Arabic Speech to Text: What Works for GCC Teams
Arabic speech to text has quietly become good enough to put into real business workflows — voice notes, sales calls, clinic dictation, committee minutes — but only if you scope it around how people in Kuwait and the GCC actually speak. The demos are impressive. The failures are specific, repeatable, and avoidable if you test the right things before you commit.
Arabic speech to text is three problems, not one
Most tools that advertise Arabic support are trained mainly on Modern Standard Arabic — the Arabic of news bulletins and official documents. Almost nobody speaks that in a sales call or on a warehouse floor. The distance between written and spoken Arabic is the single biggest reason a vendor demo that looked flawless collapses on your own recordings.
Before you buy anything, separate the problem into three:
- Dialect. A typical Kuwait office contains Khaleeji, Egyptian, Levantine and Iraqi Arabic in the same meeting. A model tuned for MSA will turn Kuwaiti speech into confident nonsense rather than admitting it did not understand.
- Code-switching. Real sentences mix languages: الـ invoice وصل بس الـ quantity غلط. Many engines lock onto one language per file and quietly drop the other half of the sentence.
- Your vocabulary. Customer names, districts like Shuwaikh and Sabhan, part numbers, drug names, vessel names. These are the words that carry the business meaning, and they are exactly the words a general model has never seen.
A system that handles all three is not harder to build than one that handles none — it is just scoped differently, because the fix is mostly in the layer around the model, not the model itself.
Where Arabic transcription pays off first
Transcription on its own is a cost. Transcription that lands in a system someone already uses is a saving. The workflows that pay back fastest in the GCC:
- WhatsApp voice notes. Suppliers, drivers and customers send 90-second voice notes instead of typing. Converting them to text, extracting the order or complaint, and pushing it into your system removes the worst bottleneck in most Kuwaiti SMEs.
- Sales and support calls. Not surveillance — call summaries, next steps and CRM notes written automatically so the rep stops writing them at 7pm, plus searchable records for disputes.
- Operational records that start as speech. A delay reported by phone, a berthing time passed on verbally, a site instruction given in person. In shipping and trading these spoken moments later become money, which is why they belong in a written log — see how laytime and demurrage are actually calculated in Kuwait for what that evidence trail has to withstand.
- Clinic and professional dictation. A doctor, engineer or lawyer talking for two minutes instead of typing for fifteen, with the output structured into the fields your records system expects.
- Board and committee minutes. The easiest technical win, because formal meetings lean toward MSA, and the hardest political one, because accuracy expectations are near-absolute.
How to measure accuracy honestly
Never accept a vendor accuracy percentage. Run your own test, and make it small enough that you will actually do it:
- Collect two to three hours of your real audio — bad phone lines, background noise, two people talking over each other, not clean studio samples.
- Have a bilingual colleague produce a correct transcript for 20 minutes of it. That is your ground truth.
- Measure word error rate on that sample, then measure something more useful: entity accuracy. Of the numbers, dates, names and quantities in those 20 minutes, how many came through exactly right? A transcript at 85% word accuracy that gets every amount wrong is worthless; one at 80% that nails every figure is production-ready.
- Test the features that fix the gap rather than hoping the base model improves. Speaker separation and custom vocabulary adaptation — feeding in your product names, districts and client list — typically move entity accuracy more than switching vendors does.
Then decide where a human sits. For minutes and medical notes, a 30-second human review step is cheaper than any accuracy improvement you can buy. For voice notes, nobody needs to review anything.
Consent, storage and who can hear the recording
Recorded speech is the most sensitive data most companies hold — it carries tone, names and things people would never write down. Three decisions to make in writing before the first recording: who is told that the call is recorded, how long the audio is kept after the transcript exists, and whether audio leaves the region at all. Transcription is also one of the paths by which company information quietly ends up in consumer AI tools on staff laptops, which is the same control problem covered in setting an AI usage control policy across a Windows estate. Deleting audio after 30 days and keeping only the transcript is usually the right default.
Cost, timeline and who should build it
The transcription itself is cheap — cloud Arabic speech to text typically runs well under KWD 0.100 per minute of audio, so a team processing 500 minutes a month is spending less on the model than on coffee. The cost is in the plumbing: connecting to WhatsApp or your phone system, extracting the right fields, writing into your CRM or ERP, and handling the cases where the audio is unusable.
A realistic first build is three to five weeks for one or two workflows, in the KWD 1,200–3,500 range for a scoped pilot, with running costs of roughly KWD 50–300 a month depending on volume. Anything bigger than that at the start usually means the scope was never narrowed. If you are weighing a freelancer against a studio against your own IT team, who to hire for AI work in Kuwait lays out the trade-offs plainly.
The part that is easy to underestimate is the bilingual product work: Arabic and English in the same interface, right-to-left layout, Arabic names rendered correctly in exports. That is the same discipline behind the Arabic-and-English pharmacy delivery product we built in Kuwait, and it is where most generic transcription deployments feel foreign to the people meant to use them. Our wider approach to AI systems and automation starts from that constraint rather than bolting Arabic on at the end.
Frequently asked questions
Does Arabic speech to text work with Kuwaiti dialect?
Yes, but not out of the box. General models handle Modern Standard Arabic well and Khaleeji dialect noticeably worse. Accuracy on Kuwaiti speech improves most through custom vocabulary (your names, districts, products), good audio capture, and a review step where correctness is critical — not by waiting for a better base model.
Can it handle sentences that mix Arabic and English?
The better engines can, and code-switching should be the first thing you test, because GCC business speech is full of it. Record a normal internal call, run it through any tool you are considering, and check whether the English words survive inside Arabic sentences or get dropped.
How much does an Arabic speech to text workflow cost to run?
The transcription cost is small — usually well under KWD 0.100 per minute of audio. Expect KWD 1,200–3,500 to build a scoped pilot covering one or two workflows, and roughly KWD 50–300 a month in running costs at normal SME volumes.
Should we keep the audio recordings?
Usually no longer than you need. A common default is to keep the transcript permanently and delete the audio after 30 days, with clear disclosure to anyone being recorded and a written decision on whether audio may leave the region.