Salesteq
ENDE

Research · Arabic speech

Arabic speech recognition, judged by the field: 1st place at NADI 2026.

Dealers in the Gulf get their customers on the phone and in WhatsApp voice notes, in Saudi, Gulf and mixed dialects. General-purpose speech engines are trained mostly on Modern Standard Arabic and lose the details that matter: a plate number, a model year, a branch name. We build our own speech models for that, trained on car-sales and service conversations, and test them in public.

Results from nadi.dlnlp.ai/2026, checked October 2026

1st

of all systems in mixed-dialect Arabic speech recognition, NADI 2026 Task 1.2

2nd

in spoken Arabic dialect identification, NADI 2026

44.22

word error rate on the shared mixed-dialect test set, ahead of 44.65 and 46.70

17.4%

word error rate on real BMW service calls in Gulf Arabic

What NADI is

NADI is the Nuanced Arabic Dialect Identification shared task, run each year for research teams and companies working on Arabic dialects. The 2026 edition was organised by researchers from the University of British Columbia, KFUPM, ELYADATA, KAUST, Avignon Université and NYU Abu Dhabi. Every team submits to the same hidden test sets and the organisers publish the leaderboards.

Salesteq placed 1st in Task 1.2, mixed-dialect (multidialectal) speech recognition, ahead of the Arabic specialists, and 2nd in Task 2, spoken dialect identification.

Why a car dealer should care

  • 01Phone calls: the agent understands what a caller says about the car, the problem and the branch, in the dialect they speak, and every call can be transcribed and searched.
  • 02WhatsApp voice notes: many Gulf customers send voice instead of text; the agent reads them like a written message.
  • 03Numbers and names: plate numbers, VINs, model years and branch names are where general engines fail and where a booking goes wrong.
  • 04Dialect identification: the same models tell a Saudi caller from an Emirati or Egyptian one, the basis for a voice agent that sounds local.

Measured on our own work, too

TestResult
NADI 2026, Task 1.2 mixed-dialect ASR1st of all systems, 44.22 word error rate
NADI 2026, Task 2 spoken dialect identification2nd
Real BMW service calls, Gulf Arabic17.4% word error rate, against 21.5% and worse for general-purpose engines

Questions about Arabic speech

What is the best speech recognition for Gulf and Saudi Arabic?+

At NADI 2026, the Arabic dialect shared task, Salesteq's speech models placed 1st of all systems in mixed-dialect Arabic speech recognition (44.22 word error rate) and 2nd in spoken dialect identification. On real BMW service calls in Gulf Arabic they measured 17.4% word error rate, against 21.5% and worse for general-purpose engines.

Who organises NADI?+

Researchers from the University of British Columbia, KFUPM, ELYADATA, KAUST, Avignon Université and NYU Abu Dhabi. Results are published at nadi.dlnlp.ai/2026.

Is the speech model available on its own?+

It runs inside Salesteq's dealer agents, for phone calls, voice notes and call analytics. Talk to us if you need Arabic speech recognition for another use.

Which dialects does it cover?+

Saudi and Gulf dialects are where it is trained and deployed in production; the NADI task covers mixed dialects from across the Arab world.

Hear it on your own calls.

We transcribe a sample of your recorded calls or voice notes and show you the result next to what you use today.

Book a demo

Or write to [email protected]

Related

Sources