PaperlayerView exact revision on arXiv

Source-linked AI summary

CARDAMOM: A Micro-Dialectal Arabic Speech Dataset for ASR

Bashar Talafha, Samar M. Magdy, Aisha Alansari, Alaa Alkhawaldeh, Abdurrahman Juma, Sharaf Makahleh, Nour Gamal, Omar Attia, Hanaa Kurdi, Najwa Rizk, Maysa Anaya, Hessah Altimyat, Layal Alhazmi, Shumukh Alotaibi, Hajar Alhadaris, Rayan Alomari, Rahaf Almalaq, Malak Alkhorasani, Sara alghamdi, Rahaf Alshamrani, Nsrin Ashraf, Ibrahim Jaradat, Nada Qardahji, Yasmin Zaraket, Elmoukhtar Brahim, Sidi Ebeidy, Oumoulmouminin Mahmoud, Yahjeb Bouha Khatraty, Meya Haroune, Mohammad Ghaddar, Mohamad Eldirany, Rashed Alamoush, Tala Chhaytle, Nuha Albadi, Yahya El Hadj, Hamzah Luqman, Fadi A. Zaraket, Mustafa Jarrar, Muhammad Abdul-Mageed

arXiv:2609.34481v1cs.CL

TL;DR

Arabic speech resources often use broad regional or country-level labels, leaving fine-grained sub-country variation underrepresented. CARDAMOM provides community-curated, transcribed speech labeled across micro-dialects and benchmarks multilingual ASR and dialect identification. Its results show uneven ASR difficulty across varieties and that micro-dialect labels are learnable identification targets.

  • Problem

    Most Arabic speech resources use broad regional or country-level labels, leaving localized sub-country variation underrepresented.

    Source: §1 · p. 2

  • Method

    CARDAMOM provides approximately 40 hours of community-curated, transcribed YouTube speech across 21 micro-dialects, with segment-level labels and speaker metadata.

    Source: §1 · p. 2

  • Results

    Zero-shot benchmarking found uneven challenges across varieties, with Mauritanian and Lebanese varieties hardest and OMNIASR-LLM strongest overall; the labels also support learnable micro-dialect identification.

    Source: §8 · pp. 8–9

  • Takeaways & Limitations

    CARDAMOM's micro-dialect labels can serve as a primary evaluation target, not only as metadata for organizing ASR error breakdowns.

    Source: §8 · p. 9

    Coverage across countries is uneven, and conclusions about sparsely represented micro-dialects are constrained.

    Source: Limitations · p. 9

Abstract

from arXiv · show

We present Cardamom, a micro-dialectal Arabic speech dataset designed to support fine-grained evaluation and adaptation of automatic speech recognition (ASR) systems. Community-curated by native speakers familiar with the represented varieties, Cardamom contains approximately 40 hours of transcribed YouTube speech spanning 21 micro-dialects across Egypt, Jordan, Lebanon, Mauritania, Palestine, and Saudi Arabia. Each segment is annotated with one or more operational micro-dialect labels, code-switching information, and utterance-level perceived gender, enabling analysis of sub-country variation that is obscured by conventional country-level labels. We describe the collection and annotation process, motivate the micro-dialect inventory linguistically, and benchmark four multilingual ASR systems in zero-shot and adapted settings. The strongest zero-shot system obtains 43.47% aggregate WER, with particularly high error rates on Mauritanian and Lebanese varieties; adaptation on Cardamom reduces its WER to 35.21%. Audio-based identification experiments further show that the annotations provide a learnable prediction target, with a dedicated classifier reaching 85.57% accuracy on 21-way micro-dialect identification. Cardamom provides a resource for studying localized dialectal variation and developing Arabic speech systems with broader regional coverage.

1 Introduction

CARDAMOM targets the underrepresentation of localized Arabic speech varieties in resources that use broad regional or country-level labels. It pairs community-curated speech and fine-grained annotations to support evaluation of ASR and dialect identification below the country level.

1 Introduction

  • Existing Arabic speech resources often use broad labels, leaving fine-grained sub-country variation underrepresented despite substantial within-country variation.Country-level labels can obscure speech continua shaped by geography, contact, migration, and local identity.Source: §1 · p. 2
  • CARDAMOM contains approximately 40 hours of transcribed YouTube speech spanning 21 localized varieties across six countries.Native speakers familiar with the represented varieties curate and annotate the corpus.Source: §1 · p. 2
  • Its segment-level annotations include one or more micro-dialect labels and speaker-level metadata, including gender, supporting analysis beyond country-level categories.The dataset is designed to examine ASR and dialect identification at both country and micro-dialect levels.Source: §1 · p. 2

2 Related Work

Prior Arabic speech resources and multilingual modeling work have expanded dialectal coverage, but most spoken resources still organize varieties at broad regional or national levels. CARDAMOM addresses the remaining fine-grained supervision gap.

