Arabic Speech Recognition Across Dialects in Call Centres
Why Gulf, Egyptian, Levantine and Maghrebi Arabic challenge speech recognition, what public research shows, and how to test transcription on your own calls.
त्वरित उत्तर
Arabic speech recognition is hard in call centres because spoken Arabic is a family of dialects, Gulf, Egyptian, Levantine and Maghrebi among them, that differ in vocabulary and pronunciation, and telephone audio adds noise and narrow bandwidth. Public research datasets cover several dialects and the call domain. The reliable way to judge any system is to test it on your own recorded calls, per dialect.
If your contact centre serves Arabic-speaking customers, the question "does your speech recognition support Arabic?" is too vague to be useful. Arabic is spoken in many forms. A system tuned on news broadcasts in one variety can behave very differently on an Emirati customer phoning about a payment, or a Moroccan caller mixing Arabic and French. This guide explains why dialects are a technical problem for transcription, what published research says, how to measure a system on your own calls, and what to look at besides a single accuracy number.
Why dialects are hard for speech recognition
Automatic speech recognition (ASR) turns audio into text by learning from many hours of speech paired with transcripts. Arabic raises three difficulties that English mostly does not.
Spoken and written Arabic diverge
Formal writing and broadcast speech use Modern Standard Arabic (MSA). Everyday conversation uses regional dialects. The dialects are not written to any single standard, so the same spoken word can be spelled several ways, and a model has to decide which spelling counts as "correct". That makes training data inconsistent and makes error measurement itself noisy: a system can produce a perfectly reasonable transcript that does not match the reference spelling.
Vocabulary and pronunciation differ by region
Everyday words, question forms, negation, greetings and the way numbers and times are said all vary between the Gulf, Egypt, the Levant and North Africa. Some consonants are pronounced differently across regions. A model that has seen mostly one variety has fewer examples of the others, so its errors concentrate there.
Code-switching is normal
Callers mix Arabic with English or French, especially for brand names, products, amounts and technical terms. In North Africa French is a common mix-in, and in the Gulf English is; the Casablanca dataset described below annotates code-switching for exactly this reason. A transcription system must handle a sentence that changes language halfway.
What public research shows
Three public research efforts are useful reference points for buyers. None of them is a benchmark of any product, and this article does not quote accuracy figures from them, because the abstracts we checked do not give a single comparable number.
MGB-3: Egyptian Arabic in the wild
The Arabic MGB-3 challenge, described by Ahmed Ali, Stephan Vogel and Steve Renals, focused on dialectal Arabic using a multi-genre collection of Egyptian YouTube videos, unlike the previous year's MGB-2, which was based on more than 1,200 hours of broadcast TV news. It included a speech transcription task, using a total of 16 hours of video across seven genres, and a dialect identification task covering four major dialects (Egyptian, Levantine, North African and Gulf) plus Modern Standard Arabic. Thirteen teams submitted ten systems. Its relevance for you: it showed dialectal, informal speech is a distinct recognition problem from broadcast MSA, and it treated telling the dialects apart as a task in its own right.
Casablanca: eight dialects
The Casablanca dataset is described by its authors as a large-scale, community-driven effort to collect and transcribe multidialectal Arabic speech. It covers eight dialects: Algerian, Egyptian, Emirati, Jordanian, Mauritanian, Moroccan, Palestinian and Yemeni, with annotations for transcription, gender, dialect and code-switching. Its relevance: it includes Emirati and Jordanian speech, the authors present it as addressing the lack of datasets for many Arabic dialects, and it records code-switching explicitly, which matches what happens on real calls.
An Arabic call-domain benchmark
A separate paper, "A New Benchmark for Evaluating Automatic Speech Recognition in the Arabic Call Domain", proposes a benchmark built for telephone conversations. Its stated aim is to cover a broad range of Arabic dialects and to emulate real call conditions, and it notes that in telephone calls, audio quality, background noise and conversational speaking styles negatively affect recognition accuracy. Its relevance: it treats call audio as a test condition of its own, so results on cleaner audio should not be assumed to carry over to calls.
The common thread is that no single Arabic exists for recognition purposes, and that call audio needs its own testing. For full results, read the papers themselves, linked in the sources.
Telephone audio adds its own errors
Even a model strong on a given dialect meets a second problem on the phone.
- Narrow bandwidth. Standard telephone audio is narrowband: it carries a limited frequency range, which removes detail that helps tell similar consonants apart.
- Compression and packet loss. Codecs and network loss add artefacts.
- Noise. Callers phone from cars, shops and streets.
- Overlap and short turns. Conversations include interruptions, back-channel sounds and fragments, unlike read speech.
- Two speakers, two channels. Agent and customer have different accents, and separate recording channels make it possible to transcribe each side independently.
Why a model trained on MSA can stumble on Gulf calls
A system trained mostly on formal MSA and broadcast speech has seen the written-style vocabulary and the careful pronunciation of newsreaders. A Gulf customer on the phone uses everyday words, shortened forms and local pronunciation, and may drop into English mid-sentence. Errors are therefore likely to show up as substituted words and mangled names rather than complete failure. This is a reasoned expectation from how the models are trained, not a measured number, which is the reason to test with your own audio and not to assume.
How to test on your own calls
A vendor demo on clean audio tells you little. A one-week test on your real recordings tells you a lot.
- Collect a sample. Take a few hundred call segments, stratified by dialect (Gulf, Levantine, Egyptian, Maghrebi), by channel (agent versus customer) and by call type. Respect your consent, recording and data-protection obligations when you handle recordings.
- Make a reference. Have a native speaker of each dialect transcribe the segments by hand, following a short written style guide, so spelling variants are treated consistently.
- Run each candidate system on the same audio with the same settings.
