Why a Gulf Arabic chatbot fails on dialect, Arabizi and voice notes, and how to fix it

A Gulf Arabic chatbot can pass every demo and still fail the first week on WhatsApp. The demo was tested on clean Modern Standard Arabic; the inbox fills with Emirati dialect, Arabic typed in Latin letters and numbers, English verbs conjugated in Arabic, area names spelled four ways, and forty-second voice notes. Each of these breaks a different layer of the bot, and each has a specific fix. This article maps the failure modes UAE retailers actually see, shows what the research says about dialect and speech accuracy, and gives you a calculator and a checklist to size the problem in your own conversations.

Sources: Casablanca: Data and Models for Multidialectal Arabic Speech Recognition (EMNLP 2024) · TechCrunch on WhatsApp voice messages · NADI 2024 findings, ACL Anthology · NYU Abu Dhabi, Gumar corpus

Where the general bilingual playbook stops

Our guide to building an Arabic-English chatbot customers use covers the architecture: intents, grounding, tools, hand-off and measurement. This article goes one level down: what happens between a customer pressing send and the bot deciding what they meant, and why that step fails so often for Gulf customers.

Research has measured this for years. At NADI 2024, the shared task on Arabic dialect identification, the winning system scored 50.57 F1 at telling dialects apart, and the organisers concluded that dialect identification and dialect-to-MSA translation “remain challenging”. DialectalArabicMMLU, a benchmark that adapted 3,000 questions into five dialects including Emirati and Saudi, tested 19 open-weight models and found persistent gaps in dialectal generalisation. If specialist systems struggle on curated test sets, a bot tested only on standard Arabic will struggle more on a real inbox.

Sources: NADI 2024: The Fifth Nuanced Arabic Dialect Identification Shared Task · DialectalArabicMMLU, arXiv 2510.27543

Seven ways real UAE messages break a Gulf Arabic chatbot

The examples below are illustrative, written the way customers commonly type, not quoted from any client. Each fails at a specific layer: intent detection, retrieval (which policy, product or order applies) or the tools behind the reply.

Notice how few of these fixes involve a bigger model. Most are input handling: a few hundred lines of normalisation, a loanword list, an area gazetteer and a test set that contains the messages customers really send. The model matters, but it can’t look up an order number it never received intact.

The model matters, but it can’t look up an order number it never received intact.

Arabizi, code-switching and spelling: normalise before you understand

Arabizi writes Arabic sounds with Latin letters and digits that resemble Arabic letters: 3 for ع, 7 for ح, 2 for the hamza. Researchers note that its use varies by region and age group, and that NLP work has often treated it as a passing phenomenon. In a UAE retail inbox it sits next to English and Arabic script in the same thread, sometimes in the same sentence.

Source: Romanized Arabic Across Dialects: Views, Usage Patterns, and Linguistic Variation, arXiv 2608.02555

The fix is a normalisation layer that runs before intent detection and produces a matching form of the message, while the original is kept for the conversation record and the reply. A sensible order:

• Digits. Map Arabic-Indic and extended Arabic-Indic digits to ASCII; strip spaces and dashes inside phone and order numbers; convert local mobile formats to international ones.
• Letters. Unify alef forms, treat ة and ه at word end as equal for matching, unify ى and ي, remove tatweel and repeated letters.
• Script. Detect Arabizi by the digit-in-word pattern and Gulf function words, then transliterate to Arabic script for matching.
• Vocabulary. Map Gulf words and Arabic-script loanwords to canonical intent terms: ابا to want, اكنسل to cancel, باجر to tomorrow.
• Entities. Match areas, malls, branches and product names against a gazetteer that lists every spelling customers have actually used.

Reference material exists. NYU Abu Dhabi’s Gumar corpus holds more than 112 million words across over 1,200 documents of Gulf Arabic, covering the varieties of all six GCC countries, and the same lab published a morphologically annotated Emirati subset. Your own anonymised conversations, though, remain the best source of the spellings your customers use.

Sources: NYU Abu Dhabi, Gumar corpus · A Morphologically Annotated Corpus of Emirati Arabic, LREC 2018

Voice notes: the input most bots quietly drop

WhatsApp users were sending about 7 billion voice messages a day when Meta last published the figure. Many text-only bots answer a voice note with “sorry, I can only read text”, which customers hear as “we aren’t listening”. Handling them starts with the format. Meta’s documentation specifies that voice messages are Ogg files encoded with the Opus codec, up to 16 MB, and the webhook marks a recorded voice note separately from an attached audio file.

Sources: TechCrunch, 30 March 2022 · Meta, WhatsApp audio messages · Meta, supported media types

Accuracy is the harder part. In the Casablanca benchmark, which includes six hours of Emirati speech drawn from television series on YouTube, Whisper-large-v3 had a 62.31 per cent word error rate on the Emirati test set with no preprocessing and 52.88 per cent after the authors’ text preprocessing. Across all eight dialects its average was 69.49 per cent. Specialised models have since narrowed the gap: on the Open Universal Arabic ASR leaderboard of June 2026, Cohere Transcribe Arabic averaged 25.87 per cent against 36.86 per cent for Whisper v3 Large, and in Cohere’s own dialect tests it scored 24.36 per cent on Gulf speech against 46.14 per cent for Whisper.

Sources: Casablanca, Table 3 (zero-shot results) · Cohere Labs, Cohere Transcribe Arabic release, July 2026. Vendor dialect figures are the vendor’s own evaluation.

Two lessons follow. A leaderboard gives you a shortlist, not a decision. And even a good transcript will contain errors on the words that matter most, names, areas and numbers, so the bot should read key details back before acting.

Choosing speech-to-text for Gulf accents

A locale code such as ar-AE tells you a vendor accepts that setting; it doesn’t prove the model was trained on Emirati speech. Treat every option below as a candidate to test on your own anonymised voice notes.