2 Related Work

  • Arabic speech resources have grown from broadcast and web-derived corpora to multidialectal, country-focused, and low-resource datasets.Examples include NIST EARS, DARPA GALE, MGB Challenge, QASR, MASC, Aswat, and SADA.Source: §2 · p. 2
  • Most spoken Arabic resources remain organized around countries or broad regions, leaving micro-dialectal supervision underrepresented.CARDAMOM adds approximately 40 hours labeled across 21 micro-dialects for within-dialect evaluation and cross-micro-dialect generalization.Source: §2 · pp. 2–3
  • Related work also addresses code-switching, orthography, and transcription evaluation, while multilingual ASR research reports continuing challenges from dialect variation and data imbalance.Examples include bilingual Arabic-English corpora and Whisper adaptation for multidialectal Arabic ASR.Source: §2 · pp. 2–3

3 Dataset Description

CARDAMOM is collected from public YouTube speech and organized around localized varieties, with community annotators assigning segment-level labels. Its release provides annotations and timestamps rather than redistributed audio.

3 Dataset Description

  • The corpus draws on naturally occurring, preferably non-scripted YouTube speech from six countries, including interviews, vlogs, and informal discussions.Candidate local varieties were elicited from native-speaker annotators, and similar or hard-to-distinguish candidates were consolidated.Source: §3 · p. 3
  • Source videos were assigned a dominant variety for collection, while final segment labels could include multiple compatible micro-dialects.The inventory reflects within-country geographic variation, with more localized categories in larger countries than in the smaller Levantine countries.Source: §3 · p. 3
  • CARDAMOM releases source video IDs, transcriptions, labels, perceived gender, code-switching status, timestamps, and reconstruction code, but not audio.Collected recordings were segmented into annotation units of at most 20 minutes.Source: §3 · p. 3
  • The 31 annotators are native speakers who primarily labeled varieties from their own micro-dialects, with some annotating multiple varieties.Counts are not mutually exclusive, and utterances may receive more than one label.Source: §3 · p. 3

4 Annotation Tasks and Guidelines

The annotation scheme combines transcription, dialect labels, perceived gender, code-switching, and timestamps at the segment level. Annotators select complete, clear speech units, preserve everyday spelling, and coordinate to resolve quality issues.

4.1 Overview

  • A unified segment-level schema supports ASR, dialect identification, micro-dialect and code-switching analysis, and aggregation by speaker metadata.Each segment includes a transcription, country and micro-dialect labels, perceived gender, a code-switching indicator, and timestamps.Source: §4.1 · p. 4, §4.2 · p. 4

4.2.1 Task 1: Segment Selection

  • Annotators select clear, complete utterances from a single speaker, preferably no longer than 30 seconds, with intact word boundaries.Mild background noise is allowed, and a segment must contain at least one word.Source: §4.2.1 · p. 4
  • Segments receive one or more plausible micro-dialect labels when speech may belong to multiple local varieties.The example utterance is labeled as both Falahi and Madani when compatible with both varieties.Source: §4.2.1 · p. 4

4.2.2 Task 2: Transcription

  • Transcriptions prioritize fidelity to spoken dialect over normalization to Modern Standard Arabic, and code-switched speech receives Arabic- and Latin-script variants.Annotators use natural written forms and may use standard orthographic markers where appropriate.Source: §4.2.2 · p. 4

4.2.3 Task 3: Speaker Gender

  • Speaker gender is labeled as perceived from the speech signal for aggregate analyses, not to identify speakers or assert self-identified gender.The available labels are male and female.Source: §4.2.3 · p. 4

4.3 Quality Control

  • Written guidelines, coordination meetings, and manual inspection were used to resolve ambiguity and check transcript completeness, segment boundaries, single-speaker constraints, and label consistency.Ambiguous cases were discussed with annotators familiar with the relevant local varieties.Source: §4.3 · p. 4, §4.1 · p. 4

5 Micro-Dialects Overview

CARDAMOM treats micro-dialects as localized sub-country varieties and uses operational categories to examine variation obscured by national labels. The inventory is motivated by phonological differences but is not presented as a definitive linguistic classification.

5 Micro-Dialects Overview

  • Micro-dialects are localized varieties recognizable to speakers through systematic differences in pronunciation, morphology, lexical choice, prosody, or grammar.The paper motivates this granularity because national labels can mask substantial within-country variation.Source: §5 · pp. 4–5
  • The taxonomy adapts prior descriptions into operational annotation categories, not a definitive classification, and its country-level groupings do not exhaust all local varieties.Categories are selected for recognizability to native-speaker annotators, frequency in public speech, and usefulness for testing generalization beyond country labels.Source: §5 · p. 5

5.1 Phonological Motivation for Micro-Dialect Annotation

  • Contrasts involving /q/, interdental sounds, velar palatalization, and affrication vary across the sampled varieties, motivating annotation below the country level.The passage illustrates variation in the realization of the word for “heart” across Cairene, Palestinian North Rural, and Hassaniya speech.Source: §5.1 · p. 5

6 Corpus Statistics and Analysis

The corpus statistics describe a substantial but uneven sample, with dialect-dependent profiles across duration, code-switching, gender, transcript characteristics, and overlap between labels. The overlap patterns are interpreted in light of the sampled speech and annotation framework.

