SHAMS: An Audio-Grounded Pronunciation Benchmark for Levantine Arabic

Version 1.0. Project page: https://shams-nlp.github.io

Summary

SHAMS (SHami Annotated Multi-dialect Speech) is an open benchmark for spoken Levantine Arabic. It has 1,300 utterances from open audio corpora. The utterances are balanced across five varieties by word count: Palestinian Urban, Palestinian Rural, Jordanian, Lebanese and Syrian. Each utterance has four aligned tiers. SHAMS is made only for evaluation.

Paper: arXiv:2610.01427. Evaluation code and model outputs: https://github.com/bensapirstein/shams

Tasks and tiers

Each utterance has four aligned tiers:

1. the audio segment;
2. unvocalized Arabic orthography, corrected to match the recording;
3. fully diacritized Arabic, with diacritics that match the pronunciation in the audio;
4. a phonetic transcription in a Semitic transcription system.

These tiers support four tasks:

• Diacritization: restore vowel and gemination marks.
• G2P: map orthography to phonetic transcription.
• ASR: transcribe audio to text.
• A2P: transcribe audio to phonetic transcription.

The paper reports Diacritic Error Rate (DER) for diacritization. It reports Character Error Rate (CER) for the other tasks and Word Error Rate (WER) for ASR.

Composition

1,300 utterances, 15,506 words, 128.2 minutes. Male/female share of words: 52.8/47.2.

Variety	Source	Utt.	Words	M/F %	Min.
Pal./Urban	Casablanca	22	267	49.8/50.2	1.9
Pal./Urban	LangMedia	208	2,908	48.5/51.5	23.1
Pal./Urban	both	230	3,175	48.6/51.4	25.0
Pal./Rural	Casablanca	234	3,153	69.0/31.0	21.8
Jordanian	Casablanca	137	1,451	60.0/40.0	10.5
Jordanian	LangMedia	168	1,722	35.0/65.0	15.0
Jordanian	both	305	3,173	46.5/53.5	25.5
Lebanese	LangMedia	316	3,008	49.9/50.1	29.1
Syrian	LangMedia	215	2,997	50.0/50.0	26.7
Total		1,300	15,506	52.8/47.2	128.2

Sources: the Levantine portions of Casablanca (scripted dialectal television dialogue) and the LangMedia archive (documentary interviews and spontaneous conversation). Casablanca's utterance segmentation and identifiers are kept. LangMedia recordings were segmented into short utterances. Segments with fewer than three words or heavy background noise were excluded. Designated training-split data is excluded. Where original transcripts did not match the audio, they were corrected by hand. Unicode is normalized across all text tiers. No orthographic normalization is applied (hamza and alif variants, tā' marbūṭa and alif maqṣūra stay as transcribed).

Limits stated in the paper: most varieties come from a single source, so source and dialect are confounded. Only Jordanian draws substantially on both sources. Casablanca actors may not be native speakers of the portrayed variety.

Fields

The release has one split, test. Columns:

Field	Notes
audio	{bytes, path} of the wav. Filled for LangMedia rows; empty for Casablanca rows (see "Casablanca rows").
seg_id	stable identifier
transcription	unvocalized orthography. Derived: strip_diacritics(vocalized_transcript), always
vocalized_transcript	gold diacritics
translit	gold phonetic transcription (dataset/README.md labels it "gold DIN 31635"; the paper describes the Semitic transcription system of the Encyclopedia of Arabic Language and Linguistics)
country	derived from subdialect at push time: Lebanese is Lebanon, Syrian is Syria, Jordanian is Jordan, Palestinian/Urban and Palestinian/Rural are Palestine
subdialect	Lebanese / Syrian / Jordanian / Palestinian/Urban / Palestinian/Rural
gender	M / F
duration	seconds, float
source	casablanca / langmedia
cut_start_ms, cut_end_ms	LangMedia only; null for Casablanca, which inherits that corpus's segmentation
title	LangMedia attribution; null for Casablanca
original_file	LangMedia source URL, the citation handle

There is no code_switch field. Latin script is used for code-switched tokens, so they can be read from the text. Build history and provenance are not published.

Annotation and reliability

Diacritics. Five categories: fatḥa, kasra, damma, shadda and tanwīn. Sukūn is not annotated. The guiding principle is to document diversity as is, not to normalize to canonical forms. The diacritized and phonetic tiers record the realization heard in each recording. Full conventions are in Appendix A of the paper.

Procedure. Both pronunciation tiers were produced semi-automatically. For 520 initial utterances, phonetic transcriptions were made by hand and aligned to the Arabic orthography. For the other 780, Levanti (Mor-Lan, 2024) made a first pass over both tiers and annotators corrected it. Annotators worked against the audio. Every utterance was reviewed by an author. 161 utterances (12.4%) carry diacritics identical to Levanti's output after human review. Because Levanti is also a baseline in the paper, it is marked as potentially in-distribution there.

