Gulf (Khaleeji) Arabic Speech Transcription: What Models Get Wrong Without Native Annotators

Why Khaleeji Arabic Speech Is the Hardest Arabic ASR Problem

Arabic automatic speech recognition has made significant progress over the past decade, driven by large-scale MSA broadcast corpora and multilingual model development. That progress largely stops at the dialect boundary. Gulf Arabic — the dialect cluster spoken across Saudi Arabia, UAE, Kuwait, Bahrain, Qatar, and Oman — presents a triple challenge for speech AI that MSA training data cannot address: distinct phonological realisations, a specialised vocabulary, and pervasive English code-switching in professional and digital contexts.

Arabic NLP and speech research consistently documents this gap. Studies applying MSA-trained ASR models to Gulf-dialect recordings report word error rate (WER) increases of 35–55% compared to MSA test sets (Ali et al., Interspeech 2021; Shatnawi et al., ACL Findings 2023). The failure is not evenly distributed across the lexicon — it is most severe on Gulf-specific function words, phonologically transformed common words, and utterance segments containing English code-switching.

For organisations building voice-enabled services for the GCC market — IVR systems, voice banking, government virtual assistants, healthcare voice intake — this performance gap is a blocker. A system with 35–55% higher WER on real user speech cannot be deployed in production without a significant Khaleeji-dialect training data programme, and that programme starts with high-quality native-speaker transcription annotation.

Five Khaleeji Speech Patterns That Break MSA ASR Models

1. Qaf realisation: /q/ → /g/ or /j/

The most phonologically distinctive feature of Gulf Arabic is the realisation of the Classical Arabic phoneme ‘ق’ (qaf, /q/). In Najdi Arabic, qaf is typically realised as /g/ — ‘قال’ (he said) becomes /gal/, ‘قلب’ (heart) becomes /galb/. In Kuwaiti and some Emirati dialects, qaf may be realised as /j/ or /dʒ/. MSA ASR models are trained on formal recitation and broadcast audio where qaf maintains its classical /q/ realisation. When a Gulf speaker says /gal/, the MSA model has no training exposure to this phone in Arabic lexical context and either drops the word or substitutes a phonologically similar but semantically unrelated MSA word.

Accurate Khaleeji transcription requires a transcriber who knows that /gal/ = ‘قال’ and who can apply the correct written form to a sound that MSA models treat as foreign. This is a native-speaker phonological mapping skill, not a language learning achievement — non-native Arabic speakers and MSA-trained annotators without Gulf-dialect exposure make this error systematically.

2. Imala vowel raising

Gulf Arabic dialects apply imala — a process that raises the Classical Arabic long /aː/ vowel toward /eː/ or /iː/ in certain phonological environments. Words like ‘كتاب’ (book, MSA /kitaːb/) become /kiteb/ or /kitiːb/ in some Khaleeji varieties. Verb stems that use a long /aː/ in MSA (/kaːna/, he was) are frequently produced as /keːn/ or /kiːn/ in Gulf conversational speech. MSA ASR models map these vowel-shifted forms to incorrect lexical entries or to out-of-vocabulary tokens, producing transcription errors that cascade through downstream NLP applications.

3. Consonant cluster reduction and elision

Khaleeji Arabic conversational speech — particularly at fast speech rates — elides consonants and reduces clusters in ways that differ from MSA phonotactics. The negative particle ‘ما’ fuses with the following verb at fast rates. Common service-domain words like ‘الحين’ (now) become /alħiːn/ → /alħiːn/ → /lħiːn/ across a speech rate continuum. Phone numbers, dates, and reference codes are often read in mixed Arabic-English digit patterns. Each elision and fusion pattern requires a Khaleeji native transcriber who can reconstruct the canonical written form from a spoken signal that bears little phonological resemblance to its MSA equivalent.

4. English code-switching in Gulf business speech

Professional Gulf Arabic speech mixes English at rates of 15–25% of tokens in business and technical contexts (Al-Khatib & Sabbah, 2008; updated estimates from GCC call-centre corpora 2024). English brand names, product categories, technical terms, and action verbs appear mid-utterance within Arabic grammatical frames. “أبي أـupgrade الـpackage تاعتي إن شاء الله” (“I want to upgrade my package God willing”) contains Arabic morphological inflection on an English verb and an English noun within a Khaleeji politeness frame.

