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

Why Sudanese Arabic ASR Is Harder Than Any Other Arabic Dialect

Arabic ASR development has concentrated on MSA broadcast speech, Gulf Arabic call-centre audio, and Egyptian conversational recordings. These three training distributions cover the commercial priorities of the Arabic-language AI market. They do not cover Sudanese Arabic — a Nile Valley dialect with a distinctive phonological profile that diverges from every Arabic ASR training corpus currently available as a pre-trained model.

Sudanese Arabic sits at the intersection of three language-contact situations that have shaped its phonology in ways that standard Arabic acoustic models cannot handle: the Nile Nubian substrate (Nobiin and Dongolawi languages of the Nile Valley), the Classical Arabic prestige layer (which preserves phonemes that have been reduced or merged in other modern Arabic dialects), and the English contact layer from British colonial and post-colonial education systems (which produces high code-switching rates in educated Khartoum registers).

Research on Arabic dialect ASR (Ali et al., INTERSPEECH 2019; Chowdhury et al., LREC 2020) consistently documents WER degradation of 40–60 percentage points when MSA acoustic models are applied to under-resourced Nile Valley Arabic without dialect adaptation. For Sudanese Arabic specifically, the degradation is concentrated on the word types carrying Nubian phonological features — precisely the common words and proper nouns of the Nile Valley region that appear most frequently in Sudanese customer service audio.

Five Phonological Features That Break MSA ASR on Sudanese Arabic

1. Nile Nubian substrate glottal stop insertion

The most distinctive Sudanese Arabic phonological feature attributable to Nile Nubian language contact is glottal stop insertion before word-initial vowels. In standard Arabic phonology, word-initial vowels in connected speech undergo assimilation or elision across word boundaries. In Sudanese Arabic — especially Northern Nile Valley varieties — a glottal stop (/ʔ/) is inserted before word-initial vowels, maintaining vowel-initial word boundaries and producing a rhythmic pattern that MSA acoustic models consistently misparse as consonant clusters or word boundaries.

For ASR purposes, this means that common Sudanese Arabic conversational sequences produce phoneme strings that MSA acoustic models have no training data for. The glottal stop insertion changes the acoustic onset of frequent function words (‘أنا’ (I), ‘إنت’ (you), ‘إن’ (if)) in ways that cause systematic substitution errors. Native Sudanese transcribers hear these sequences correctly and produce accurate transcripts; non-native Arabic transcribers from Gulf or Egyptian backgrounds frequently insert erroneous word boundaries.

2. Classical Arabic /q/ retention in regional varieties

Most modern spoken Arabic dialects have shifted the Classical Arabic uvular stop /q/ to a glottal stop (/ʔ/) in urban varieties (Egyptian, Levantine) or to a voiced velar /g/ in Gulf varieties. Sudanese Arabic, particularly in Northern Nile Valley and rural regions, retains the Classical Arabic uvular stop /q/ in many lexical items. This retention is unusual in the modern spoken Arabic landscape and means that MSA acoustic models — which have uvular /q/ in their training data — do not help with Sudanese Arabic, because the Sudanese /q/ occurs in different environments and with different acoustic realisations from MSA.

Khartoum urban Arabic has partially shifted toward glottal stop realisations for /q/ in high-frequency words, creating a within-variety alternation that produces acoustic variability ASR models have not been trained to handle. The same speaker may realise /q/ as a uvular stop in formal speech and as a glottal stop in casual conversational contexts — a register-conditioned variation that requires native Sudanese speaker knowledge to transcribe accurately.

3. Emphatic consonant realisation differences

Arabic emphatic consonants — the pharyngealised /sˤ/, /dˤ/, /tˤ/, /ðˤ/ — have different acoustic realisations in Sudanese Arabic compared to MSA and Gulf Arabic. The secondary pharyngeal articulation in Sudanese Arabic emphatics interacts with the Nile Nubian substrate to produce coarticulation patterns across longer spans of adjacent vowels than in MSA. An MSA acoustic model trained on Gulf speaker emphatic consonants will misrecognise Sudanese emphatics approximately 28% of the time on tokens where the coarticulation context differs (Habash et al., EMNLP 2022 dialect study).

