Arabic speech recognition is hard because the Arabic people speak is rarely the Arabic they write. Modern Standard Arabic (MSA) is the language of news, education and writing. Everyday conversation happens in regional dialects that differ from MSA in sounds, vocabulary and grammar, a split linguists call diglossia.
The dialects also have no standard spelling, so two careful transcribers can write the same sentence differently, and many speakers mix in French in North Africa or English in Egypt, Jordan and the UAE. A model trained on MSA or broadcast news loses much of its accuracy on dialect speech. In the MGB-3 challenge, the best system scored 13.2% word error rate (WER) on Al Jazeera broadcasts and 37.5% on Egyptian YouTube videos.
This guide covers what the benchmarks for Arabic dialect speech recognition measured, what to put in a transcription spec, which dialect to plan for in each market and how to test. It is written for teams building Arabic voice agents, call analytics and dictation.
Why Arabic ASR breaks on dialect speech
The dialects are not accents of MSA. The paper that introduced CODA, a spelling convention for dialectal Arabic, describes the dialects as differing from MSA "on all levels of linguistic representation, from phonology and morphology to lexicon and syntax". MSA is the language of education. The dialects are what people speak natively.
Sounds. The letter qaf (ق) shows how far apart the varieties are. In Cairo, "palace" (qaṣr in MSA) is pronounced ʔaṣr, with a glottal stop, and the letter jim (ج) is a hard g, as CODA's Egyptian guidelines note. The CODA* guidelines give qaf as a glottal stop in Damascus, where "road" is pronounced ṭarīʔ, and as g in Riyadh, where "dish" is pronounced ṭabag.
Words and grammar. The same guidelines show that Egyptian Arabic marks the future with the prefix ḥa- where MSA uses sa-, and its everyday word for "not", miš, is purely dialectal. The same paper notes that NLP tools built for MSA perform poorly when applied directly to dialect text.
Dialects differ from each other, too. CODA* covers 28 city dialects from Rabat to Muscat, grouped into the Maghreb, the Nile Basin, the Fertile Crescent and the Arabian Peninsula. Its authors write that the dialects differ significantly in phonology, morphology and lexicon from one another as well as from MSA.
Training data skews formal. The training set used across the MGB challenges, MGB-2, holds 1,200 hours from 19 Al Jazeera programs. Its authors estimate that more than 70% of the speech is MSA.
Code-switching. In Casablanca, a dataset of TV series episodes from eight Arab countries, Algerian and Moroccan speech each had more than 500 code-switched segments, mostly into French, while Egyptian and Jordanian speakers switched into English. In Mixat, 15 hours of podcast speech from the UAE, 36% of sentences switch between Emirati Arabic and English.
On Casablanca's code-switched segments, Whisper large-v3 did not output a single code-switched word: told the audio was English, it translated it, and otherwise it wrote the foreign words in Arabic letters. Our article on code-switching in speech recognition explains why single-language models trip on mixed speech.
What the MGB challenges measured
The Arabic Multi-Genre Broadcast (MGB) challenges gave every team the same fixed training data, so they show cleanly what dialect does to accuracy.
| Challenge | Test audio | Best result | What it shows |
|---|---|---|---|
| MGB-2 (2016) | Al Jazeera TV programs, more than 70% MSA | 14.7% WER | Broadcast speech, mostly MSA, is the easier case |
| MGB-3 (2017) | 16 hours of Egyptian YouTube videos in seven genres | 37.5% average WER, 29.3% multi-reference WER | The same system scored 13.2% on MGB-2 |
| MGB-5 (2019) | 13 hours of Moroccan YouTube videos in seven genres | 59.4% average WER, 37.6% multi-reference WER | Moroccan Arabic is harder still, even for its transcribers |
The MGB-3 baseline makes the point most directly. It was trained on all 1,200 hours of MGB-2 broadcast data and scored 22.6% WER on the MGB-2 development set. On the Egyptian development set, without adaptation, it averaged 63.2%. Teams then had 4.6 hours of Egyptian adaptation data to close the gap.
MGB-3 and MGB-5 had every file transcribed by four people, who were allowed to spell words as they judged correct. Results are reported as average WER across the four references and as multi-reference WER, which accepts a word if any reference spells it that way. The gap between the two, about 8 points in MGB-3 and 22 in MGB-5, shows how much a score depends on whose spelling it is checked against.
