Saudi dialect datasets for AI: Najdi, Hijazi, Southern and Eastern Arabic
Saudi Arabic is several dialects in the middle of a merger. This guide sets out the features that trip models up, the public datasets and what their licenses allow, and what you still have to collect yourself.
In short: no single Saudi dialect dataset covers the Kingdom's main varieties (Najdi, Hijazi, Southern, Northern and Eastern) with sub-dialect labels and a commercial license. The biggest public resources, SDAIA's roughly 667-hour SADA speech corpus and the 141-million-tweet STMC, are non-commercial; the openly licensed sets are small, synthetic or translated. Use public data for baselines and collect your own Saudi data for production.
Saudi Arabic is not one dialect
Researchers divide Saudi speech into regional groups. The SDC-5 corpus (Al-Shenaifi et al., 2024) uses five: Hijazi, Najdi, Southern, Northern and Eastern. The SauDial localization dataset works with four: Najdi, Hijazi, Janoubi (Southern) and Eastern. Each group hides more variation: Ha'il in the north, for example, speaks a Najdi variety with features of its own.
The map is also moving. A 2022 review by Najla Alghamdi in King Saud University's Journal of Research in Language & Translation reviews studies from Jeddah, Mecca, Abha, Ha'il, Medina and Dammam. In each city, speakers are found dropping marked local features for neutral ones. Saudis call the result اللهجة البيضاء (al-lahja al-bayḍāʾ, "the white dialect"), speech without obvious regional markers. In Alghamdi's sample of 26 Saudi TV advertisements, this koine had a Najdi pole and a Hijazi pole, and the Najdi one was more prominent.
For data teams, a model trained on media and social posts learns the levelled koine well and the regional varieties poorly. In these studies younger speakers (young women in Abha, for example) lead the move away from local features, so older and rural users are the likeliest to keep them. A "Saudi" label tells you little until you know which speech it contains.
Najdi, Hijazi, Southern and Eastern: features that matter to models
The contrasts below come from the studies Alghamdi (2022) reviews, plus a few very common everyday words. Each row is a tendency, not a rule: every feature varies by age, setting and contact with other dialects.
| Feature | Where | Example | Why models care |
|---|---|---|---|
| Interdental sounds kept or turned into stops | Najdi speakers, and Ghamdi speakers from Al-Baha, keep [θ, ð, ðˤ]; urban Hijazi uses [t, d, dˤ] | ثلاثة thalātha vs talāta, "three"; chat spelling can follow speech (تلاتة) | Normalizers must map both spellings to one word; a strong Hijazi cue |
| Affricated feminine "your" | Najdi, also parts of the north, east and south; being levelled in city speech | كتابك "your book (f.)" said kitābits rather than kitābik | Transcribers need a rule: spell the sound or the underlying ك |
| Gender in plural verbs | Najdi keeps it; urban Hijazi uses the masculine form for everyone | ياكلون yāklūn (m.) vs ياكلن yāklin (f.), "they eat" | Reviewers from other regions may "correct" valid Najdi feminine forms |
| Diphthongs | Kept in some southwestern speech; Meccan speech uses long vowels | بيت bait vs bēt, "house"; لون laun vs lōn, "color" | Invisible in text, audible in speech: a speech-only cue |
| Definite article m- | Southern: Tihami Qahtani speech in ʿAsir | m- in place of ال al- | Tokenizers and morphological analyzers built for al- will not segment it |
| Lateral ض | Southern: the same Tihami Qahtani speech | ض said as an emphatic lateral sound, not [ðˤ] | Rare in training audio, so ASR is likely to mishear it |
| ج as a y sound | Eastern: traditional Dawāsir speech in Dammam | ج said as a y-glide instead of j | Being levelled away, so rare in young speakers' audio |
| Raised feminine ending | Northern: Ha'il (a Najdi variety) | ثلاثة said thalātheh | Another sub-regional cue that a "Najdi" label hides |
| Everyday words | Najdi vs Hijazi | "What": وش wish vs إيش ēsh. "Now": الحين al-ḥīn vs دحين daḥīn. Hijazi future حـ ḥa-, as in حيدرس "he will study" | The cheapest text signals for dialect ID, and the first words a generator gets wrong |
Transliterations are simplified. Note that several strong markers, such as diphthongs, the lateral ض and the y-glide, exist only in speech. Standard spelling erases them, so a text-only dataset cannot teach them, and a sample of young urban social-media users will under-represent them.