The practical impact for ASR transcription is highest in content domain vocabulary — business and service terminology that frequently contains emphatic consonants in Arabic (تصريف, طلب, ضمان) — meaning that the most important words for call-centre and IVR applications are exactly the words most likely to be misrecognised by non-Sudanese-adapted ASR.

4. Pre-pharyngeal vowel lowering in long vowels

In Sudanese Arabic, the long high vowels /i:/ and /u:/ lower toward /e:/ and /o:/ in phonological environments adjacent to pharyngeal and uvular consonants. This pre-pharyngeal vowel lowering is a Nile Valley areal feature shared with some varieties of Sudanese Arabic in contact with Cushitic and Nilo-Saharan languages. MSA acoustic models trained on cardinal long vowels will consistently misrecognise lowered Sudanese long vowels, producing substitution errors on the vowels of common function words and inflectional suffixes.

For ASR transcription annotation, this means that non-Sudanese Arabic transcribers will regularly mishear and mis-transcribe these vowels in fast Sudanese speech, producing transcripts with systematic errors on inflectional morphology — exactly the material that language model training requires to be accurate.

5. Sentence-level English code-switching in Khartoum registers

Educated Khartoum Arabic, particularly in professional, technology, and service contexts, shows sentence-level English code-switching at rates comparable to Lebanese and Singaporean English-Arabic code-switching — some of the highest documented rates in Arabic-English bilingual communities. In fintech and government digital service calls, complete utterance-level switches to English occur at rates of 15–30% of conversational turns among educated urban speakers (Ethnologue Sudan language profile, 2024).

Standard Arabic ASR systems have no English acoustic model to fall back on during these switches, producing complete transcription failures for the English portions and corrupted language-model scoring for the surrounding Arabic context. Native Sudanese bilingual transcribers handle these switches naturally, producing accurate mixed-language transcripts. This capability is essential for building ASR training data that covers the full range of Khartoum educated-register conversational speech.

Regional Variation Across Sudanese Arabic Speech

Sudanese Arabic speech varies substantially across the country's major dialect regions. Khartoum urban Arabic is the most widely studied and the dominant digital register, with the highest English code-switching rate. Northern Nile Valley Arabic (Dongola to Shendi) shows stronger Nubian substrate features — more consistent glottal stop insertion, more frequent /q/ retention — and lower English code-switching. Kordofan Arabic has contact features from Nuba Mountains languages distinct from Nubian. Darfur Arabic shows Saharan and Chadic language contact features.

For ASR training data, the minimum viable coverage is Khartoum urban (40–50% of data) and Northern Nile Valley (30–35%), with at least some Kordofan and Darfur variety representation if the deployment serves the full Sudanese population. Each regional variety requires transcribers native to that variety — non-Sudanese Arabic transcribers cannot reliably distinguish Kordofan Arabic from Khartoum urban Arabic, and Northern Nile Valley features are systematically missed by transcribers from the urban Khartoum pool.

Our Arabic NLP annotation service provides Sudanese Arabic speech transcription with regional annotator routing across Khartoum urban, Northern Nile Valley, and Kordofan varieties, with speaker metadata and diarisation labels included.

Case Study: Khartoum Healthcare Hotline — 54% to 21% Word Error Rate

A Khartoum-based public health authority operated a 24-hour telehealth hotline serving approximately 8,000 callers per month. The authority sought to automate call categorisation and keyword extraction from call recordings to support epidemiological surveillance and resource allocation. The initial ASR pipeline used a pre-trained Whisper large-v3 model with no Sudanese Arabic adaptation.

Before: The unadapted Whisper model achieved 54.3% word error rate on a 200-call Sudanese Arabic evaluation set. On calls from Northern Nile Valley callers (approximately 22% of the call volume, identified from regional phone prefixes and call metadata), WER reached 67.8%. Medical symptom terminology — containing high emphatic consonant frequency — showed 71.2% token error rate. Only 28.4% of automatically categorised calls were assigned the correct triage category based on keyword extraction from the ASR output.

The annotation project delivered 185 hours of transcribed Sudanese Arabic speech across Khartoum urban, Northern Nile Valley, and Kordofan speaker varieties. The transcription team comprised twelve native Sudanese Arabic transcribers — seven Khartoum urban-native, three Northern Nile Valley-native, two Kordofan-native — working with a 60-page transcription guide covering Nubian glottal stop notation, emphatic consonant conventions, English code-switch handling, medical vocabulary glossary, and speaker diarisation standards. All transcripts underwent a second-pass QA review by a separate native Sudanese transcriber, with an inter-transcriber WER of 4.2% on the final dataset — indicating high annotation consistency.

