Egyptian Arabic datasets for speech and LLMs: what's usable commercially

In short: the main Egyptian Arabic datasets are CALLHOME Egyptian (telephone speech), BOLT SMS/Chat (text and Arabizi), MGB-3 (YouTube speech), MADAR (parallel sentences from Cairo, Alexandria and Aswan), ArzEn-MultiGenre (Egyptian–English parallel text) and Egyptian Wikipedia. Very few can go into a commercial model as they stand. Most are research-only, non-commercial, LDC-licensed or unclear on rights, and the largest free text source is heavily machine-generated. Treat public Egyptian data as material for evaluation and research. For product training, plan to license or collect your own.

This guide lists what exists, what each license allows, why models over-predict Egyptian, and the three gaps most teams hit: customer-service speech, Upper Egyptian speech and organic Arabizi.

Why is Egyptian Arabic data still scarce?

Egyptian Arabic (Masri, مصري) is heard across the Arab world through Egyptian film, television and music. That visibility makes it look well resourced. In NLP, it is not. Three things shrink the usable pool:

• The biggest free corpus is not organic. The Egyptian Arabic Wikipedia edition (arz) looks large, but Alshahrani and colleagues describe it as large but unrepresentative. A follow-up LREC-COLING 2024 study documents the massive automatic creation of its articles by template-based translation from English, with no human involvement. Size is not the same as signal.
• Much of the curated data is old or research-only. The best-known telephone corpus dates from 1997, and the main multi-city parallel corpus is licensed for internal research only (details below).
• "Egyptian" usually means Cairo. Many resources are Cairene or carry no regional label. Alexandrian, Delta and Upper Egyptian speech are thinly represented, even though they differ in pronunciation and vocabulary.

The same pattern shows up across dialects, which is why dialect coverage decides a model's ceiling. Egyptian is the case where the gap between reputation and usable data is easiest to miss.

Egyptian Arabic datasets for speech and LLMs

The table lists widely used Egyptian resources and their license terms as published on each source page, as of September 2026. This is a starting point for your own legal review, not legal advice.

Two things stand out. Every speech resource is either small, non-commercial, LDC-gated or unclear on rights. And the only large permissive text source is the one researchers flag as unrepresentative. Recent Egyptian LLM work builds its own data. The Nile-Chat paper (Shang et al., 2025) adapts models to Egyptian in both Arabic and Latin script, and its authors state that all their resources are publicly available.

Can you use Egyptian Arabic datasets commercially?

Read the license before the dataset card's marketing line. Each of these patterns has a different answer:

1. LDC corpora. LDC's licensing policy is explicit: non-members, government members and not-for-profit members cannot use LDC data to develop or test commercial products. For-profit members can, for data received while they were members. A startup that licenses CALLHOME as a non-member cannot train its product on it.
2. NC and ND licenses. NonCommercial rules out product training. NoDerivatives adds a further problem: you cannot redistribute a modified version, including a re-segmented or corrected one.
3. "Internal research and evaluation only." MADAR's terms fit a benchmark, not a training set. Commercial rights require a separate agreement with the licensor.
4. Permissive license, third-party content. A CC BY 4.0 label on ArzEn-MultiGenre covers the release. It does not by itself settle rights in the novels, subtitles and song lyrics the sentences come from. Get advice before you train on it.
5. Scraped web audio or text with no stated license. Treat "not stated" as "not granted" until the provider confirms otherwise.
6. Personal data inside the corpus. Speech and chat identify people. If you collect Egyptian speech yourself, Egypt's personal data protection law and your own consent wording apply; get local legal advice. Our PDPL and data residency checklist covers the regional transfer rules.

Why do models label every Arabic dialect as Egyptian?

Because Egyptian is the variety they have seen most, and a zero-shot model falls back on it when unsure. LabelBench v0.1, our reproducible audit, measured this on a balanced 500-item, five-region dialect manifest. A local language model (Qwen 2.5 3B, zero-shot) reached 22.4% exact-match accuracy on the fixed held-out manifest, where chance is 20%. The errors were not random:

The 85% recall on Egyptian looks strong until you add up the column. From the published row percentages, 380 of the 500 predictions were "Egyptian", and only 85 of those were correct. So on this manifest, an item the model called Egyptian was genuinely Egyptian only about 22% of the time. Levantine is the second most common prediction for every region.