Newer models show the same pattern. In the Casablanca paper (EMNLP 2024), Whisper large-v3 without fine-tuning averaged 63.0% WER across eight dialects after the paper's text normalization, from 39.7% on Jordanian to over 83% on Algerian, Moroccan and Mauritanian.
The authors also fine-tuned Whisper on MSA, on MGB-2, on MGB-3 Egyptian and on MGB-5 Moroccan data. On Moroccan speech, the Moroccan model scored 61.6%, against 88.0% for the MSA model and 84.9% for the Egyptian one. The model fine-tuned on 1,200 hours of MGB-2 did worst overall, which the authors attribute to its studio recordings and to topics such as politics and economics, far from daily conversation.
Bulbul, a 2026 preprint, had speakers from 11 countries record both dialect sentences and MSA or Classical Arabic sentences. Its best system averaged 41.7% WER on the dialect recordings and 13.3% on the formal ones.
Three lessons follow for anyone buying data:
- Matching the dialect beats adding hours. A large corpus in the wrong variety or domain can do less than a small one in the right dialect.
- A neighboring dialect is not a substitute. Egyptian data barely helped on Moroccan speech.
- Part of every Arabic WER is spelling. Fix the convention before comparing models or vendors.
Spelling is part of the error rate
MSA has an established spelling. The dialects do not: CODA's authors note there are no Arabic dialect academies and no large body of edited dialect writing to set a standard. The CODA* paper counted 27 spellings of one Egyptian word, mabiʔulhāš ("he does not say it"), in use online, and concludes that, typos aside, no spelling of a dialect word can be considered incorrect.
Diacritics add a second layer. Arabic script marks short vowels and doubled consonants with diacritics that are optional, so transcripts may or may not include them. In MGB-2, whose original transcripts could contain punctuation and diacritics, the best system scored 21.1% WER against them as written and 15.0% once those marks were removed. The official 14.7% also mapped variant spellings of numbers and common words, and normalizing alef, yaa and taa marbuta brought it to 14.1%. The recognizer's output was the same throughout; the scoring rules moved the result by seven points.
Then there is Arabizi, Arabic written in Latin letters. Letters with no Latin equivalent are often written as digits that look like them, such as 3 for ع. It is common in social media, SMS and chat, so it turns up in web text gathered for language models. The Linguistic Data Consortium, for example, transliterated 1,856 Arabizi conversations into Arabic script and normalized the spelling to CODA for its Egyptian SMS and chat corpus.
CODA was designed for exactly this problem: a conventional orthography for building computational models of the dialects, first defined for Egyptian Arabic and extended by CODA* to 28 city dialects. It uses Arabic script only, gives each dialect word one spelling, follows MSA spelling rules where it can and keeps dialect grammar, such as the Egyptian future prefix. Some letters keep their MSA root spelling even where the dialect pronounces them differently, so Egyptian "palace" is written with qaf although it is said with a glottal stop. That makes CODA transcripts consistent and readable, but they don't show every pronunciation.
What to put in an Arabic transcription spec
Paste these lines into your spec or RFP, next to the general rules for fillers, overlap and noise that our guide to verbatim and clean verbatim transcription covers.
- Variety label: tag each segment as MSA, the target dialect or mixed, and record each speaker's country and region.
- Orthography: CODA* for the named dialect, or your own spelling list, with one spelling per word and a lexicon of frequent dialect words (negation, future and progressive markers, demonstratives).
- Script: Arabic script for all Arabic words. No Arabizi in transcripts.
- Diacritics: none, unless the data will also train text-to-speech. If you use them, apply them to every word, not some.
- Code-switching: English and French words either in Latin script with a language tag, or transliterated into Arabic script. Pick one for training and evaluation references.
- Numbers, names and brands: digits or words, plus a spelling list for product, brand and place names.
- Clitics: whether particles such as the negation ma-…-š and the future ḥa- attach to the word or stand alone.
- Scoring normalization: the same alef, yaa and taa marbuta normalization applied to references and model output.
- QA: a share of files transcribed twice to measure disagreement, and two independent references for at least part of the test set.
