Arabic voice notes: what speech AI gets right and wrong in Kuwait

In Kuwait a great deal of business arrives as a voice note. A supplier confirms a delivery, a customer explains a complaint, a colleague dictates an order while driving. Speech AI can turn those into text well enough to be useful — and it fails in specific, predictable ways that decide whether the project is worth doing.

The short answer

Transcription of clear Arabic speech is good. Transcription of Kuwaiti dialect mixed with English, over car noise, with names and numbers in it, is not good enough to act on unattended. Build for a human to confirm the parts that matter — the amount, the date, the name — and automate everything around them.

Why voice notes are a business input here, not a novelty

Typing Arabic on a phone is slower than speaking it, and the keyboard pushes people towards Modern Standard forms they would not use out loud. So the voice note is not laziness; it is the fastest honest way for a Gulf customer to tell you something complicated.

Which means a channel most companies treat as unstructured noise carries a large share of their actual instructions. That is a reason to process it, and a reason to be careful with it.

What the technology genuinely does well

Three things have stopped being research problems.

Clear speech into readable text. One speaker, close to the microphone, ordinary vocabulary: the output is usually clean enough to read instead of listening.

Getting the gist. Even a flawed transcript usually supports a correct summary and a correct routing decision. “This is a complaint about a late delivery” survives a transcript with several wrong words in it.

Search. Turning a year of voice notes into text you can search is valuable even when the text is imperfect, because you are looking for the note, not quoting it.

Where it still goes wrong

The failures cluster, and they cluster exactly where money lives.

• Dialect. Most systems are trained overwhelmingly on Modern Standard Arabic. Kuwaiti speech differs in vocabulary and in verb forms, and the transcript drifts towards a standard word the speaker did not say. This is the same gap that makes assistants hand over to a human.
• Code-switching. Half a sentence in Arabic, a product name in English, then back. Systems asked to pick one language per file lose the other half.
• Numbers. Prices, quantities and phone numbers are where errors cost real money, and they are also where context is thinnest — a wrong digit is still a plausible sentence.
• Names. Personal and company names are exactly the words the model has least reason to expect.
• Noise. A note recorded in a car or a warehouse degrades faster than people assume.

Do not accept a published accuracy figure

Vendors quote word error rates from academic evaluation sets. OpenAI’s own speech-to-text documentation gives no per-language accuracy figures at all and says plainly that accuracy varies by language, pointing instead at techniques for improving recognition of uncommon words. The Whisper paper does publish per-language results, but on read-aloud benchmark audio — not a voice note from a car park in Shuwaikh.

So the number you need is the one measured on your own recordings. Take fifty real voice notes, transcribe them by hand once, and compare. That is the same test-set discipline we would apply to any AI claim, and here it is unavoidable because no published figure describes your audio.

The design that works: transcribe freely, confirm narrowly

The pattern that survives contact with real voice notes is to let the machine do the bulk and make a person confirm only the load-bearing values.

Part of the note	Who handles it	Why
Full transcript	Machine, unreviewed	Used for search and context, not for action
Category and urgency	Machine	Survives imperfect words; errors are reversible
Summary for the team	Machine	Gist is robust; a human reads it anyway
Amounts, dates, quantities	Machine extracts, human confirms	A single wrong digit is expensive and plausible
Customer and product names	Machine proposes, matched against your own list	Matching beats transcribing for known values

That last row is the one that changes results most. You almost never need the system to spell a customer’s name correctly — you need it to pick the right row out of a list you already have. Matching a fuzzy transcript against your own customer table is a far easier problem than transcription, and it is the trick most projects miss.

Five things that move accuracy more than changing the model

Teams reach for a different engine first. These five are cheaper and usually matter more.

Give it your vocabulary. Most transcription services accept a prompt or hint list. Feeding in your product names, branch names and the twenty terms your business actually uses removes a large share of errors at no extra cost.

Ask for the language you expect, and allow the other one. Forcing a single language is what silently deletes the English half of a mixed sentence.

Fix the recording end. Asking staff to record away from the road, in one take, does more than any model change. It is unglamorous and it works.

Match, do not spell. Every value that exists in one of your own tables — customers, products, branches — should be resolved by matching, with the transcript treated as a rough query.

Show the audio next to the extraction. When a person confirms an amount, let them replay those four seconds. Confirmation takes seconds instead of a minute, which is what makes the human-in-the-loop design affordable at all.

A worked example

A distributor receives around forty order voice notes a day on WhatsApp. Staff listen to each one and type the order. Average handling is four minutes, so roughly two and a half hours a day.

Transcribed and summarised automatically, with product lines matched against the catalogue and quantities shown for confirmation, handling drops to about ninety seconds — the time to read a filled form and correct one field. That is roughly an hour and a half saved per day.

Note what was not automated: nobody removed the human from the order. The saving came from turning typing into checking. Had the design tried to post orders unattended, the first mis-heard quantity would have cost more than the month’s saving, and the team would have gone back to listening.

The privacy question this raises immediately

A voice note is personal data, and often more revealing than the text equivalent — background conversation, other people’s voices, a customer’s tone when complaining. Before any of this is processed, decide where the audio goes, how long you keep it, and whether the transcript replaces the recording. We set out the framing for that in where your company data actually goes.

One practical rule: keep the audio for as long as you would keep the equivalent email, and no longer. Most companies discover they have kept every voice note since the day they opened, simply because nobody decided otherwise.

What we would not promise

We would not promise a transcript accurate enough to quote in a dispute, unattended dialect recognition equal to a native listener, or reliable speaker separation when three people talk over each other. We would also not promise the cost stays flat as volume grows — audio is priced by the minute, which makes usage a recurring line rather than a build cost.

How we scope it

A voice pilot starts with fifty of your own notes, transcribed by hand, so the decision rests on your audio. Send us a typical voice note and we will tell you honestly which half of it is safe to automate.

Frequently asked questions

Will it understand Kuwaiti dialect?

Partly. Expect good results on the shape of the request and weaker results on specific dialect vocabulary. The reliable fix is not a better model but a narrower job: match what was said against lists you already hold.

Can it handle a note that switches between Arabic and English?

Better than it used to, and still the most common cause of a dropped sentence. Test with your own mixed notes rather than trusting a language list.

Do we have to keep the audio?

No, and often you should not. Decide deliberately whether the transcript replaces the recording, then apply the same retention period you use for email.

Is it cheaper than having someone listen?

Usually, but the saving comes from replacing typing with checking, not from removing the person. Price the minutes of audio, not the project, because that line recurs every month.

What about phone calls rather than voice notes?

Technically similar, legally not. Recording a call brings consent obligations that a voice note someone chose to send you does not, so treat it as a separate decision.