This matters in two ways for Egyptian data work:

• LLM-filtered "Egyptian" corpora are contaminated. If you mine a web crawl for Egyptian text with a zero-shot classifier like the one evaluated, expect much of what it keeps to be Gulf, Levantine or Maghrebi. The real share depends on your crawl's mix, so measure it.
• Per-class recall flatters. Report the full confusion matrix, not Egyptian accuracy alone.

Specialization helps but does not close the gap. Arabic-specialized Fanar 1 9B reached 47.8% on the same items, still well below a small supervised character model at 69.8%, and no evaluated generative run reached non-zero Safe Automation Coverage at a 95% quality floor under the v0.1 tied-confidence protocol. That is a result for the evaluated local models on one manifest, not a verdict on every LLM.

What Egyptian Arabic data is missing?

Customer-service speech

None of the public speech sets above is agent-and-customer call audio. CALLHOME consists of unscripted calls placed from North America to overseas numbers, MGB-3 is YouTube content, and ASR-EgArbCSC is 5.5 hours of conversation. Contact-center speech has its own vocabulary, English code-switching and interruptions: الإنترنت مش شغال (il-internet mish shaghal, "the internet isn't working") or عايز أغير الباقة (ʿāyiz aghayyar il-bāʾa, "I want to change my bundle"). If you deploy on phone lines, you need recordings from phone lines. Our Arabic speech data collection and transcription service covers transcription, timestamping and diarization for that kind of audio.

Upper Egyptian (Sa'idi) speech

Cairene pronounces ق as a glottal stop, so قال ("he said") sounds like ʾāl. Across much of Upper Egypt it is a hard g: gāl. An ASR model trained on Cairene audio will meet sounds and words it rarely saw. Among the public research corpora above, MADAR is one of the few to separate Aswan from Cairo and Alexandria, and its license is research-only. If your users are in Asyut, Sohag or Aswan, sample speakers from there and record their region as metadata.

Organic Arabizi

Egyptians write a lot of Latin-script Arabic: 3ayez for عايز ("I want"), 7aga for حاجة ("thing"), msh for مش ("not"). The main organic corpus, BOLT, was collected for a DARPA program, released in 2017 and is LDC-gated. The large open one, Arabizi-Egypt, was produced by a model transliterating Arabic-script text, and it is non-commercial. Machine-transliterated Arabizi reproduces one model's spelling habits, not the variation real users type. That is the synthetic-data trap in a new script.

How to annotate Egyptian Arabic data

If you build or extend an Egyptian dataset, these rules prevent the most common rework. They extend the general Arabic text annotation playbook:

1. Tag the sub-variety, not just "Egyptian". Record speaker or author region (Cairene, Alexandrian, Delta, Upper Egyptian) as metadata. You cannot fix the Cairo skew later if you never logged it.
2. Fix a spelling policy. The same word appears as عايز or عاوز, and إيه or ايه. Keep the raw form, and add a normalized layer if you need one. LDC normalized BOLT's transliterations to the CODA convention.
3. Keep Arabizi as Arabizi. Store the Latin-script original, put any transliteration in a separate field, and never strip digits that stand for letters.
4. Mark code-switching. English and French terms in support calls are part of the data, not errors to "correct" into Arabic.
5. Use native annotators for polarity. Egyptian negation wraps around the verb: مفهمتش (ma-fhimt-ish, "I didn't understand"). Non-native raters and MSA-trained models miss it, and sentiment labels flip.
6. Transcribe speech to a written standard. Decide how to write dialect words, hesitations and English insertions before the first file. Our Arabic transcription guidelines give a template.
7. Evaluate dialect ID with a confusion matrix. Build a stratified gold set with adjudicated labels and check what the "Egyptian" column absorbs.

This is the work Bayanat Labs does. Egyptian is one of the dialect families we cover, alongside Gulf (including Saudi/Najdi, Emirati and Kuwaiti), Levantine, Iraqi and Maghrebi, within 25+ Arabic varieties. Contributors are vetted native speakers who pass dialect screening before they touch client data. Their work is calibrated against gold standards, with adjudication, rework loops and an audit trail on every task. Our Arabic dialect data collection and annotation service starts with a fixed-scope pilot with gold-standard QA and a quality report, usually within two weeks, before you scale.