Iraqi Arabic Speech Transcription: What Models Get Wrong Without Native Annotators

Why Iraqi Arabic Is the Hardest Major Arabic Dialect for ASR

Iraqi Mesopotamian Arabic presents ASR challenges that are qualitatively different from the challenges posed by Gulf, Egyptian, or Levantine dialects. Gulf Arabic diverges from MSA primarily in vocabulary and some morphological patterns; Egyptian Arabic shows consistent phonological shifts (qaf realised as hamza, jeem as /g/ in some registers) that are well-documented and have reasonable coverage in Arabic dialect ASR research. Iraqi Arabic diverges from MSA in ways that are more extensive, more phonologically disruptive, and far less represented in the research and commercial data that underpins Arabic ASR development.

Research from the MADAR shared task (Bouamor et al., 2019) and subsequent Interspeech Arabic dialect ASR evaluations documents 45–62% higher word error rate on Iraqi Mesopotamian Arabic speech compared to MSA when using the same acoustic models — the highest WER degradation of any major Arabic dialect in comparative multi-dialect benchmarks. The MGB-3 Arabic multi-genre broadcast challenge (Ali et al., 2017) similarly found Iraqi dialect speech producing word error rates 2.3–3.1 times higher than Levantine and Gulf dialect speech under comparable acoustic conditions.

The ASR training data situation compounds the acoustic model problem. The largest publicly available Arabic dialect speech corpora — the MADAR corpus, the Egyptian Conversational Speech corpus, the MGB-2 and MGB-3 datasets — contain minimal Iraqi Arabic content. A survey of Iraqi Arabic ASR data resources found that fewer than 80 hours of labelled Iraqi Arabic speech exist in publicly available research corpora as of 2025 (Habash et al., 2022), compared with over 1,000 hours for Egyptian Arabic and 400+ hours for Levantine Arabic. For production Iraqi Arabic ASR, custom transcription annotation is not an enhancement — it is a prerequisite.

Five Phonological Failure Modes That Break Iraqi Arabic ASR

1. The Baghdadi qaf-to-gaf phoneme shift

The most pervasive and consequential phonological difference between Baghdad Arabic and every other major Arabic dialect is the systematic realisation of the classical Arabic consonant قاف (qaf, /q/) as گاف (/g/) — a voiced velar stop that is absent from the phoneme inventory of MSA and all other major Arabic dialects. In Baghdad and central Iraq, this substitution is categorical and highly stable: native Baghdadis produce /g/ where all other Arabic dialects produce /q/, /ʔ/ (hamza, as in Egyptian), or the original /q/. Words like قلب (heart → glb in Baghdadi realisation), قريب (near → grib), and قديم (old → gdim) are immediately recognisable to native Iraqis and entirely unexpected in any standard Arabic acoustic model.

The scale of the impact is not limited to one or two high-frequency words. Qaf is a common consonant in Arabic morphology — it appears in verb roots, noun patterns, and function words across the language. Every instance of qaf in everyday Baghdadi speech is realised as gaf. MSA-trained acoustic models, encountering /g/ where they expected /q/, have no trained mapping for this sound and fall back to whatever phoneme in their inventory is acoustically closest — often producing a /k/ or a /dʒ/ transcription that is systematically wrong. A transcription trained by native Baghdadi transcriptionists correctly represents the gaf realisation in the training data, enabling acoustic model fine-tuning that resolves this categorical mismatch.

2. Baghdadi vowel elision and consonant cluster formation

Baghdadi Arabic exhibits systematic short vowel elision — the deletion of unstressed short vowels — that produces consonant clusters unattested in MSA and structurally different from the elision patterns in Egyptian and Gulf Arabic. Words and phrases that in MSA or Egyptian Arabic retain a short vowel between consonants are produced in Baghdadi speech with the vowel deleted and the resulting consonant cluster pronounced as a single phonological unit. The MSA phrase "أريد أن أذهب" (I want to go) undergoes multiple rounds of vowel elision in Baghdadi to produce a collapsed form that sounds almost monosyllabic to an MSA-trained decoder.

This vowel elision pattern interacts badly with MSA-trained acoustic models in two ways. First, the acoustic model expects vowel segments where there are none, causing forced-alignment failures that disrupt word boundary detection and produce cascading transcription errors across the sentence. Second, the resulting consonant clusters are acoustically distinct from any MSA or Egyptian cluster that the acoustic model was trained on, causing the cluster to be parsed as multiple incorrect phoneme sequences rather than the intended consonant pair plus elided vowel.

3. Basrawi emphatic consonant spread

Southern Iraqi Arabic — particularly Basrawi — shows a distinctive pattern of emphatic consonant spread where the coarticulation effect of emphatic consonants (ص، ض، ط، ظ) extends across more syllables than in MSA or northern Iraqi Arabic. In MSA, emphasis typically affects the vowel immediately adjacent to the emphatic consonant. In Basrawi Arabic, emphasis can spread across an entire word and sometimes across a word boundary, changing the quality of vowels that are phonologically distant from the emphatic segment.

