ArNum-TTS: does the number survive?

The same TTS engine recovers 11 of 15 numbers typed as 2026 and 1 of 15 typed as ٢٠٢٦. Same sentences; only the digit script differs.

هل ينجو الرقم؟ ٤٥ جملة عربية لقياس قدرة أنظمة تحويل النص إلى كلام على نطق الأرقام بثلاث صيغ: أرقام غربية، وأرقام هندية، وأرقام مكتوبة بالحروف.

The question: when an Arabic text-to-speech engine reads a sentence with a number in it, can a listener recover the number? Arabic writes the same year as 2026 or ٢٠٢٦, and both are correct. TTS demos rarely contain numbers, so this failure goes unnoticed until a date, price or phone number is spoken wrong in production.

Design: 15 sentences × 3 numeral forms = 45 utterances. Each sentence is identical across the three forms except for how the number is written. The forms are western (2026), arabic_indic (٢٠٢٦) and spelled (ألفين وستة وعشرين). Categories: year, percentage, decimal, count, currency, time, date, phone, ordinal and range. Scoring: TTS → Whisper (small) → Arabic number parser (arnum.py, 34 tests) → was the expected value recovered?

Results

Computed from the recovered field of each results file. python rescore.py re-derives the same numbers from the stored transcripts.

numeral form	fish s2.1-pro-free	Apple Majed (ar_001)	ArTST MBZUAI/speecht5_tts_clartts_ar
western 2026	11/15 (73%)	14/15 (93%)	0/15 (0%)
arabic_indic ٢٠٢٦	1/15 (7%)	14/15 (93%)	0/15 (0%)
spelled ألفين وستة وعشرين	13/15 (87%)	12/15 (80%)	5/15 (33%)
overall	25/45 (56%)	40/45 (89%)	5/45 (11%)

• Apple handles both digit forms equally (14/15 each). This shows the problem can be solved.
• fish handles Western digits and collapses on Arabic-Indic: 11/15 vs 1/15. The failures are noise, not mispronunciations. For example, ٢٨ in a date came back as تخاناتر.
• ArTST doesn't speak either digit form. ٠-٩ are missing from its 87-token vocabulary, so all 15 Arabic-Indic numerals are cut to <unk> before synthesis (data/artst_tokens.json). Western digits do reach the model (15/15) but produce almost no audio. In a Whisper-free duration control (data/artst_duration_control.json), digits add +0.19 s of audio for ArTST vs +2.44 s for Apple (data/apple_duration_control.json).
• A larger ASR model doesn't change the digit result. Re-scored with whisper-medium (data/artst_asr_medium.jsonl), ArTST gets western 0/15, arabic_indic 0/15 and spelled 7/15.
• Phone numbers written as digits fail everywhere: 0/6 across the three engines. Only Apple's reading of the spelled form survives, digit by digit. Full analysis: FINDING_artst_numerals.md.

Practical rule: normalise every numeral before it reaches an Arabic TTS engine. For fish, convert to Western digits; for ArTST, convert to Arabic words. arabic-tts-frontend (pip install arabic-tts-frontend) does this. Its release notes report a re-run on this set.

Usage

from datasets import load_dataset

sentences = load_dataset("syamjithnk/arnum-tts", "sentences", split="test")      # 45 rows: id, category, form, text, expect
fish      = load_dataset("syamjithnk/arnum-tts", "results_fish", split="test")   # + heard, recovered
apple     = load_dataset("syamjithnk/arnum-tts", "results_apple", split="test")
artst     = load_dataset("syamjithnk/arnum-tts", "results_artst", split="test")  # + n_unk, numeral_reached_model, seen_by_model

# score by numeral form
import collections
c = collections.Counter((r["form"], r["recovered"]) for r in fish)

To benchmark your own engine, run python build_set.py, then python run_bench.py --engine <name> (see run_bench.py). Run python rescore.py to re-score the stored transcripts without re-synthesising.

file	what
data/sentences.jsonl	the 45 test utterances
results_{fish,apple,artst}.jsonl	per-utterance transcript and verdict, one file per engine
arnum.py, test_arnum.py	Arabic number-word → value parser, 34 tests
run_bench.py, build_set.py, rescore.py	synthesise → transcribe → score, plus re-scoring
artst_duration_control.py, data/*duration_control.json	Whisper-free audio-duration check
data/artst_tokens.json, data/artst_asr_medium.jsonl	what ArTST's tokenizer saw; whisper-medium re-score

No audio is redistributed. ArTST weights are CC BY-NC 4.0.

