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

Why Yemeni Arabic Breaks Standard Arabic ASR Models

Standard Arabic ASR models — including Whisper (Arabic), Wav2Vec2-Arabic, and commercial Arabic ASR APIs — are trained predominantly on MSA broadcast speech, Quranic recitation, and Egyptian or Gulf Arabic conversational data. Yemeni Arabic diverges from all of these at the acoustic, phonological, and lexical levels simultaneously.

Yemeni Arabic is one of the most phonologically conservative Arabic dialect clusters. It preserves consonant contrasts that MSA and most other dialects have merged or simplified: the distinction between qaf (ق) pronounced as a uvular stop and its pharyngealised variant, emphatic consonants with distinct acoustic realisations, and vowel length contrasts that are phonemically significant in Classical Arabic but absent from most modern dialect speech. These distinctions mean Yemeni speech contains acoustic events that MSA acoustic models have never been trained to recognise.

At the lexical level, Yemeni Arabic retains a significant stratum of vocabulary from Old South Arabian substrate languages — the ancient South Semitic languages of Yemen that predate Arabic — that have no equivalents in MSA or other Arabic dialects. These words produce high out-of-vocabulary rates in standard Arabic ASR language models, generating substitution errors that cascade through downstream NLP tasks. For voice AI applications serving Yemeni users, these OOV words often appear in exactly the highest-value semantic positions: names, place names, products, and service terms.

Evaluations at the Interspeech 2022 Dialect Recognition Challenge and the ODYSSEY 2024 Arabic Speaker and Language Characterisation workshop report WER degradation of 45–58% on Yemeni dialect speech for models optimised on MSA data, with the sharpest OOV rates in rural and Hadrami content (Biadsy et al., 2022; Dehak et al., 2024). Yemeni Arabic is among the least-resourced Arabic dialect in publicly available ASR corpora — the Common Voice Arabic dataset contains fewer than 0.5% Yemeni speaker contributions.

Five Phonological Challenges That Raise WER on Yemeni Arabic Speech

1. San'ani qaf and uvular stop phonology

San'ani Arabic preserves the Classical Arabic qaf (ق) as a pharyngealised uvular stop — a sound that has merged with glottal stop or velar stop in most Arabic dialects, including Egyptian (glottal) and Gulf (velar/uvular without pharyngealisation). MSA ASR acoustic models trained on news broadcast Arabic model qaf as an MSA uvular without the San'ani pharyngealisation feature. In San'ani speech, this produces consistent substitution errors wherever qaf appears — a high-frequency consonant in Arabic.

Correct transcription of San'ani speech requires transcribers who can hear and orthographically represent this distinction. For ASR training data, the decision about how to orthographically represent the San'ani qaf — standard Arabic ق with or without diacritics, or a dialect-specific marker — must be made in the transcription style guide before production and applied consistently. Inconsistent orthography for the same phoneme across the training dataset produces acoustic model confusion that reduces WER improvements.

2. Old South Arabian substrate vocabulary with no MSA equivalents

A significant portion of everyday Yemeni Arabic vocabulary derives from Old South Arabian — the ancient South Semitic language family of pre-Islamic Yemen, which includes Sabaean, Minaean, Qatabanian, and Hadhramatic. These words have survived in Yemeni dialect as household, agricultural, cultural, and geographic terms with no MSA equivalents. In ASR, they appear as OOV events in the language model, producing substitution errors as the decoder forces the unknown sound sequence onto the closest MSA phonological match.

For voice AI applications in agriculture, traditional commerce, healthcare in rural Yemeni communities, or any domain involving geographic names, these OOV events can dominate WER because they concentrate in the highest-information positions of utterances. Building Yemeni ASR training data that covers this vocabulary layer requires native Yemeni transcribers who recognise and can correctly orthograph these terms — non-native transcribers often leave blanks or use phonetically similar MSA substitutes that train the ASR model with incorrect targets.

3. Hadrami distinctive phonology and diaspora prosody

Hadrami Arabic — spoken in Yemen's Hadramawt governorate and by a large diaspora in Saudi Arabia, UAE, East Africa, and Southeast Asia — has distinctive phonological features that differentiate it from both San'ani Arabic and Gulf Arabic. Hadrami preserves certain interdental fricatives (ث, ذ, ظ) in a manner closer to Classical Arabic than San'ani does, uses different vowel harmony patterns, and in the Gulf diaspora has absorbed Gulf Arabic prosodic features (stress patterns, speech rate norms) that differ from homeland Hadrami.

