Dialects
Khaleeji Arabic call transcription
The Arabic of the GCC — Saudi, Emirati, Qatari, Kuwaiti, Bahraini, and Omani business calls.
Last updated: April 2026
Khaleeji — the Gulf dialect cluster spoken from Kuwait City to Muscat — is what most CallScribe customers actually need. It is the working language of GCC contact centres, government hotlines, banking dispute lines, and B2B sales calls across Saudi Arabia, the UAE, Qatar, Kuwait, Bahrain, and Oman. It is also the dialect family that English-first ASR models handle worst, because Khaleeji diverges from Modern Standard Arabic in ways that matter acoustically.
What makes Khaleeji acoustically distinct
The most-cited Gulf phonological feature is the realisation of MSA ج (jīm) — pronounced [d͡ʒ] in MSA — as [j] in much of the southern Gulf and as [ɡ] in Najdi Saudi speech. A speaker saying "هذا الجواب" might produce [haːða l-jawaːb] in Manama and [haːða l-ɡawaːb] in Riyadh. Generic Arabic ASR tuned on Levantine or Egyptian audio mistranscribes both. Then there is the affrication of ك (kāf) → [t͡ʃ] in front-vowel environments common to Bahraini, Kuwaiti, and rural Najdi speech: "kīf" (how) often surfaces as "tʃīf". The MSA ق (qāf) is realised as [ɡ] across most Bedouin-rooted Gulf varieties — opposite to the urban Levantine glottal-stop realisation.
Vowel inventory differs too. Khaleeji preserves short /i/ and /u/ in word-final position where Egyptian neutralises them. Pharyngealised consonants (the "emphatic" series ص ض ط ظ) trigger broader vowel-colouring in Khaleeji than in MSA — an acoustic model trained on news broadcasts reads these as different phonemes entirely.
Lexical and morphological features that break generic ASR
Khaleeji uses a different demonstrative system from MSA: هذا (haːða) is often replaced by "ðaː" or "haːda" in Saudi, "haːða" with reduced vowel in Emirati, and the feminine form contracts to "ðiː". The negation particle مو (muː) replaces MSA لا/ليس; the future marker بـ (b-) prefixes verbs ("byirūḥ" for "he will go") in a way that does not exist in MSA. Possessive pronouns are reordered: GCC speakers say "māl-i" or "ḥaqq-i" where MSA would use a bound suffix.
Lexically, Gulf Arabic borrows extensively from Persian (especially Bahraini and coastal Omani), Hindi-Urdu (across all six countries due to large South-Asian populations), Swahili (Omani coast), and English. A Gulf customer-service call routinely contains "checkbook" (تشيك), "appointment" (موعد or "appointment"), "billing", "complaint number", and proper nouns of products and brands in three languages within a 90-second exchange.
Code-switching: the actual hard problem
On a typical Dubai or Riyadh customer-service call, agents and customers code-switch English-Arabic mid-utterance — sometimes mid-word with English roots and Arabic morphology ("بنعمل reschedule لك"). CallScribe handles this via a hybrid pipeline: language-ID at the segment level, then dialect-aware acoustic decoding with a bilingual lexicon. Generic ASR systems either fall back to romanised Arabic (unreadable) or drop the English entirely. We benchmark on real GCC call recordings — not on MSA news where code-switching is rare.
Sentiment cues specific to Khaleeji
Sentiment in Gulf Arabic is communicated through prosody and lexical choice differently from MSA. Polite negative-intent markers ("والله العظيم" as a frustration intensifier, not just a religious oath; "يا أخي" in falling intonation as exasperation) are dialect-specific signals our sentiment model is trained to weight. Generic English-trained sentiment scoring misses these almost entirely, scoring frustrated GCC customers as "neutral" because the surface vocabulary is religious rather than profane.
How CallScribe is tuned for Khaleeji
The base model is Whisper large-v3-turbo, then fine-tuned on a corpus of GCC call recordings covering Saudi, Emirati, Qatari, Kuwaiti, Bahraini, and Omani speakers. We hold out a per-country evaluation set so a regression on Omani audio cannot be hidden by a gain on Saudi audio. Word-error rate on our internal Gulf benchmark is 8-12% for clear telephony audio (8 kHz, single speaker per channel) and rises to 14-18% on noisy multi-party audio with heavy code-switching. See our Model Card for the full per-dialect WER table.
At a glance
- ✓8-12% WER on clear Gulf telephony audio
- ✓Country-aware: Saudi, UAE, Qatar, Kuwait, Bahrain, Oman
- ✓Code-switching with English handled at segment level
- ✓Khaleeji-specific sentiment lexicon
- ✓No per-minute fees — Business plan flat $29/mo
FAQs
Does CallScribe distinguish Saudi Arabic from Emirati Arabic?▾
Yes — within the Khaleeji model, a country signal is included in the acoustic conditioning. You can tag a project as "SA", "AE", "QA", "KW", "BH", or "OM" to bias the lexicon toward country-specific vocabulary (e.g., "ʿaīb" emphasis in Saudi, "yaʿnī" filler frequency in Emirati). The acoustic model itself is unified across the Gulf with country-specific layer adapters.
What about Hijazi and Najdi as separate sub-dialects?▾
Hijazi (western Saudi — Mecca, Medina, Jeddah) and Najdi (central Saudi — Riyadh) differ phonologically: Hijazi keeps [d͡ʒ] for ج while Najdi uses [j] or [d͡z]; Hijazi often treats ق as [ɡ] but with less consistency than Najdi. Our model handles both as Saudi sub-varieties without requiring you to label them upfront.
How does CallScribe handle South-Asian-accented Gulf Arabic?▾
The Gulf has large Indian, Pakistani, Filipino, and Bangladeshi worker populations whose Arabic carries L1 phonotactic interference. Our Gulf training set includes such non-native speech intentionally, because it is a large fraction of real call-centre traffic. WER is a few points higher than for native speakers but the model does not silently fail.
Can I get speaker-attributed transcripts on Khaleeji audio?▾
Yes — diarization runs on the raw audio (pyannote-based) before the language model sees it, so dialect does not affect speaker labelling. Two-channel telephony recordings (one speaker per channel) bypass diarization for near-perfect attribution.
Where do you process Khaleeji audio?▾
EU infrastructure (Hetzner, Helsinki and Falkenstein). For GCC customers requiring in-country residency, an enterprise tier is on the roadmap — talk to us. No audio is sent to third-party cloud ASR APIs; everything runs on our own GPUs.
Does the Free plan work for Khaleeji testing?▾
Yes — the Starter (free) plan gives you 5 minutes per month, which is enough to upload one or two short calls and verify accuracy on your specific audio before committing. The same Khaleeji model serves Free and paid tiers.
5 min/mo free · No credit card · 8-12% WER on Khaleeji