Limitations

• Small: 45 utterances, 3 engines, one voice per engine. The comparison between numeral forms is robust (7% vs 87% isn't transcription noise); the absolute percentages are less so.
• Automatic scoring: TTS → Whisper → parser. An error anywhere in that chain counts against the TTS engine. There has been no human listening pass yet.
• ArTST is run off-domain. It was trained on Classical Arabic (ClArTTS), and this set is Modern Standard Arabic news copy. Its numeral-specific effect is the drop from 33% to 0%, not the whole 0%.
• Not yet tested: ElevenLabs, Azure and Google. Don't generalise these results to "Arabic TTS" as a field.
• Clock-time rules depend on Whisper: they only fire when the transcript contains ساعة. A garbled transcript can therefore still score a correct half-past reading as lost.

Corrections

Scorer fixes are re-derived from stored transcripts, with no re-synthesis. rescore.py refuses to write if a digit-form score moves without an explicit reason.

• 2026-08-20: four spelled-form parser gaps were fixed (النصف/الربع, the tanween alef, two-word ordinals, clock grouping). fish spelled went from 11 to 13/15, Apple spelled from 8 to 9/15, and ArTST was unchanged.
• 2026-08-21: بالمئة had been read as 100, and spoken decimals weren't parsed. Apple went from 12 to 13/15 on both digit forms.
• 2026-10-09: the card's Apple overall cell still read 33/45 after the 21 Aug fix. The results files give 35/45. Only the card was corrected; no data changed. (FINDING_artst_numerals.md keeps its original figures, with a dated note at the top giving the current ones.)
• 2026-10-27: three parser gaps found in an external audit, re-derived from the stored transcripts with rescore.py. (1) Whisper's hamza spellings ماءة, مأة, مأتين and مأتي were not folded to مائة / مئتين, so correctly spoken numbers scored LOST. (2) Unit words read digit by digit (خمسة خمسة خمسة) were summed to 15 instead of read as 555. (3) In للسادسة the لل prefix hid the ordinal. Apple western and arabic_indic went 13 → 14/15 (the 6:30 item, so the digit-form guard was overridden with a recorded reason), Apple spelled 9 → 12/15, ArTST spelled 4 → 5/15, and the whisper-medium ArTST spelled control 6 → 7/15. fish is unchanged, and so is the digit-script finding. Ten new tests cover the fixes (44 pass).

Part of "Does Arabic survive the pipeline?"

• ArNum-TTS (this dataset): does the number survive speech synthesis?
• ArShape: does Arabic survive text rendering?
• ArPDF: does Arabic survive a PDF round trip?
• Arabic Corpus Integrity Audit: does reshaped Arabic reach public training data? (341 Hub datasets)
• arabic-lint Space: paste Arabic and watch a tokenizer turn corrupted text into [UNK]

Links

• Code: github.com/Syamjith-NK/arnum-tts · fix: arabic-tts-frontend · linter: arabic-lint · also by the author: shotdrift-comfyui
• Write-up: One of these reads ٢٠٢٦ as a year. The other reads it as noise.
• Blog: The Arabic fix everyone recommends is now the bug · Is Arabic training data corrupted?

Citation

@misc{syamjithnk_arnum_2026,
  author       = {Syamjith NK},
  title        = {ArNum-TTS: numeral handling in Arabic speech synthesis},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/syamjithnk/arnum-tts}},
  note         = {Data CC BY 4.0; code MIT}
}

License: data CC BY 4.0, code MIT (see LICENSE). Author: Syamjith NK.