- Score with word error rate (WER). WER counts substitutions, deletions and insertions against the reference, divided by the number of reference words. Normalise both sides first (diacritics, alef and ya variants, punctuation, digit style) so formatting differences do not count as errors. Report WER per dialect, never only an average.
- Read the errors, not only the score. Sort the mistakes into names, numbers, code-switched words and everyday words, and judge which would change a business outcome.
- Repeat on a second sample a few weeks later to check the result holds up.
What to measure beyond WER
WER treats every word as equal. In a contact centre they are not.
| Measure | यह क्यों मायने रखती है | How to check |
|---|---|---|
| Names | Identity checks and CRM notes depend on them | Count correct names out of names spoken in your sample |
| Numbers, amounts and dates | A wrong digit changes a payment or an appointment | Check every number against the reference |
| Opt-out and consent phrases | A missed "stop calling me" is a compliance risk | List the ways callers refuse in each dialect and test for them |
| Code-switched words | Brand and product names are often English | Tag these in the reference and score them separately |
| Dialect and language detection | The right model must be chosen per call | Compare the detected label with the native-speaker label |
| विलंब | Live assistance and voice agents need speed | Time from speech to text on your network |
DialerBee कैसे काम करता है
The transcription feature lists support for Arabic as Gulf, Levantine, Egyptian and Maghreb varieties among its 11 languages, detects the language automatically, and transcribes agent and customer audio separately. The page itself says accuracy depends on audio quality, language, accent and call environment, and we do not publish an accuracy figure here. Run the test above on your own calls. Dialect also matters for answering-machine detection, covered in our article on अरबी बोली का उत्तर देने वाली मशीन का पता लगाना और पर AI AMD page.
For voice, the अरबी AI वॉइस एजेंट page describes an agent that has run on live calls in Arabic and English, with Arabic handled as dialects. For any other language, test on a demo first. If you design phone menus for Gulf callers, see IVR design for Gulf callers, and for the case for language-aware systems over English-only ones, see multilingual AI versus English-only.
अक्सर पूछे जाने वाले प्रश्न
Why is Arabic speech recognition harder than English?
Spoken Arabic is a set of regional dialects that differ in vocabulary and pronunciation, while formal writing follows Modern Standard Arabic. Dialects have no single spelling standard, and callers often switch to English or French mid-sentence. Telephone audio then adds noise and narrow bandwidth.
Does speech recognition work on Gulf Arabic?
It depends on the system and its training data. Public research such as the Casablanca dataset includes Emirati speech, but results vary by system. Test any candidate on recordings of your own Gulf callers, per dialect, before relying on it.
What is the best way to compare Arabic transcription systems?
Use the same sample of your own call recordings for every system, have native speakers write the reference transcripts, normalise spelling variants, and report word error rate separately for each dialect. Add checks for names, numbers and opt-out phrases.
Is a low word error rate enough?
No. Word error rate counts all words equally, but a wrong digit or a missed opt-out matters more than a wrong filler word. Measure the business-critical items separately.
What is the Arabic call-domain benchmark?
It is a published benchmark for evaluating Arabic speech recognition on telephone conversations, designed to cover dialect variety and emulate real call conditions such as noise and varied audio quality. Its authors note that these call conditions negatively affect recognition accuracy.
Does DialerBee publish an Arabic transcription accuracy figure?
No. The transcription page lists the supported Arabic varieties and notes that accuracy depends on audio quality, language, accent and call environment. We recommend a test on your own calls.
सूत्रों का कहना है
- Speech Recognition Challenge in the Wild: Arabic MGB-3 (Ali, Vogel, Renals, 2017) (arXiv, checked November 7, 2026)
- Casablanca: Data and Models for Multidialectal Arabic Speech Recognition (2024) (arXiv, checked November 7, 2026)
- A New Benchmark for Evaluating Automatic Speech Recognition in the Arabic Call Domain (2024) (arXiv, checked November 7, 2026)
Note: This article summarises the abstracts of the papers above and general practice for evaluating speech recognition. It does not report accuracy for any product. Statements about why models struggle are explanations, not measured results.
संबंधित पठन
- कॉल प्रतिलेखन
- अरबी बोली उत्तर देने वाली मशीन का पता लगाना
- अरबी AI वॉइस एजेंट
- खाड़ी देशों के कॉल करने वालों के लिए IVR डिज़ाइन
- How to measure call transcription accuracy (WER)
- Arabic-English code-switching in voice AI and transcription
- Arabic voice agent testing checklist before you go live
- Arabic TTS for numbers, dates and prices in voice agents
संबंधित लेख: Arabic Sentiment Analysis on Calls: What Works
संबंधित लेख
Arabic TTS for Numbers, Dates and Prices in Voice Agents
पढ़ने में 10 मिनट लगेंगे
तकनीकीWriting Prompts for an Arabic Voice Agent
पढ़ने में 11 मिनट लगेंगे
गाइडArabic Voice Agent Testing Checklist Before You Go Live
पढ़ने में 12 मिनट लगेंगे
कलेक्शंसAre AI Payment Reminder Calls Legal?
पढ़ने में 10 मिनट लगेंगे
DialerBee को काम करते हुए देखने के लिए तैयार हैं?
15 मिनट के लाइव डेमो का अनुरोध करें, या 14 दिन का फ़्री ट्रायल शुरू करें — न स्लाइड, न कोई प्रतिबद्धता। कॉलिंग तब शुरू होती है जब आपका अपना SIP ट्रंक या PBX जुड़ जाए और मंज़ूर हो जाए।
14 दिन का फ़्री ट्रायल · मुफ़्त AI क्रेडिट · कोई कार्ड नहीं · 11 भाषाएँ · BYOC · अनुपालन में सहायक नियंत्रण