Saudi dialect datasets: what exists
Details come from each dataset card or paper, as of September 2026. Licenses change; check the card before you download.
| Dataset | What it contains | Saudi coverage | License | Watch out for |
|---|---|---|---|---|
| SADA (SDAIA NCAI + Saudi Broadcasting Authority) | About 667 hours of transcribed audio from 57+ TV shows, with roughly 20 hours held out for validation and testing (Gerazov et al., 2025) | Mostly Saudi dialects; speaker dialect, gender and age metadata | CC BY-NC-SA 4.0 | Non-commercial; broadcast, not phone, speech. Best WER in Gerazov et al. (2025): 40.9% on the clean test set (fine-tuned MMS 1B with a 4-gram language model) |
| STMC (Saudi Tweets Mega Corpus) | 141M+ tweets used to pretrain SaudiBERT | Saudi dialect overall; no sub-dialect labels | CC BY-NC 4.0 (dataset card) | Unlabeled pretraining text; non-commercial |
| SauDial | 1,000 English–MSA–dialect entries for game localization, with context notes | Najdi, Hijazi, Janoubi, Eastern | CC BY 4.0 | GPT-4o output revised by one expert localizer (Alanazi et al., 2025); one domain |
| Sa`7r irony dataset | 19,810 tweets, 8,089 labeled ironic | Saudi dialect; ironic/non-ironic labels only | README says CC BY 4.0; the repository license file is CC0 | Clarify license and X terms before commercial use |
| SDC-5 (Al-Shenaifi et al., 2024) | 790,000 tweets, annotated with dialect seed words and geolocation | Five groups: Hijazi, Najdi, Southern, Northern, Eastern | See the paper's data-availability statement | The only set here labeled across five Saudi regions; text only |
| MADAR | Parallel travel-domain sentences in 25 city dialects plus MSA | Riyadh and Jeddah | Research license; check the terms before any commercial use | Translated sentences: good for Riyadh-vs-Jeddah contrasts, not natural speech |
| MASSIVE ar-SA | Voice-assistant intents and slots | Saudi locale | CC BY 4.0 | Localized from English; a locale tag is not native dialect |
| Saudi-Dialect-ALLaM instruction set (Barmandah, 2025) | 5,466 synthetic instruction–response pairs, split evenly between Hijazi and Najdi | Hijazi, Najdi | Not released (code and a datasheet are) | Method only; data not reusable |
Read by license, the picture is stark: the large corpora are non-commercial, and the open ones are 1,000 synthetic lines, a translated intent set and a tweet set with conflicting license statements. None combines sub-dialect labels, natural speech and a commercial license.
What is missing for production AI
- Sub-regional speech. Gerazov et al. (2025) describe SADA as mainly Najdi, Hijazi and Khaleeji (Gulf) speech, plus some Yemeni, Egyptian and Levantine. Southern and Northern Saudi varieties are not among its main dialects, and the SauDial authors note that research on Southern varieties is limited.
- Conversational and telephony audio. Even the spontaneous, noisy parts of a TV corpus are not a banking call or a voice note, with their channel, overlap and pace.
- Domain vocabulary. Tweets, TV and games do not cover how people describe symptoms or dispute a bank transfer.
- Instruction and preference data. Barmandah (2025) reports that ALLaM-7B-Instruct answered in Saudi dialect for 47.97% of prompts, with 32.63% MSA leakage. LoRA fine-tuning on Saudi instruction data moved those to 84.21% and 6.21%. Targeted data works; that dataset is not public.
- Sub-dialect evaluation sets. Without a test set split by region, a model can look fine on "Saudi" while failing users outside Riyadh and Jeddah; see why dialect coverage sets your model's ceiling.
Annotating Saudi dialects: pitfalls and examples
Do not trust a general model to pre-label dialect. In our LabelBench v0.1 audit, on the fixed held-out manifest, a local language model (Qwen 2.5 3B) reached 22.4% exact-match accuracy across five dialect regions, where chance is 20%. It predicted Egyptian for 70% to 80% of the items from every non-Egyptian region, and its Gulf recall was 4%. Arabic-specialized Fanar 1 9B did better (47.8%, with 75% Gulf recall). LabelBench's Gulf class is broad Gulf Arabic; Saudi sub-dialects are a finer distinction it did not test. More in our explainer on Arabic dialect identification.
The same everyday question in three registers:
- MSA: ماذا تريد الآن؟ mādhā turīd al-ān? "What do you want now?"
- Najdi: وش تبي الحين؟ wish tibī al-ḥīn?
- Hijazi: إيش تبغى دحين؟ ēsh tibgha daḥīn?
An annotator from either region understands both dialect lines. That is the trap: understanding is not the same as judging whether a model's answer sounds native. Common pitfalls in Saudi label schemes:
- "Saudi" or "Gulf" as the only label. Use a hierarchy (Gulf, then Saudi, then Najdi, Hijazi, Southern, Northern or Eastern). Add MSA, mixed, "pan-Saudi/levelled" and "cannot tell" classes.
- Treating locale as dialect. An ar-SA tag or a Riyadh geotag says where a text came from, not how its author speaks.