MSA ASR models handle English code-switching poorly for two reasons: their language model assigns very low probability to English tokens in Arabic context, and their acoustic models are trained on monolingual Arabic audio. Whisper-style multilingual models handle code-switching better but still exhibit systematic errors on the Arabic morphological inflection of English verbs — precisely the pattern most common in Gulf business speech.

5. Gulf-specific function words absent from MSA corpora

Khaleeji Arabic uses function words and discourse markers that are absent from MSA ASR training corpora. ‘عشان’ (because/so that), ‘اللي’ (which/that, Gulf form), ‘الحين’ (now), ‘شلون’ (how, Gulf form), and ‘وين’ (where) are extremely high-frequency in Gulf conversational speech but appear in MSA ASR language models at near-zero probability, causing systematic recognition failures on some of the most common words in Khaleeji dialogue.

Sub-Dialect Variation Across the GCC: Why One Khaleeji Pool Is Not Enough

Khaleeji Arabic is a dialect cluster, not a single dialect. Sub-dialect differences in phonological realisation, vocabulary, and code-switching rate are significant enough to cause 15–25% additional WER degradation when an ASR model trained primarily on one sub-dialect is applied to another.

Najdi Arabic (Central Saudi Arabia, Riyadh) is the most phonologically conservative Khaleeji sub-dialect — it retains the /q/ realisation more frequently than other Gulf varieties and has distinct vowel patterns for common verb stems. Hejazi Arabic (Jeddah, Makkah, Madinah) is more influenced by contact with Egyptian and Levantine Arabic through Hajj pilgrimage traffic, has higher rates of MSA influence in formal speech, and a distinct rhythm pattern. Emirati Arabic (UAE) has the highest English code-switching rates in professional contexts, a gaf (/g/) realisation of qaf that is consistent rather than variable, and distinct intonation patterns on questions and statements.

For ASR training data collection and transcription annotation, this means sub-dialect routing is not optional — it is a data quality requirement. Emirati audio transcribed by Najdi annotators produces systematic errors on qaf realisation consistency and code-switching boundary detection. For GCC-wide products, a transcription team with annotators representing at least Najdi, Emirati, and Hejazi varieties is the minimum viable configuration.

The Gulf voice AI market is expanding rapidly alongside the region's digital government initiatives. Saudi Arabia's Ministry of Interior AI voice portal handled over 12 million citizen voice interactions in 2025 (SDAIA Digital Services Report, 2026). The UAE's federal chatbot infrastructure processed 8.4 million Arabic voice sessions across ministries in the same period. Both require ASR models trained on native-speaker Gulf-dialect transcription data to meet production accuracy standards.

Case Study: UAE Federal Voice Assistant — 48% WER to 14% WER

A UAE federal government agency deployed a voice-enabled citizen services assistant to handle enquiries across six ministries, covering residency, health, education, and social services. The initial ASR component used a multilingual Whisper-large-v3 model with no Gulf-dialect fine-tuning, augmented with a generic Arabic language model.

Before: The off-the-shelf model achieved 48.3% WER on live citizen voice sessions recorded at ministry kiosks. Code-switching segments (English brand or service names embedded in Arabic speech) showed 74.6% WER — essentially unusable for downstream intent extraction. Speaker diarisation accuracy on multi-speaker interactions (citizen + agent) stood at 61.2%, insufficient for turn-level intent analysis. The downstream chatbot intent model, receiving noisy transcripts from the high-WER ASR, achieved only 44.7% intent accuracy on live sessions despite 88% accuracy on the scripted test set.

The transcription annotation project collected 340 hours of de-identified citizen voice sessions and produced timestamped, speaker-diarised transcripts with code-switching annotations. The transcription team comprised nine native Khaleeji annotators — four Emirati-native, two Najdi-native, two Hejazi-native, one Kuwaiti-native — with a phonological transcription guideline covering qaf realisation variants, imala, and code-switching orthography conventions. Each hour of audio averaged 5.2 hours of annotation time. Final transcript accuracy (measured against double-blind expert review) was 98.4%.

