On Saudi speech, today's best Arabic ASR model gets roughly one word in three wrong. On Gulf speech, it is closer to one in two.

Those numbers come from the Open Universal Arabic ASR Leaderboard presented at Interspeech 2025. The top model scored 19% word error rate (WER) on Modern Standard Arabic. On Saudi dialects, the same model jumped to between 36% and 48%.

The gap in numbers

Variety

WER, best model

Gap vs MSA

MSA

19.2%

---

Najdi

36.3%

+17 points

Hijazi

37.0%

+18 points

Khaliji (Gulf)

48.2%

+29 points

Test set: SADA. Model: NVIDIA Conformer-CTC with a language model.

Fine-tuning helps less than teams hope. A 2025 study trained models directly on SADA and still reached only 40.9% WER at best.

Saudi Arabia does not speak one dialect

Najdi (Riyadh, Qassim). ق is said as "g", so قال becomes "gāl". Some speakers shift ك toward "ts", so كيف sounds like "tsēf".

Hijazi (Jeddah, Makkah, Madinah). ث often becomes "t", so ثلاثة sounds like "talāta". People say دحين for "now".

Eastern Gulf (Dammam, Al-Ahsa). Many speakers turn ج into "y", so رجّال becomes "rayyāl".

Southern (Jazan, Asir). Some varieties replace the article ال with "am", so البيت becomes امبيت. Almost no public data covers these.

Why models break

Too little real-world audio. SADA, the largest public Saudi corpus, holds about 668 hours from Saudi TV shows. That is useful, but it is broadcast speech, not customers on a phone line.

Phone audio is different. Contact centers often record at 8 kHz. Most training audio is 16 kHz or higher, so call recordings sound unfamiliar to the model.

There is no standard spelling. One transcriber writes وش, another writes ويش. Inconsistent transcripts teach the model noise and inflate error rates.

Callers mix languages. "أبي أسوي reset للباسوورد" is a normal sentence at a Saudi bank. Many models stumble at the switch.

What kind of data fixes it

  • Tag by region. Label every clip Najdi, Hijazi, Eastern or Southern, and balance age and gender.

  • Record where users talk. Phone lines, cars and busy shops, not only quiet rooms.

  • Fix spelling rules first. Agree on how to write dialect words and English before the first file.

  • Score each dialect separately. A single "Arabic" WER hides a 29-point gap.

  • Keep consent and data in-country. Voice is personal data under Saudi PDPL.

The takeaway

If your model works in testing but fails with Saudi callers, start with the data. A small pilot with dialect-tagged phone audio and a separate test set per dialect will show you where the errors come from.

MenaData collects and transcribes dialect-specific Arabic speech and builds evaluation sets for teams shipping voice AI in the Gulf. Request a pilot.