Which dialect to plan for in each market
Plan from where your users are and how they speak, not from "Arabic". The table maps markets to dialect groups and to what the SpeechData.ai catalog has. It follows the CODA* grouping, which treats Hijaz (Jeddah), Najd (Riyadh) and the Gulf cities of Abu Dhabi, Manama and Doha as separate sub-regions, and splits the Levant into north (Beirut, Damascus, Aleppo) and south (Jerusalem, Amman, Salt). Hours, speakers and prices are list figures from each dataset page.
| Market | Plan for | In the catalog | Watch for |
|---|---|---|---|
| Saudi Arabia | Gulf, Najdi and Hijazi speech, by region | Gulf Arabic: 500 hours, 50 speakers, recorded in Saudi Arabia, $90 per hour | Check the speakers' regions in the metadata against your users |
| UAE, Qatar, Bahrain | Gulf, in each country's local form | Gulf Arabic, recorded in Saudi Arabia | Emirati Arabic has features of its own, and young Emiratis mix in English; test on local speakers first |
| Egypt | Egyptian | Egyptian Arabic: 2,000 hours, 200 speakers, recorded in Egypt, $65 per hour | English mixing; CODA* treats Cairo, Alexandria and Aswan as separate city dialects |
| Lebanon, Syria | North Levantine | Levantine Arabic: 500 hours, 50 speakers, recorded in Lebanon, $65 per hour | Recorded in Lebanon; check the regional mix before serving Syrian users |
| Jordan, Palestine | South Levantine | Levantine Arabic, recorded in Lebanon | Recorded in the north Levant; test on Jordanian or Palestinian speakers before relying on it |
| Morocco | Moroccan Darija | Darija: 1,500 hours, 150 speakers, recorded in Morocco, $65 per hour | French mixing; Arabizi in any text you scrape |
| Algeria, Tunisia, Libya | Each country's own dialect | Not in the catalog; recorded to spec | In Casablanca, a Moroccan-tuned model beat the MSA and Egyptian ones on Algerian speech, but it is still a different dialect |
| Iraq, Sudan, Oman, Yemen | Each country's own dialect | Not in the catalog; recorded to spec | Allow time to recruit native speakers |
| Any market: news, announcements, read text | Modern Standard Arabic | MSA: 2,000 hours, 200 speakers, recorded in Saudi Arabia, $90 per hour | MSA is spoken with a regional accent |
The last row matters for products that handle formal speech or read text aloud. Accent still counts there. In Bulbul, the best system's WER on accented MSA and Classical Arabic was 8.1% for Syrian speakers, while Sudanese-accented speech stayed above 20% for every system tested.
Across the five Arabic datasets the catalog holds 6,500 hours from 650 speakers, at $65 to $90 per hour of audio. Every dataset is spontaneous two-person conversation recorded in-country, with one channel per speaker, time-aligned and human-verified transcripts, and age, gender, region and device metadata for every speaker.
A voice agent for Saudi and Egyptian callers could start with 500 hours each of Gulf and Egyptian Arabic speech data: 500 × $90 plus 500 × $65 comes to $77,500 at list price, before any volume or multi-language discount. Ask for a free sample and check its transcripts against your spelling rules before you buy, so you know how much mapping your pipeline needs.
Dialects outside the catalog, such as Iraqi, Sudanese or Tunisian Arabic, can be recorded to spec by native speakers, scripted or spontaneous, with time-aligned transcripts.
Test every dialect separately
A single "Arabic" WER hides the problems that matter. Build a test set for each dialect you ship from conversations that sound like production: speakers from the target cities, the same devices and channel, the same topics, and nobody who also appears in training. Report MSA separately.
Score every test set with the same spelling convention and normalization as your training references. Give at least part of each test set a second, independent reference and report multi-reference WER next to ordinary WER, as the MGB challenges did. Report character error rate as well: in MGB-5, word-level disagreement between transcribers was about 45% but character-level disagreement only about 17%, so CER is less affected by spelling variation. Include code-switched segments at the rate your users produce them, tagged by language.
Our guide to building an ASR test set covers sizing, speaker splits and slicing results by group.
To compare the Gulf, Egyptian, Levantine, Darija and MSA datasets side by side, with hours, speakers and prices for each, see our page on Arabic speech data.