Annotators. Three native speakers of Palestinian Arabic, led by a PhD student specializing in Arabic dialectology, added diacritics. A co-author with a PhD in Semitic linguistics produced the phonetic transcriptions.

Reliability. A second native speaker of Palestinian Arabic independently re-annotated a stratified subset of 150 utterances, 30 per variety, under the same audio-guided protocol. The unit is the diacritic slot, one per Arabic letter (7,290 paired observations). Agreement is 93.2% and Cohen's κ is 0.898 [0.886, 0.909]. Intervals are 95% cluster bootstrap intervals over utterances.

Subset	Utt.	Slots	Agree	κ [95% CI]
Overall	150	7,290	93.2%	0.898 [0.886, 0.909]
Pal./Urban	30	2,237	93.1%	0.898 [0.881, 0.917]
Pal./Rural	30	1,484	95.4%	0.930 [0.913, 0.944]
Jordanian	30	1,008	91.3%	0.867 [0.823, 0.901]
Lebanese	30	1,078	93.2%	0.896 [0.863, 0.927]
Syrian	30	1,483	92.4%	0.886 [0.861, 0.912]
Palestinian	60	3,721	94.0%	0.911 [0.898, 0.924]
Non-Palestinian	90	3,569	92.4%	0.883 [0.865, 0.901]

The second annotator was shown the first annotator's consonantal skeleton. The comparison therefore measures diacritic reliability only. Words with different skeletons are excluded (142 of 1,881 words). Agreement on Palestinian material (κ = 0.911) is slightly higher than elsewhere (0.883); the difference is 0.027 (95% CI [0.006, 0.050]). The paper states: "Inter-annotator agreement is reported for the diacritization tier over a 150-utterance subset (§2.4); the phonetic and orthographic tiers were each produced by a single expert."

Multi-annotator data

The paper says: "The subset and both annotators' labels are released with the benchmark." (§2.4, Protocol). The subset is the 150 re-annotated utterances, 30 per variety. The paper's sentence covers both annotators' labels for these utterances.

The subset and both annotators' labels are in iaa/ of the evaluation code repository (iaa_subset.jsonl, iaa_annotator_b.jsonl).

Casablanca rows

Casablanca rows (source = casablanca, 393 rows: 22 Palestinian Urban, 234 Palestinian Rural, 137 Jordanian) carry identifiers and annotations. They carry no audio. This follows §2.5: "Annotations linked to Casablanca utterances are released with identifiers without redistributing audio."

The annotations on these rows are the three text tiers (transcription, vocalized_transcript, translit) and the row metadata (seg_id, subdialect, country, gender, duration, source). cut_start_ms, cut_end_ms, title and original_file are null.

To get the audio, run reconstruct.py from this repository. It pulls the 393 segments from Casablanca on Hugging Face, checks each one against casablanca_checksums.csv, and merges them with these annotations by identifier. The result is SHAMS with audio for all 1,300 rows, saved as a datasets dataset.

Casablanca is access-gated. Accept its terms at https://huggingface.co/datasets/UBC-NLP/Casablanca, log in, and download the script with its checksums:

pip install datasets huggingface_hub
hf auth login                                     # older huggingface_hub: huggingface-cli login
hf download shams-nlp/shams reconstruct.py casablanca_checksums.csv \
    --repo-type dataset --revision v1.0 --local-dir .
python reconstruct.py --out shams_full --revision v1.0   # add --wav-dir wavs/ for one wav per row

from datasets import load_from_disk
shams = load_from_disk("shams_full")

The Casablanca audio stays under Casablanca's own licence (CC BY-NC-ND).

Licensing

In the words of §2.5:

SHAMS annotations are released under terms consistent with each source corpus's license. Utterances sourced from LangMedia and accompanying annotations are released in compliance with LangMedia's CC BY-SA license. Annotations linked to Casablanca utterances are released with identifiers without redistributing audio; written correspondence with its authors confirmed this format is consistent with their CC BY-NC-ND license. Full license terms are provided with the released dataset.

Citation

@misc{sapirstein2026shams,
  title         = {{SHAMS}: An Audio-Grounded Pronunciation Benchmark for {L}evantine {A}rabic},
  author        = {Sapirstein, Ben and Mattar, Roy and Mor-Lan, Guy and
                   Mohamed, Ahlam and Cerqueglini, Letizia and Alper, Morris},
  year          = {2026},
  eprint        = {2610.01427},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.01427}
}