For Gulf-facing voice AI products serving the Hadrami diaspora, training data from homeland Hadrami speakers alone will miss the prosodic features of diaspora speech. The acoustic model will perform well on heritage speakers (older diaspora with homeland-proximate phonology) but poorly on second-generation diaspora whose Hadrami Arabic has been influenced by Gulf Arabic prosody. Transcription annotation should tag speakers with a homeland versus diaspora register flag so that training data can be stratified appropriately.

4. Adeni code-switching with South Asian and English phonology

Adeni Arabic — the dialect of Aden, Yemen's port city — has been significantly influenced by Indian Ocean trade contact, with loanwords from Hindi, Urdu, Gujarati, Swahili, and British English incorporated into everyday speech. Adeni speakers produce code-switching at the word level in commercial, medical, and logistics contexts, inserting South Asian or English lexical items with their source-language phonology into otherwise Arabic sentences. MSA ASR models have no acoustic model coverage for this phonological inventory and produce high substitution error rates on Adeni code-switched speech.

Transcription guidelines for Adeni speech should specify how code-switched segments are to be handled: whether to transcribe them in their source orthography (Latin for English, Devanagari for Hindi) or phonetically in Arabic script. For ASR training, source-orthography transcription allows a multilingual ASR model to handle Adeni code-switching without requiring Adeni-specific acoustic models for every contact language.

5. Conflict-era vocabulary producing OOV spikes in Yemeni speech data

Since the escalation of the Yemen civil conflict in 2014, new vocabulary has entered everyday Yemeni Arabic speech — terms for factions, displacement conditions, humanitarian organisations, and conflict-era technologies. In spoken data collected from Yemeni users during or after the conflict period, this vocabulary appears frequently in news commentary, community discussion, and even service interactions. It produces OOV spikes in standard Arabic ASR language models because none of this vocabulary is present in MSA or pre-conflict Arabic ASR training corpora.

For ASR training datasets collected from Yemeni users in 2018 onwards, conflict-era vocabulary should be treated as a mandatory domain adaptation layer. Transcription style guides should include the most common conflict-era Yemeni terms with their agreed orthographic forms, so transcribers handle these consistently rather than each making individual spelling decisions that fragment the language model's vocabulary.

Case Study: Yemeni Diaspora Healthcare Hotline — WER 61% to 24%

A humanitarian healthcare organisation operating a phone triage hotline for Yemeni displaced communities in Saudi Arabia and UAE sought to automate call triage using an Arabic ASR system. The system needed to identify caller location (governorate of origin), symptom category, and urgency level from spoken Yemeni Arabic calls. The existing system used a commercial Arabic ASR API optimised for MSA and Egyptian Arabic.

Before: The ASR system achieved a word error rate of 61.3% on held-out Yemeni Arabic call recordings. Location identification accuracy — identifying the caller's governorate of origin from geographic names — was 19.4% because Old South Arabian-derived place names were systematically substituted with the nearest MSA phonological match, producing wrong or non-existent place names. Symptom category assignment accuracy was 34.7%, driven by OOV errors on Yemeni dialect terms for body parts and symptoms that had no MSA equivalents. The triage system was producing incorrect urgency classifications in more than half of calls, requiring 100% human review and defeating the purpose of automation.

The transcription annotation project produced 180 hours of transcribed Yemeni Arabic telephone speech across San'ani (90 hours), Hadrami homeland and Gulf-diaspora (60 hours), and Ta'iz–Ibb (30 hours) sub-dialects. Transcription was performed by a team of fourteen native Yemeni transcribers — six San'ani-native, five Hadrami-native (two with Gulf-diaspora backgrounds), and three Ta'iz-native — working with a Yemeni dialect orthography style guide developed in collaboration with a Yemeni linguist at the University of Aden. The style guide specified orthographic conventions for 340 Yemeni dialect terms — including 120 Old South Arabian substrate terms and 85 conflict-era terms — with no standard MSA spellings. Speaker metadata tagged each recording with sub-dialect, age range, gender, audio condition (phone vs recorded), and displaced versus homeland speaker status.