After fine-tuning on the annotated data: Overall WER improved from 48.3% to 14.1% on held-out citizen audio. Code-switching WER fell from 74.6% to 22.3% — a 52.3 percentage point improvement. Speaker diarisation accuracy reached 89.7%. With clean transcripts feeding the intent model, intent accuracy on live sessions improved from 44.7% to 83.4%. Citizen session completion rate (interactions resolved without live agent intervention) rose from 28.6% to 61.3%.

The annotation project cost AUD $82,000 for audio collection, transcription, QA, and delivery across all 340 hours. The agency estimated AED 4.8 million in annual live-agent cost reduction from the improved session completion rate — a 10-month payback on the annotation investment.

Transcription Annotation Protocol for Khaleeji Speech Projects

Producing ASR training transcripts for Khaleeji Arabic requires a structured protocol that goes beyond standard transcription task design. The key elements are:

Phonological transcription convention guide. The annotation guideline must specify the written Arabic form for each major phonological variant — how to represent /gal/ (→ ‘قال’), how to transcribe imala-affected vowels, and whether to use dialect-specific spelling variants or normalised MSA orthography. The choice between dialect-faithful orthography and MSA normalisation affects downstream language model training and must be made before annotation begins, not discovered mid-project when annotators have applied different conventions.

Code-switching orthography protocol. When English words appear in Arabic speech, the guideline must specify: which script to use for the English token (Latin, Arabic transliteration, or both), how to handle Arabic morphological inflection on English verbs (write the stem in Latin and the suffix in Arabic?), and how to handle brand names and acronyms. Inconsistent code-switching transcription introduces unnecessary language model perplexity that directly increases WER.

Speaker diarisation annotations alongside transcription. For multi-speaker audio (call-centre recordings, IVR interactions, interviews), speaker turn boundaries should be marked at the time of transcription — not as a separate post-processing pass. Transcribers who are listening to the audio are best placed to identify speaker transitions, including when speakers talk over each other or when a single speaker shifts register (formal Arabic → Gulf colloquial) within a turn.

Sub-dialect metadata tagging per segment. For ASR training data intended to cover multiple Khaleeji sub-dialects, each speaker or conversation segment should be tagged with the transcriber's identified sub-dialect. This metadata allows controlled sub-dialect sampling during model training and enables targeted evaluation of per-dialect WER after model deployment.

AI Taggers' Gulf Arabic annotation service covers Khaleeji speech transcription across all five major GCC sub-dialects, with pre-built phonological transcription guidelines, code-switching orthography protocols, and PDPL-compliant audio de-identification procedures developed across multiple Saudi and UAE government voice projects.

PDPL and Compliance in Khaleeji Speech Transcription

Voice recordings of Saudi individuals are personal data under Saudi PDPL. When those recordings can identify a speaker from voice characteristics, they qualify as biometric data under PDPL Article 2 — the most sensitive data category, requiring explicit consent and heightened protection measures. Call-centre recordings, government kiosk sessions, and healthcare voice interactions all potentially contain biometric personal data.

The required compliance steps before sending audio to annotation teams outside KSA are: obtain data subject consent for annotation use under PDPL Article 5; apply voice de-identification where technically feasible (voice anonymisation tools that preserve linguistic content while masking biometric identity); remove explicit identifiers from the audio content (names mentioned in speech, account numbers read aloud, addresses); document the data transfer agreement specifying KSA data residency requirements; and maintain access logs for annotator interactions with the audio data.

UAE voice data under ADGM or DIFC free zone frameworks is subject to the DIFC Data Protection Law or UAE Federal Data Protection Law, both of which include biometric data protections similar to PDPL. For pan-GCC voice AI projects, voice de-identification before annotation is the most robust compliance position across all relevant jurisdictions.

For multilingual speech transcription projects covering Arabic alongside other languages, see our multilingual speech transcription annotation case study. For the full Arabic data pipeline, see our end-to-end Arabic data labelling case study.

Related Reading

• Gulf Khaleeji Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators
• Gulf Khaleeji Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators
• End-to-End Arabic Data Labelling Pipeline Case Study
• Speech Transcription & Multilingual Annotation Service

Frequently Asked Questions