Speech-to-text struggles with Arabic dialects because they differ from Modern Standard Arabic (MSA) in sounds, words and spelling, and speakers mix in English or French. To know how a provider will do, build a small test set per dialect from your real audio and score each dialect separately.
A Riyadh caller, a Cairo voice note and a Casablanca support call are all “Arabic”, and one model can handle the first well and the third badly.
This post explains what makes each major dialect group hard for speech recognition, and how to build a test that shows how a provider will do on your dialects. For the broader buying checklist (price, limits, privacy), see the Arabic speech-to-text buyer’s guide.
Why are Arabic dialects hard for speech recognition?
Four problems stack up:
- Different sounds. The same letter is pronounced differently by region, so the acoustic model must map several sounds to one spelling.
- Different words. Everyday words differ by dialect. A model trained mostly on MSA may “correct” a dialect word into its formal equivalent, which counts as an error if your reference is verbatim.
- No standard spelling. Dialects are written informally, so even two human transcribers disagree. Training data is uneven too: a 2024 study of dialectal coverage notes that existing ASR systems “mainly cover the modern standard Arabic (MSA) variety and few high-resource dialects” (arXiv 2411.05872).
- Code-switching. Speakers mix English or French mid-sentence. The Casablanca dataset, which covers eight dialects, annotates code-switching alongside dialect and gender.
The Open Universal Arabic ASR Leaderboard paper (Interspeech 2025) evaluates open-source models across several multi-dialect test sets for exactly this reason: one Arabic score hides the spread.
How does each Arabic dialect trip up speech recognition?
| Dialect group | Where | Typical ASR trouble spots |
|---|---|---|
| Gulf (Khaleeji) | Saudi Arabia, UAE, Kuwait, Qatar, Bahrain, Oman | English product names, numbers and brand terms mixed in; regional pronunciation of qaf and kaf |
| Egyptian | Egypt | Jeem pronounced as a hard g; qaf often a glottal stop; very common in media, so models usually see more of it |
| Levantine | Syria, Lebanon, Jordan, Palestine | Urban qaf as a glottal stop; French or English mixed in, especially in Lebanon |
| Maghrebi (Darija) | Morocco, Algeria, Tunisia, Libya | Furthest from MSA; short vowels dropped; heavy French code-switching; Amazigh loanwords |
| Iraqi | Iraq | Distinct vocabulary and pronunciation; less public training audio |
These are tendencies, not rules. City and rural speech differ inside each group, and younger speakers code-switch more. The point is that each group fails in its own way, so you score each group on its own.
How do you build an Arabic dialect test set?
1. Sample real audio per dialect
Pull 30 to 50 clips per dialect from the channel you will run in production: phone calls, WhatsApp voice notes, meeting recordings. Phone audio at 8 kHz behaves differently from studio audio. Label each clip by dialect, and tag clips that contain English or French.
2. Agree a spelling convention first
Before anyone writes a reference, decide:
- Verbatim dialect (write what was said, in dialect spelling), or
- Normalised (map to a consistent form).
Write the choice down with examples. If references are verbatim and the model outputs MSA-like words, WER will look terrible even when the meaning is right. Neither choice is wrong. It has to match what your downstream system needs.
3. Write references for 10 clips per dialect
Have native speakers of that dialect transcribe them. For code-switched clips, decide whether English words stay in Latin script, and apply it everywhere.
4. Normalise both sides, then score
Strip diacritics, unify alef forms (إ أ آ to ا) and final yaa (ى to ي), remove punctuation, and convert Eastern Arabic digits to Western ones. Then compute:
| Metric | Why |
|---|---|
| WER per dialect | The standard measure; strict on clitics and prefixes |
| CER per dialect | More forgiving of spelling variation |
| Entity accuracy | Share of numbers, dates and names transcribed exactly |
| Code-switch group WER | Shows whether English or French words survive |
The jiwer package computes WER and CER; Arabic WER explained has a normalisation function you can reuse and a worked example of how much it changes the score.
5. Read the transcripts
Numbers decide whether a call-centre or banking workflow works. Read 20 transcripts per dialect and check every number, date and name by hand.
What do published Arabic WER numbers tell you?
Public benchmarks give a direction, not a promise. BroadNet publishes three WER figures for BroadNet Transcribe Arabic in its model catalogue: 3.8 on FLEURS-ar (clean read speech), 21.1 on MASC-ar (mixed real-world speech) and 23.4 on its own dialect set. That is a sixfold jump for the same model between read speech and dialect speech, which is why you test on your own dialects.
When you compare providers, compare them on the same test set. A leaderboard such as the Open Universal Arabic ASR Leaderboard is useful because every model is scored the same way. A vendor’s own number on its own set is not comparable with another vendor’s.
Which LLM should read dialect transcripts?
Transcripts usually feed an LLM: to summarise a call, route a ticket or answer a question. That model needs to read dialect too. Our guide to choosing an LLM for Arabic covers how to test it, and Qwen3.8 27B is the model BroadNet tested on Gulf and Egyptian dialects.
Where BroadNet Inference fits
BroadNet Transcribe Arabic is built for Arabic including dialects, costs $0.0042 a minute ($0.25 an hour), and BroadNet reports it #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard, and #1 among models whose licences permit serving through an API. For numbers, dates and names in noisy audio, BroadNet offers a fine-tuned version on request. The $5 starting credit covers about 1,190 minutes, enough for the full test above; the speech-to-text cost calculator turns your production volume into a monthly bill. Create a key to start.
BroadNet does not publish separate WER figures for each dialect group. That is exactly why you should run this test: its dialect-set figure is one number, and your customers speak one or two dialects, not all of them.