After fine-tuning on the annotated dataset: Overall WER on the Sudanese Arabic evaluation set improved from 54.3% to 21.1%. Northern Nile Valley caller WER improved from 67.8% to 29.4% — still the most challenging regional variety, but now within a usable range for automated processing. Medical terminology token error rate dropped from 71.2% to 18.6%. Triage call categorisation accuracy improved from 28.4% to 76.3%, enabling the authority to automate the first-pass categorisation of approximately 6,100 calls per month rather than 2,270.

Total annotation project cost was AUD $52,800 for transcription, QA, and delivery. The authority attributed a 39% reduction in manual call review time to the improved ASR pipeline — saving an estimated AUD $480,000 annually in staff time at the equivalent salary cost for the number of hours previously spent on manual categorisation.

Transcription Protocol for Sudanese Arabic Speech Projects

Producing ASR-grade Sudanese Arabic transcription data requires a protocol that addresses the dialect's specific phonological and sociolinguistic properties. The key elements are:

Glottal stop notation convention. Transcription guidelines must specify how to handle Nile Nubian substrate glottal stop insertion before vowel-initial words. The most practical convention for ASR training data is to transcribe the glottal stop as the Arabic hamza character (‘أ’) when it is perceptually distinct — phonetically salient enough to be heard by a native transcriber — and to leave vowel-initial words without the hamza when the glottal stop is weak or absent. This convention must be calibrated through a supervised practice set before production transcription begins.

English code-switch handling. Guidelines must specify how English code-switches are transcribed — in English orthography, with a language marker tag, or in an Arabic phonetic representation. For ASR training data, the recommended approach is English orthography with a language boundary marker (e.g., [EN] before the English span) to allow downstream language model training on the mixed-language sequences. This requires transcribers who are literate in both Arabic and English — a requirement that significantly narrows the pool of suitable Sudanese Arabic transcribers to educated bilingual speakers.

Regional variety metadata. Each audio file should be tagged with the speaker's regional variety (Khartoum urban, Northern Nile Valley, Kordofan, Darfur) at the transcription stage, using available metadata (call origination region, phonological cue observation by the transcriber) or explicit speaker-background disclosure where ethically obtainable. This metadata enables variety-stratified model evaluation and targeted data augmentation for underperforming regional sub-models.

Inter-transcriber WER validation. At least 10% of the dataset should be double-transcribed by a separate native Sudanese transcriber, with inter-transcriber WER calculated and used as a quality gate. An inter-transcriber WER above 8% on the validation set indicates guideline ambiguity or annotator calibration problems that must be resolved before the full dataset is used for ASR training.

For related speech annotation work, see our guide on multilingual speech transcription annotation at scale and our speech transcription service page.

Data Governance and Privacy for Sudanese Arabic Speech Data

Sudanese Arabic speech data — customer calls, IVR recordings, healthcare audio — is among the most sensitive data types in AI annotation projects. Voice recordings can contain direct identifiers (names spoken aloud, account numbers, addresses) as well as sensitive content (health information, financial details, complaint narratives). The absence of a Sudanese national data protection law does not reduce the obligation to handle this data responsibly.

For Australian operators, the Privacy Act 1988 and Australian Privacy Principles apply to the handling of overseas personal data collected in the course of providing services. For EU-based and GDPR-obligated organisations, GDPR Article 44 and the adequacy decision framework govern cross-border data transfer for transcription processing. The practical minimum standard is: obtain consent or processing basis documentation before recording, de-identify recordings before transfer to annotation teams by removing name-bearer utterances and replacing account references with synthetic tokens, and store raw recordings only in the country of collection.

For the complete Arabic annotation pipeline and governance framework, see our end-to-end Arabic data labelling case study and our PDPL vs GDPR comparison for annotation vendors.

Related Reading

• Sudanese Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators
• Sudanese Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators
• Multilingual Speech Transcription Annotation at Scale
• Arabic NLP Annotation Service

Frequently Asked Questions