For ASR, Basrawi emphatic consonant spread produces vowel qualities that MSA acoustic models have no training coverage for in the non-adjacent positions affected by spread. The model produces transcription errors not on the emphatic consonant itself — which it may identify correctly — but on the surrounding vowels that have been modified by the spread pattern. These errors are difficult to diagnose without knowledge of southern Iraqi Arabic phonology: the transcription output looks plausible to a non-specialist, but native Basrawi speakers immediately hear the mismatch.

4. Kurdish phoneme imports in northern Iraqi speech

Northern Iraqi Arabic — spoken in Mosul, Kirkuk, and the surrounding regions — has been in centuries-long contact with Kurdish and incorporates Kurdish phonemes that do not exist in Arabic phonology. The Kurdish voiced velar fricative /ɣ/ (different from the Arabic غ realisation), Kurdish uvular consonants distinct from Arabic uvulars, and Kurdish nasal patterns appear in northern Iraqi Arabic speech in words borrowed from Kurdish and in code-switched segments of northern Iraqi conversation. Arabic-only acoustic models have no representation for these phonemes and produce systematic transcription errors on any word or segment containing them.

For commercial ASR applications serving northern Iraqi users — customer service for national platforms, government document processing in the Kurdistan Region of Iraq, healthcare transcription for facilities in Mosul and Kirkuk — the Kurdish phoneme import problem is a material accuracy gap that cannot be resolved without training data that accurately represents northern Iraqi phonological patterns. Native Mosuli transcriptionists accurately represent these phonemes in transcription output; non-native Iraqi transcriptionists from Baghdad or southern Iraq treat Kurdish-influenced phonological output as acoustic noise or incorrect pronunciation and normalise it toward Baghdad patterns, producing training data that teaches the ASR model to ignore the phonological features that distinguish northern Iraqi speech.

5. Arabic-Kurdish code-switching in Mosuli and Kirkuki speech

Northern Iraqi callers frequently switch between Iraqi Arabic and Kurdish within a single utterance, producing mixed-language speech where phonological patterns shift between Arabic and Kurdish phonology within a sentence. An Arabic-only ASR model encountering a Kurdish-language segment produces complete transcription failure for that segment — the acoustic output is in Kurdish phonology, the model has no Kurdish acoustic training, and the transcription output is either silence, a string of random Arabic phonemes, or a garbled Arabic-like sequence that corresponds to nothing the speaker said.

For ASR transcription annotation of northern Iraqi speech, this means that single-language Arabic transcription guidelines are insufficient. Transcriptionists must be able to identify code-switched segments, accurately transcribe the Kurdish portion in a consistent orthographic representation (typically Kurdish Sorani script or a standardised romanisation), and mark the language boundary with the language tag expected by the ASR fine-tuning pipeline. Arabic-Kurdish bilingual transcriptionists — native Iraqi Arabic speakers with Kurdish literacy — are the only annotation resource that produces accurate transcription of northern Iraqi mixed-language speech at production quality.

The Iraqi Arabic ASR Training Data Gap and Why It Cannot Be Filled With Synthesis

Synthetic speech generation for Iraqi Arabic is appealing but does not solve the acoustic model problem. Text-to-speech systems for Iraqi Arabic are extremely limited in 2026 — no major TTS provider offers a production-quality Baghdadi dialect voice that accurately realises the qaf-to-gaf substitution, the Baghdadi vowel elision patterns, and the full phonological feature set of natural Iraqi speech. The few experimental Iraqi Arabic TTS systems that exist in research produce speech that native Iraqis describe as heavily accented and unnatural — useful for generating augmentation data at the margins but not for building the acoustic model core.

Data augmentation from other Arabic dialects is similarly inadequate. Egyptian Arabic /g/ (the jeem-to-gaf shift in Egyptian) is superficially similar to Baghdadi gaf but phonologically distinct — using Egyptian gaf data to train a Baghdadi gaf model produces an acoustic representation that partially overlaps with Baghdadi phonology but introduces Egyptian phonological artefacts that degrade accuracy on distinctively Baghdadi phonological patterns. Gulf Arabic data augmentation is even further from Baghdadi phonology and introduces qaf-as-hamza patterns that conflict with Baghdadi qaf-as-gaf representation.

The practical implication is that Iraqi Arabic ASR improvement requires real Iraqi speech, accurately transcribed by native Iraqi speakers. There is no shortcut through synthesis or cross-dialect augmentation that reaches production-grade Iraqi Arabic ASR accuracy. Our multilingual speech transcription service provides native Iraqi Arabic transcription for ASR training across all three major Iraqi sub-dialects, with Arabic-Kurdish bilingual coverage for northern Iraqi content.

Case Study: Baghdad Government Services Call Centre — WER From 48.7% to 19.1%

A Baghdad-based government services call centre — handling citizen enquiries for a national social welfare agency — deployed an Arabic ASR transcription system to automate call logging, complaint categorisation, and case creation for 14,000 daily citizen interactions. The system used a commercial Arabic ASR engine described by its vendor as "multi-dialect Arabic" — but trained primarily on Gulf and Egyptian Arabic with no Iraqi-specific acoustic data.

