A widely used speech model produced a 38 percent word error rate when tested on 50 hours of Gulf Arabic calls from a retail banking contact center. On clean English speech, the same model delivers word error rates in the low single digits. That 38 percent result is not a software defect. It follows from training mainly on Modern Standard Arabic and broadcast-quality audio, then applying the model to the way customers speak when calling a bank in Riyadh.
Arabic NLP has moved forward considerably over the past three years. Multilingual models now handle Arabic better overall, more Arabic BERT variants are available, and researchers have invested seriously in Arabic speech and text processing. Still, enterprise teams applying these models to real MENA contact center work encounter a familiar group of obstacles. This article examines those obstacles, their sources, and the field's actual position in 2025.
Diglossia is structural, not solved by more data
Arabic is harder to process than many other major languages because of diglossia. Modern Standard Arabic, or MSA, serves formal writing and broadcast. In everyday speech, including contact center calls, people use colloquial forms such as Khaleeji, Egyptian, Levantine, and Moroccan Arabic. These are more than small variants of one language. Khaleeji Gulf Arabic and Moroccan Darija may share roughly 60 to 70 percent of their vocabulary with one another and with MSA at the lexical and phonological levels, yet the differences are large enough for a model trained on one to degrade noticeably on another.
For enterprise NLP, that means a vendor's claim of "Arabic support" needs closer examination than a claim of "French support" or "Spanish support." Spanish has dialect variation, but for speech processing the distance between Castilian and Mexican Spanish is far smaller than that between Gulf and Moroccan Arabic. When a provider says its system supports Arabic, the useful follow-up is: which dialect, trained on which audio, and tested against what corpus?
Code-switching within one call
Across MENA contact centers, especially in banking and telecom, agents and customers regularly move between Arabic and English within the same sentence. This is ordinary speech rather than deliberate stylistic mixing. Among educated Gulf professionals, English financial and technical terms have become part of everyday Arabic conversation.
A customer may say: "I want to cancel my subscription, the automatic renewal charged me twice and I want a refund." In practice, parts of that sentence may be Arabic, with English used for "subscription," "automatic renewal," and "refund," because those terms feel natural in context. A speech model or NLP pipeline unable to manage such intra-sentence changes can create broken transcripts, dropping, garbling, or flagging the English sections as errors.
This problem is difficult partly because the switch points cannot be predicted. They vary with the speaker's background, the subject, the register, and the vocabulary of the domain. Handling them calls for labeled examples containing natural code-switching, leading to the next gap.
The shortage of labeled data
Compared with English, Arabic NLP research has long had fewer large, high-quality labeled datasets. Formal written MSA is in a better position, thanks to public Arabic text corpora and multilingual models trained on extensive Arabic web text. Spoken dialectal Arabic in focused domains such as financial services remains a different matter.
A useful Gulf Arabic contact center speech dataset would require natural recordings rather than studio-read text, accurate transcripts that retain dialect pronunciation instead of normalizing it to MSA spelling, and ideally intent or topic labels. Creating it also calls for people who can transcribe Gulf Arabic precisely, understand the relevant domain, and access the recordings. Assembling all three is difficult and costly. Consequently, published Arabic speech datasets tend to favor MSA or Egyptian Arabic, the two dialects supported by the largest academic research communities.
For enterprise teams, the result is predictable: general-purpose models often perform poorly on Gulf Arabic contact center data, and domain-specific fine-tuning is usually needed for production-level accuracy. That process depends on labeled examples from the actual use case, so data collection and annotation must happen before the outputs reach a useful quality level.
Morphology and tokenization
Arabic carries substantial grammatical information inside word forms. English generally expresses relationships through word order, while Arabic often uses affixes attached to roots. One Arabic verb can encode subject, tense, aspect, gender, number, and mood, information that might take several English words to express.
That structure creates a tokenization problem. Word-level methods that work well for English can create very large Arabic vocabularies filled with rare forms. Subword methods such as BPE and WordPiece help in part, but were largely optimized for Latin-script languages and do not consistently divide Arabic morphemes in linguistically useful ways.
In contact center NLP, this matters especially for intent classification and entity extraction. A customer seeking cancellation may use several morphological versions of the same root, all expressing the same intent. Without sufficient Arabic morphological coverage in training, a system will miss some forms. In churn detection, those misses can hide genuine churn signals.
Arabic contact center named entities
Arabic named entity recognition remains behind English, with the gap particularly visible for financial services entities. Identifying an account type, product name, agent name, or branch location in Arabic requires training on Arabic financial and telecom text annotated for those categories. Most Arabic NER systems learn from news, where people, places, and organizations appear in proportions and contexts unlike those found in bank calls.
In practice, we combine general Arabic NER with domain patterns for the entity types most relevant to our use case: account types, transaction types, product names, and explicit complaint or request language. The general model covers names and locations, while the pattern layer covers specialized vocabulary. This is an engineering workaround for an immature tool rather than a final answer, but it performs better than using general NER alone.
The current position
Arabic NLP has advanced in real ways. Multilingual transformer models understand Arabic substantially better than they did three years ago. Arabic-specific pre-trained models now provide usable starting points. Arabic speech recognition has also improved, especially for Gulf dialects, as major ASR providers have added more MENA-focused training data.
We are not implying Arabic NLP is still at its 2019 level. It has progressed. The gap between Arabic and English NLP capability in enterprise contact center applications remains real and material. An English analytics team can often adopt general-purpose tools and reach good results with modest tuning. An Arabic team needs a dialect-aware baseline, planned fine-tuning, and domain components for the entity types relevant to its use case.
Teams assessing Arabic NLP vendors or developing internal capability should ask what data trained the model, which dialects it covers, what the evaluation set contains, and whether testing used audio resembling their own contact center recordings. Evidence of varied dialect coverage, domain-specific evaluation, and realistic audio conditions matters. References only to academic benchmarks, without dialect or domain detail, warrant further investigation.
Arabic NLP for real enterprise settings is a solvable engineering problem. The work begins by treating Arabic as linguistically rich and dialectally varied, rather than reducing "Arabic support" to a checkbox. That perspective guided the construction of intella and remains central to the field's hardest technical questions. Practitioners are still finding that dialect coverage and realistic call audio, more than headline model size, determine whether a system holds up in production.