Sources: OpenAI, speech to text guide · Google Cloud, Speech-to-Text supported languages · Microsoft, Azure Speech language support · Cohere Labs · DialectalArabicMMLU. Documentation as published on 7 October 2026.

Build a dialect evaluation set from real conversations

Average accuracy hides the problem. A bot that looks accurate overall can be right far less often on Arabizi or voice notes, and those customers simply stop writing. The fix is an evaluation set that’s sliced the way your inbox is.

• Sample. Pull three months of conversations and take a stratified sample: English, Modern Standard Arabic, Gulf dialect, other dialects, Arabizi, mixed script and voice notes, in roughly the proportions you receive them.
• Anonymise. Replace names, phone numbers, emails, addresses and order numbers with realistic placeholders in the same script and digit style, so the test still exercises the normalisation layer.
• Label. Have a native Gulf Arabic speaker tag each message with intent, entities and the ideal reply register. For voice notes, write a human transcript.
• Score by slice. Report intent accuracy, entity accuracy and word error rate for each slice, every week, and block a release that improves the average but worsens any slice.

You don’t need thousands of examples to start: a few hundred labelled messages give a useful first read, provided every slice has real examples. Add every misunderstood live conversation to the set once it’s anonymised, and keep the set under the same access controls as the conversations themselves. Within a quarter it describes your customers better than any public benchmark.

Within a quarter it describes your customers better than any public benchmark.

Replying: Modern Standard, light Gulf tone, or the customer’s script

Understanding dialect doesn’t mean replying in it. A bot that imitates Emirati dialect and gets a word wrong sounds like a parody. The safer default is clear, simple Arabic close to Modern Standard, warmed with a few Gulf courtesy phrases customers expect, such as حيّاك الله (welcome) or أبشر (consider it done). Put these in a style guide rather than leaving tone to the model.

Mirror the customer’s script for short confirmations: Arabizi gets a short reply they can read, Arabic script gets Arabic script. Policy wording is the exception. Returns, refunds and delivery terms should come from approved Arabic or English text, because a generated Arabizi paraphrase of a refund policy is where errors creep in. For voice notes, reply in text and read back the key facts: “Order 48213, delivery to Al Barsha tomorrow afternoon, correct?” The customer corrects a misheard detail in one tap instead of discovering it at the door.

Hand-off rules need to be dialect-aware too. Send the conversation to a person when transcription confidence is low, when the bot has asked for clarification twice, when the customer repeats themselves in another script, or when the tone turns to complaint. Pass the original message, the normalised version and the transcript, so nobody asks the customer to start again.

How many conversations is dialect costing you?

Enter your monthly volume, the mix of messages, and the first-message accuracy your bot achieves on each slice of your evaluation set. The calculator shows how many conversations are misread and how many of those come from the dialect and voice gap.

Illustrative only. The defaults are examples, not market data; use accuracy figures from your own evaluation set. The gap is the number of misreads that would disappear if dialect and voice accuracy matched your standard-text accuracy.

Run the defaults and roughly one conversation in five is misread, more than half of them because of the gap. That is the number to bring to a budget meeting and to track weekly as the fixes land. If your volume is high, our breakdown of what an AI chatbot costs in the UAE helps put a price on closing it.

Also read: AI customer service in the UAE

Dialect readiness check

Tick what’s true for your bot today.

• We know what share of our messages are Gulf dialect, Arabizi, mixed script and voice notes
• Our evaluation set contains real, anonymised examples of every one of those slices
• We report accuracy per slice every week, not only an average
• Arabic-Indic and extended Arabic-Indic digits are converted before every order or phone lookup
• Spelling variants and elongated letters are normalised for matching
• Arabizi is detected and transliterated rather than treated as English
• We have a gazetteer of delivery areas and branches with every spelling customers use
• Voice notes are transcribed by a model we tested on our own Gulf audio
• The bot reads back names, areas and numbers taken from voice notes before acting
• A reply style guide sets register and approved Gulf courtesy phrases, reviewed by a native speaker

Seven or more: the bot probably understands most of your customers, and the work is tuning. Four to six: one slice, often voice or Arabizi, is quietly failing. Under four: start with the inbox audit and evaluation set before changing any model.

Questions UAE buyers ask

Why does an Arabic chatbot fail on Gulf dialect?

Demos use clean Modern Standard Arabic, while real inboxes fill with Emirati dialect words, Arabizi, English verbs written in Arabic script and area names spelled several ways. Each breaks a different layer: intent detection, retrieval or the tools behind the reply.

Can a chatbot understand WhatsApp voice notes in Emirati Arabic?

Yes, but test it first. WhatsApp voice messages are Ogg files with the Opus codec, up to 16 MB. In the Casablanca benchmark, Whisper-large-v3 had a 62.31 per cent word error rate on Emirati speech with no preprocessing. So compare speech-to-text options on your own anonymised voice notes and have the bot read key details back.

How many example messages do we need to test a dialect bot?

You don’t need thousands. A few hundred labelled messages give a useful first read, provided every slice has real examples: English, Modern Standard Arabic, Gulf dialect, Arabizi, mixed script and voice notes. Add each misunderstood live conversation once it’s anonymised.

Should an Arabic chatbot reply in Gulf dialect?

The safer default is clear, simple Arabic close to Modern Standard, with a few Gulf courtesy phrases. Policy wording such as returns, refunds and delivery terms should come from approved Arabic or English text, not a generated paraphrase.

When should the bot hand a conversation to a person?

Hand over when transcription confidence is low, when the bot has asked for clarification twice, when the customer repeats themselves in another script, or when the tone turns to complaint. Pass the original message, the normalised version and the transcript.