Open the language page of any global transcription service and Arabic is there, usually next to a reassuring number: 4, 5 or 6 percent word error rate. Then record a real meeting with one speaker from Riyadh, one from Jeddah and a third who drifts into English every other sentence, and what comes back needs rewriting, not proofreading. Both numbers are true. The problem is what gets measured.
We read the official documentation and pricing pages of twelve tools and four published independent benchmarks, and asked one question of each tool: was it trained on Gulf speech, or did it just put "Saudi Arabia" in a dropdown?
Why the advertised numbers mislead without lying
Most Arabic error rates quoted by global vendors are measured on a test set called FLEURS: Modern Standard Arabic sentences read aloud, cleanly, by Egyptian speakers. That is about as far from a business call in Dammam as audio gets. Test the same models on real Gulf speech and the picture inverts.
The clearest evidence is the Open Universal Arabic ASR Leaderboard from the Saudi company Elm, which scores open models on several test sets including SADA, the Saudi Audio Dataset for Arabic released by SDAIA from real broadcast recordings. The best open model reaches 19 percent error on Modern Standard Arabic and 48 percent on Gulf dialect in the same dataset. Every model does best on MSA and degrades on Egyptian, then Gulf.
A University of British Columbia study tested Whisper, the free model inside dozens of apps, on dialects without fine-tuning: on Emirati speech its error rate passed 100 percent, meaning it produced more wrong words than it heard. And in a test of Arabic contact-centre calls by the Jordanian company Maqsam, the big cloud services landed between 49 and 72 percent.
The conclusion that governs the rest of this guide: the number that matters is not "Arabic error rate" but the error rate on the dialect your team speaks, and almost nobody publishes it.
Group one: models actually trained on Gulf speech
Cohere Transcribe Arabic is the exception that publishes the number. An open-weights model under the Apache 2.0 licence, released in July 2026, whose notes name the dialects: Gulf, Najdi, Hijazi. Its Gulf-only error rate is 24.4 percent, and on SADA it scores 37.5 percent against 56 percent for the largest Whisper model. Not small numbers, but roughly half the error, and the only model in this group you can run for free on your own servers.
Munsit, from CNTXT AI in Abu Dhabi, is the closest competitor. In the company's paper it scored 27.7 percent on SADA, while the same test put Azure, ElevenLabs and OpenAI's transcription model between 40 and 46 percent. Vendor-run, but the only head-to-head on Saudi audio in print. Sold as an API and apps with on-premises deployment; pricing through sales only.
HUMAIN Voice is the Saudi option that answers the compliance question before the accuracy one: processing is documented as staying in the Kingdom, and prices are public in riyals, from a free tier to 649 riyals a month, with batch transcription at about one riyal per audio hour. Its model handles Arabic and English in the same recording. What it lacks is a published error rate.
Hams.AI is also Saudi, puts Saudi and Gulf dialects in the first line of its product page, and processes in Saudi data centres. But it publishes neither a benchmark nor prices, and its accuracy wording is muddled. A candidate to test with your own recording, not a proven choice.
Group two: global vendors that name the Gulf
Speechmatics leads this group for two reasons. One Arabic model it describes as covering the Gulf, Egypt and the Levant without a country selector. And, more important, a bilingual Arabic-English mode built for recordings where the speaker jumps between languages mid-sentence, which describes most Gulf business meetings. Its claimed figures are 4.5 percent on its own test and 6.3 percent on mixed speech. Pricing from about 12 cents an hour, with on-premises deployment.
Deepgram added an Arabic model in January 2026 with separate codes for Saudi Arabia, the UAE, Qatar and Kuwait, and claims up to 40 percent fewer errors than "competing systems" it does not name. The lowest price in this guide, about 0.43 cents a minute with a 200-dollar credit. The developer's choice for a cheap prototype.
AssemblyAI earns a mention for candour: it tiers its own Arabic as "good", defined as 10 to 25 percent error, which is closer to reality than any single-digit figure. Diarization and streaming are mature; pricing 15 to 21 cents an hour.
ElevenLabs Scribe suits anyone already paying for ElevenLabs voices and separates up to 32 speakers. The company puts it in the 10 to 20 percent tier, but CNTXT's dialect test placed it near 40 percent. Good for podcasts in careful Arabic, not for crowded calls.
Group three: the big clouds, and the locale-code trap
Azure lists eighteen Arabic locales, gives the Saudi one specifically a post-stream refinement step and the fullest customisation options, and has a UAE North region, the only Middle East speech region among the hyperscalers. Google lists nineteen Arabic locales in preview, served from US and EU regions only. Amazon Transcribe has just two entries, and here is the trap: its ar-SA code means Modern Standard Arabic, not Saudi dialect; Gulf Arabic lives under the UAE code ar-AE.
The rule: a locale code in a list is not a dialect model. None of the three clouds publishes a per-locale error rate, and independent tests put them between 45 and 70 percent on Arabic calls. Their real value is integration and compliance, not accuracy.
Whisper, and the meeting apps that run on it
Whisper is free and open source, and the engine behind many "we support 90 languages" claims. Its Gulf numbers are above, which is why it pays to know what runs inside your app. tl;dv, for example, lets you select "Arabic (Saudi Arabia)" or "(UAE)", but its help pages say transcription runs on Whisper, so those are labels, not models. Fireflies accepts Arabic as one language and quotes 95 percent for all non-English languages as a blanket figure. Otter and Descript, the most recommended names in English lists, do not accept Arabic at all.
What to choose
- Saudi-dialect calls and data that must stay in the Kingdom: HUMAIN Voice for a quick start with public pricing, or Hams after testing it on your own recording.
- You have an engineer and want the best measured Gulf accuracy: Cohere Transcribe Arabic on your servers, or Munsit if you want a vendor.
- Meetings that mix Arabic and English: Speechmatics in bilingual mode.
- Cheap prototype with Gulf locale codes: Deepgram.
- Already on Azure: enable the Saudi locale from the UAE region and test before relying on it.
- Non-technical meeting notes: tl;dv or Fireflies, expecting to edit.
One last piece of advice that outweighs everything above: record ten minutes of a real meeting of your own team, with its voices, dialects and code-switching, and run every candidate on it before paying. The number you get is the only one that matters.