Before: The ASR system achieved 48.7% word error rate on Baghdadi citizen calls — measured against a manually transcribed reference set of 500 calls. WER on calls from Mosuli citizens (approximately 11% of call volume) was 71.3%, driven by Kurdish phoneme imports and code-switching that the Arabic-only acoustic model could not handle. At 48.7% WER, automated complaint categorisation accuracy was 29.4% — below the agency's threshold for automation trust, requiring 100% human review of ASR output. The automated call logging system was producing more transcription errors per call than the agency's manual transcriptionists were correcting, effectively adding cost rather than reducing it.

The transcription annotation project delivered 185 hours of Iraqi Arabic speech annotation across the agency's call volume profile: 68% Baghdadi, 19% Basrawi, and 13% Mosuli (with Kurdish code-switching in 41% of Mosuli calls). Transcription was conducted by a team of twelve native Iraqi transcriptionists — seven Baghdadi-native, three Basrawi-native, and two Mosuli Arabic-Kurdish bilingual — following transcription guidelines that included explicit notation for the qaf-to-gaf substitution (transcribed as گ in the Arabic-script output and as <g> in the romanised parallel), Kurdish-language segment tagging, Basrawi emphatic vowel quality notation, and filled-pause retention for prosodic alignment. Final transcription accuracy (measured against a double-blind reference transcription of a random 5% sample) was 97.3% on Baghdadi content and 94.1% on Mosuli code-switched content.

After fine-tuning on the Iraqi transcription dataset: Overall WER on Baghdadi speech fell from 48.7% to 19.1%. Mosuli speech WER fell from 71.3% to 34.7% — still higher than Baghdadi due to the Kurdish code-switching complexity, but within the threshold for automated complaint categorisation with human-in-the-loop review of Kurdish segments only. Automated complaint categorisation accuracy improved from 29.4% to 71.8%. The agency moved to automated case creation for 68.3% of calls (up from 0%), with human review focused on Mosuli Kurdish-segment calls and high-complexity complaint categories.

Call logging cost per interaction fell from AUD $4.30 (fully manual) to AUD $1.40 (automated with selective human review). At 14,000 daily interactions, the annualised cost reduction was AUD $14.6M — driven primarily by the shift to automated case creation and the elimination of full-transcript human review for Baghdadi and Basrawi calls. The transcription annotation project cost AUD $52,000 for transcription, guideline development, quality assurance at 5% sampling rate per transcriptionist, and delivery of timestamped transcriptions in both Arabic-script and forced-alignment-compatible CTM format.

Iraqi Arabic Speech Transcription: What the Annotation Protocol Requires

Iraqi Arabic ASR transcription annotation requires protocol elements that generic Arabic transcription guidelines do not address.

Explicit qaf-to-gaf transcription convention. Annotation guidelines must specify how the Baghdadi gaf realisation is transcribed — whether as the Arabic گ character, as the qaf character (normalising to MSA orthography), or as a phonetic annotation in a parallel field. For ASR training, representing the actual phonological realisation (gaf, not qaf) in the transcription produces a more accurate acoustic model for Iraqi speech. Guidelines that normalise to qaf produce training data that teaches the model to expect a phoneme that Baghdadi speakers do not produce.

Vowel elision retention rather than normalisation. Transcription guidelines must specify that Baghdadi vowel elision patterns are transcribed as produced — not normalised to MSA vowel-full forms. Transcriptionists who normalise elided forms to MSA produce training data where the acoustic output (elided) does not match the transcription label (MSA vowel-full), introducing systematic alignment errors in forced-alignment ASR training pipelines. Native Iraqi transcriptionists naturally produce elision-accurate transcription; non-native transcriptionists normalise to MSA forms by default and require explicit guideline instruction to transcribe elided forms as heard.

Kurdish language-switch markup for northern Iraqi content. Transcription of northern Iraqi speech requires standardised markup for Kurdish-language segments: a language-switch tag, Kurdish transcription in Sorani Arabic script or standardised romanisation, and a return-to-Arabic tag at the end of the Kurdish segment. The markup format must be compatible with the downstream ASR fine-tuning pipeline's handling of mixed-language data — different ASR frameworks handle multilingual training data differently, and transcription annotation should specify the exact markup schema the pipeline expects.

Filled pause and disfluency retention. Iraqi Arabic speech transcription for ASR training should retain filled pauses (آه، أممm، يعني used as filler), false starts, and repetitions rather than cleaning them to fluent text. ASR models trained on fluent-text transcriptions produce higher WER on real conversational speech than models trained on disfluency-retained transcriptions, because real speech contains disfluencies that the model must learn to handle. Our Arabic NLP annotation service provides disfluency-retained Iraqi Arabic transcription with explicit protocol for each disfluency category and consistent notation across all transcriptionists.

Related Reading

• Iraqi Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators
• Iraqi Arabic Sentiment Analysis: Loanwords, Tribal Register, and Kurdish Code-Switching
• Multilingual Speech Transcription Annotation Done at Scale
• Where Do Arabic NLP Datasets Come From — and How Do You Build Your Own?
• Arabic Data Labeling Service

Frequently Asked Questions