام ٢٠٢٦
spelled	في عام ألفين وستة وعشرين

Holding the sentence constant is the whole point: any difference in score is caused by numeral form and nothing else. Categories are the ones that break real work — years, percentages, decimals, currency, times, dates, phone numbers, ordinals, ranges, large counts.

Results — three engines

numeral form	fish s2.1-pro-free	Apple Majed (ar_001)	ArTST speecht5_tts_clartts_ar
western 2026	11/15 (73%)	12/15 (80%)	0/15 (0%)
arabic_indic ٢٠٢٦	1/15 (7%)	12/15 (80%)	0/15 (0%)
spelled ألفين وستة وعشرين	13/15 (87%)	9/15 (60%)	4/15 (27%)
overall	25/45 (56%)	33/45 (73%)	4/45 (9%)

fish + Apple measured 2026-08-18 · ArTST added 2026-08-20 · spelled-form figures revised 2026-08-20 by a scorer correction — see the Changelog.

Three engines, three different behaviours on the same sentences:

Apple normalises the numeral before speaking it and is therefore form-agnostic — identical 80% on 2026 and ٢٠٢٦. It is the existence proof that this is solvable.

fish handles Western digits and collapses on Arabic-Indic — 73% to 7% on sentences that differ in nothing but how the number is typed. The failures are not mispronunciations but noise: ٢٠٢٦ in a date came back as تخاناتر, ٢:٤٥ as انفافس اسم مفاعس.

ArTST speaks neither digit form — 0/30. And because it is open weights, the failure can be explained rather than just observed: the Arabic-Indic digits ٠-٩ are absent from its 87-token vocabulary, so ٢٠٢٦ collapses to a single <unk> and is deleted before synthesis starts. Western digits are in the vocabulary, tokenise cleanly, and still produce ~0.19 s of audio where the same model spends 1.41 s saying the number spelled out. Two different mechanisms, one outcome.

Full analysis, including a Whisper-free duration control: FINDING_artst_numerals.md.

Practical rule this yields: normalise every numeral before it reaches an Arabic TTS engine — to Western digits for fish, and all the way to Arabic words for ArTST.

What this benchmark does NOT yet establish

Stated plainly, because the absolute percentages above are weaker than the comparison:

• The measurement chain is TTS → Whisper → a number parser, and an error anywhere in it is charged to the TTS. Three rounds of parser fixes moved the Western scores from 47% to 73–80% without a single audio file changing, which shows how much of the early "failure" was the harness. arnum.py now handles split hundreds, ordinals (Arabic says times and dates ordinally), spelling variants and summed times — the ordinal gap in particular was silently penalising Apple for speaking correctly.
• The relative gap is robust — 7% vs 80% cannot be produced by transcription noise, and the Arabic-Indic transcripts are visibly gibberish rather than near-misses.
• Three engines have been run (fish, Apple, ArTST). The set is engine-agnostic (--engine) and still needs ElevenLabs, Azure and Google before any claim about Arabic TTS as a field is made.
• ArTST is being run off-domain. It is fine-tuned on ClArTTS (Classical Arabic) and this set is Modern Standard Arabic news copy, which is why its spelled-form baseline is only 27%. Its numeral-specific effect is the drop from 27% to 0%, not the whole of the 0% — see the finding for why that distinction matters.
• No human listening pass yet. A 45-item set should be human-verified before the absolute numbers are quoted anywhere.
• The scorer gaps listed here on 2026-08-20 have since been fixed and the affected numbers revised — see the Changelog. What replaces them is a narrower gap: the clock rules are gated on the transcript literally containing ساعة, so when an engine's output is garbled enough that Whisper writes الساعدة, a correct half-past reading is still scored LOST. That is deliberate — نصف means "half" in any context and reading it as 30 unconditionally would be a worse bug than the one it fixes.

Files

build_set.py	generates data/sentences.jsonl
run_bench.py	synthesise → transcribe → score
arnum.py	Arabic number-word → value normaliser
test_arnum.py	34 tests for the normaliser
rescore.py	re-derives every result from the stored transcripts, no re-synthesis
artst_duration_control.py	Whisper-free check that ArTST emits no audio for digits
results_{fish,apple,artst}.jsonl	per-utterance results, one file per engine
data/artst_tokens.json	what ArTST's tokenizer actually received per utterance

arnum.py exists because the first version of this benchmark scored only digit strings, and therefore marked correct spoken Arabic (ستمائة وأربعة وسبعون) as a failure. That bug understated every form. It is called out here because a benchmark whose scorer is wrong is worse than no benchmark.

Changelog

2026-08-20 — scorer correction, spelled form only

Four gaps in arnum.py were charging correct speech as a failure. They were published as known limitations before they were fixed; they are now fixed, and every result has been re-derived from the transcripts already stored in the results files — no audio was re-synthesised, so nothing moved underneath the change. rescore.py reproduces this, and refuses to write if a digit-form score moves.

What was wrong:

1. Fractions of an hour were absent. النصف and الربع were in no dictionary, so 6:30 spoken correctly as الساعة السادسة والنصف parsed as {6} and scored LOST. Now read as 30 and 15 — only when the sentence contains ساعة, because نصف is "half" in every other context.
2. Tanween left a trailing alef. The diacritic was stripped but not the alef under it, so سبعونًا missed the tens table and 674 parsed as 604. Lookup now retries without one trailing ا.
3. Ordinals written as two words could not match. الثانية عشرة (twelfth) can only live in a dictionary pre-joined, so the pair is now joined before lookup. This also fixes الحادية عشرة, which used to read as 11 + 10 = 21.
4. The group never closed after a clock hour. الثانية عشرة وخمسة وأربعين accumulated 12+5+40 into a single 57. The hour now closes the group, so the hour and the minutes are separate values. Gated the same way, which is what keeps the date ordinal الثامن والعشرين equal to 28 rather than {8, 20}.

Old → new:

cell	before	after
fish spelled	11/15 (73%)	13/15 (87%)
fish overall	23/45 (51%)	25/45 (56%)
Apple spelled	8/15 (53%)	9/15 (60%)
Apple overall	32/45 (71%)	33/45 (73%)
ArTST spelled	4/15 (27%)	4/15 (27%) — unchanged
ArTST whisper-medium control, spelled	5/15 (33%)	6/15 (40%)
all six digit-form cells	—	unchanged, as required

Three rows flipped in total, every one of them spelled, every one of them LOST → OK: time_0630 for fish and Apple, and big_674 for fish. No row went the other way, and the digit forms are byte-identical — which is the guarantee that matters, because the headline finding (7% vs 80% on Arabic-Indic digits) lives entirely in those forms and is untouched. test_arnum.py grew from 19 tests to 34; the original 19 still pass.

Neither the ranking nor any conclusion in this card changes. fish is still stronger on the spelled form than Apple, Apple is still the only form-agnostic engine, ArTST still scores 0/30 on digits.

Reproduce

python build_set.py
FISH_API_KEY=... python run_bench.py --engine fish
python run_bench.py --engine apple

# ArTST needs torch + transformers to synthesise and faster_whisper to score; if those
# live in different environments, split the run:
python run_bench.py --engine artst --synth-only
python run_bench.py --engine artst

python test_arnum.py

# re-derive every score from the stored transcripts, without re-synthesising
python rescore.py