- No transcription rule for variation. Decide in advance whether talāta is written تلاتة or ثلاثة, and whether kitābits keeps standard ك. Mixed conventions inflate WER and blur dialect cues.
- Reviewers "fixing" valid forms. Najdi feminine plurals and Hijazi future حـ are correct in their own dialects. Put them in the gold set so reviewers stop editing them away.
- Mismatched raters. A Hijazi speaker should judge a Hijazi answer; route tasks on each annotator's region of upbringing.
Saudi–English code-switching
Saudi speakers often mix English into Arabic within one utterance, and data to model it is scarce: Alharbi et al. (2024) name the lack of suitable data as the main obstacle for code-switching ASR. They built a Saudi Arabic–English evaluation set (the Saudilang Code-switch Corpus, announced for release) and found that LLM-generated code-switched text cut WER by 5.5% relative, on average, on two test sets.
For annotation, decide three things up front: whether English is transcribed in Latin or Arabic script, whether language tags are word- or span-level, and how to tag English stems carrying Arabic prefixes or suffixes. Our guide to Arabizi and code-switching annotation rules covers each.
Collecting Saudi dialect data in-Kingdom
For most commercial systems public data will not do. A collection plan should settle these points before the first recording:
- Quotas and metadata. Set quotas per variety, age band and gender from where your users are, not who is easiest to recruit in Riyadh. Record each speaker's region of upbringing and current city: dialect follows biography, not ID cards.
- Channel realism. Record on the channels you will serve (phone audio, voice notes, chat), not only in a studio.
- Consent and residency. Voice recordings and chat logs are usually personal data under Saudi Arabia's PDPL, so settle consent, retention and processing location first; see our PDPL data residency checklist.
- Rights you own. Consent and contracts should give you the commercial rights the large public corpora do not.
- Native validation per variety. Speakers of each variety check its transcripts and labels against a gold set that includes its marked features.
This is the work Bayanat Labs does from Riyadh. We cover 25+ Arabic varieties, including Gulf Arabic with Saudi/Najdi, with vetted native speakers who pass dialect screening before they touch client data, gold-standard calibration with adjudication, and in-region hosting by default. Our Arabic dialect data collection and annotation service starts with a fixed-scope pilot and a quality report, usually within two weeks. The pilot is where to confirm which Saudi varieties you need and how each is staffed. For fully in-Kingdom work, see data annotation in Saudi Arabia.
- "Saudi" is a family, not a label. Its regional varieties differ in sound, grammar and words, and city speech is levelling into a koine.
- The big public corpora are non-commercial. SADA and STMC are for research; the open sets are small, synthetic or translated.
- Many markers exist only in speech. Diphthongs, the lateral ض and the y-glide for ج disappear in standard spelling.
- Do not let a general model pre-label dialect. In LabelBench v0.1 a local model recalled only 4% of Gulf items, predicting Egyptian for most of them.
- Collect for your users. Regional quotas, speaker metadata, real channels, owned rights and same-variety validators.
Frequently asked questions
Is there a free Saudi dialect dataset I can use commercially?
A few, but they are small or indirect. SauDial (CC BY 4.0) has 1,000 lines produced with GPT-4o and revised by one expert. MASSIVE ar-SA (CC BY 4.0) is translated from English. The largest resources, the SADA speech corpus and the STMC tweet corpus, carry non-commercial licenses. For production, collect and license your own data.
What are the main dialects of Saudi Arabia?
Researchers usually group them by region. The SDC-5 corpus uses five groups: Najdi (central), Hijazi (west), Southern, Northern and Eastern. Within each, urban and Bedouin speech differ; urban Hijazi in Jeddah and Mecca, for instance, is not the Bedouin speech of the same region. City speech is also levelling toward a shared Saudi koine.
Is Saudi Arabic the same as Gulf Arabic?
Partly. Eastern Province speech is usually grouped with Gulf Arabic, and Najdi shares many features with it. Urban Hijazi is further away: stops in place of the interdental sounds, long vowels for diphthongs, and its own everyday words such as إيش ("what") and دحين ("now"). So a dataset labeled only "Gulf" can leave Jeddah users under-represented.
Can large language models understand and write Najdi and Hijazi?
It varies by model, so measure it. A 2025 study found that ALLaM-7B-Instruct gave Saudi-dialect answers for only 47.97% of prompts before dialect fine-tuning, with 32.63% drifting into MSA. Targeted instruction data raised the Saudi rate to 84.21%. Test your own model per sub-dialect, with native Saudi raters, before trusting it with users.
How much public Saudi speech data is there?
The main open corpus is SADA, from SDAIA's National Center for AI and the Saudi Broadcasting Authority: about 667 hours of transcribed TV audio, mostly Saudi dialects, licensed CC BY-NC-SA 4.0. For wider options, see our comparison of Arabic speech recognition datasets.