After ASR fine-tuning on the transcribed dataset: Word error rate improved from 61.3% to 23.8%. Location identification accuracy improved from 19.4% to 79.3% — the Old South Arabian place name vocabulary was now in the language model. Symptom category assignment improved from 34.7% to 81.2%. Urgency classification accuracy reached 76.4%, reducing required human review from 100% to 28% of calls and enabling the triage system to handle the call volume without additional staffing. The transcription project cost AUD $68,000 for 180 hours of native-transcribed, QA-reviewed, style-guide-consistent Yemeni Arabic audio. The healthcare organisation attributed an estimated AUD $290,000 in operational saving in the first year from reduced manual triage workload, with a secondary benefit of faster triage times for high-urgency callers.

Transcription Protocol for Yemeni Arabic Speech Annotation Projects

Yemeni Arabic speech transcription annotation requires a structured protocol that differs from generic Arabic ASR transcription in five operational respects.

Dialect orthography style guide before production. The single most important project artefact is a Yemeni dialect orthography style guide agreed before the first batch of transcription begins. The guide should specify: orthographic conventions for Yemeni phonemes without standard MSA representation (particularly qaf variants and Old South Arabian vocabulary); code-switching handling for Adeni South Asian and English segments; conflict-era term spellings; and prosodic marker conventions (pause, emphasis, laughter, noise). Without this guide, transcribers make individual spelling decisions that fragment the vocabulary and reduce language model quality.

Sub-dialect-separated transcription pools. San'ani and Hadrami speech should be transcribed by sub-dialect-matched transcribers. Cross-dialect transcription on Old South Arabian substrate words — which differ between San'ani and Hadrami — produces systematic substitution errors that cannot be caught by automated QA because both the error and the correct form are non-MSA. Minimum pools: six San'ani-native transcribers for San'ani content, four Hadrami-native for Hadrami homeland content, three Hadrami with Gulf-diaspora experience for Gulf-diaspora Hadrami content.

Speaker metadata collection. Each recording in the ASR training dataset should be tagged with: sub-dialect (San'ani, Hadrami, Adeni, Ta'iz–Ibb), speaker gender and age range, audio condition (telephone, microphone, noisy environment), and homeland versus diaspora status. This metadata enables sub-dialect-stratified evaluation, targeted data augmentation for underrepresented speaker demographics, and acoustic model analysis when WER is higher on specific speaker groups.

Double-transcription for evaluation reference sets. ASR evaluation reference transcriptions — held-out test sets used to measure WER — should be double-transcribed: two native transcribers from the same sub-dialect independently transcribe each recording, with disagreements adjudicated by a senior linguist. Single-transcriber evaluation references on Yemeni Arabic produce noisy WER measurements because individual transcribers make different orthographic decisions for the same OOV items, making WER comparisons between model versions unreliable.

For multilingual speech transcription covering Arabic dialects beyond Yemeni, see our guide on multilingual speech transcription annotation at scale and our overview of multilingual speech transcription services.

Data Sourcing for Yemeni Arabic Speech Corpora

Yemeni Arabic speech corpora are among the most challenging to source ethically and at scale. The conflict context means that audio recordings of Yemeni speakers from some sources — particularly social media, news footage, and humanitarian organisation recordings — may contain sensitive personal information, trauma-adjacent content, or politically sensitive material. These issues require specific data governance attention before collection and annotation begins.

For read-speech corpora — where speakers read from prompts — Yemeni dialect text prompts should be developed in collaboration with native Yemeni linguists to ensure natural and representative vocabulary coverage, including Old South Arabian substrate terms that a Yemeni speaker would produce naturally but that would be absent from MSA-derived prompt text. Read-speech corpora developed from MSA prompts produce artificially low OOV rates in the training data, building an ASR model that overestimates its own coverage of natural Yemeni speech.

Consent collection from Yemeni speakers — particularly displaced communities — must follow humanitarian research ethics standards and be conducted in Yemeni Arabic, not MSA. Consent forms and audio release agreements should specify the AI training purpose, the anonymisation approach, and the data retention policy in plain Yemeni dialect. Consent obtained in MSA from speakers whose primary language is a Yemeni sub-dialect may not meet informed consent requirements under applicable data protection frameworks.

See our overview of Arabic NLP annotation services and our guide to sourcing Arabic NLP datasets for broader corpus governance guidance.

Related Reading

• Yemeni Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators
• Yemeni Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators
• How Is Multilingual Speech Transcription Annotation Done at Scale?
• Arabic NLP Annotation Service

Frequently Asked Questions