6.1 Micro-dialect and Transcript Statistics

  • 27,420 segments totaling approximately 40 hours make up CARDAMOM, with substantial differences in representation across micro-dialects.Some varieties, including Egyptian Cairene and Jordanian dialects, are well represented, while Lebanese B’albaki and Saudi Northern are less represented.Source: §6.1 · pp. 5–6
  • Segment duration, code-switching, and perceived-gender distributions vary by micro-dialect; Mauritanian segments are longest, while Egyptian segments are shorter.Mauritanian, Jordanian Madani, Lebanese, and Egyptian varieties contribute the largest amounts of code-switching, while several Palestinian, Saudi, and Jordanian Bedouin subsets contain very little.Source: §6.1 · p. 6
  • Lebanese and Mauritanian subsets have long transcripts and high vocabulary rates, while Egyptian and some Saudi varieties are faster and Palestinian and Lebanese varieties slower.The paper cautions that source-domain composition and segmentation may influence these statistics, so they may not reflect linguistic variation alone.Source: §6.1 · p. 6

6.2 Quantitative Micro-dialectal Overlap

  • Mauritania has the largest pairwise overlap counts, especially among its three regional labels, as much available speech was broadly shared rather than clearly localized.Annotators therefore often judged Mauritanian segments compatible with multiple regional labels.Source: §6.2 · p. 6
  • Jordan’s Badawi–Falahi and Falahi–Madani co-assignments number 178 and 171 segments, respectively, versus 23 for Badawi–Madani.Egyptian overlap centers on Cairene with Alexandrian (130) and Port Saidi (67); Saudi intersections are mostly sparse, which does not establish categorical linguistic separation.Source: §6.2 · p. 6
  • Multi-label assignments can reflect shared lexical or discourse material, limited variety-specific cues in short utterances, and overlap shaped by the annotation framework.Co-assignment counts may also be affected by differences in annotators’ use of multiple labels and by recurring speakers in YouTube material.Source: §6.2 · pp. 6–7

7 Zero-Shot Benchmarking

The paper benchmarks multilingual ASR on CARDAMOM and examines adaptation, audio-based micro-dialect identification, and the relationship between dialectness and recognition difficulty. Adaptation lowers aggregate error for the strongest system, while dialectness alone does not account for recognition difficulty.

7.1 Automatic Speech Recognition (ASR)

  • OMNIASR-LLM-7B-V2 has the best WER on 20 of 21 micro-dialects and best CER on 17 of 21; Mauritanian varieties remain hardest, with best-model WER around 76.Lebanese B’albaki is the WER exception, where Whisper-v3 is best at 62.2; all systems struggle on that variety.Source: §7.1 · p. 7

7.2 Adaptation

  • 35.21% WER and 16.04% CER for adapted OMNIASR-LLM compare with 43.47% WER and 20.67% CER zero-shot.Adaptation reduces both error rates for Whisper-v3 and the two OMNIASR variants; Seamless-v2’s WER reduction is smaller, from 53.13% to 52.48%.Source: §7.2 · p. 8

7.3 Micro-Dialect Identification

  • The dedicated Whisper classifier reaches 95.67% country, 85.57% global micro-dialect, and 88.20% country-conditioned accuracy.LoRA fine-tuning improves global micro-dialect identification by 58.46 points, and hierarchical consistency rises from 33.67% to 95.48%.Source: §7.3 · p. 8

7.4 Dialectness Analysis

  • MSA-only segments score lowest on dialectness (8.5%) and multi-label intersections highest (up to 97.1%), but dialectness alone does not explain recognition difficulty.Jordan-Madani and Egypt-Cairene remain comparatively easy to recognize despite high dialectness scores.Source: §7.4 · p. 8

8 Conclusion

CARDAMOM provides community-curated speech data for fine-grained Arabic ASR evaluation, and its conclusion reports uneven challenges across micro-dialects, with adaptation narrowing but not closing the gaps. The limitations and ethics discussion bounds the resource’s coverage, source composition, and intended use.

Ethics and Data Statement

  • CARDAMOM covers 21 localized varieties across six countries, with utterance-level annotations and standoff release artifacts rather than redistributed audio.The release includes human-produced transcriptions, micro-dialect labels, perceived gender, code-switching status, timestamps, and reconstruction code; gender is perceived from speech and not self-identified.Source: §8 · p. 8, Ethics and Data Statement · pp. 10–11

8 Conclusion

  • Zero-shot ASR results show uneven challenges, with Mauritanian and Lebanese varieties hardest; adapting on CARDAMOM narrows but does not close these gaps.Micro-dialect labels are also learnable as a downstream identification target, supporting their use for evaluation rather than only organizing ASR error breakdowns.Source: §8 · p. 9

Limitations

  • Perceived-gender distributions are skewed in some subsets, so the metadata support aggregate description rather than balanced subgroup comparisons across every micro-dialect.The authors attribute these imbalances to the availability of suitable public speech rather than a principled sampling strategy.Source: Limitations · p. 9
  • Cross-border varieties can share phonological and lexical features, making some micro-dialect boundaries difficult to distinguish; multi-label annotation accommodates plausible overlap.Jordanian Badawi and Saudi Northern are given as examples of varieties with shared Bedouin-type features.Source: Limitations · p. 10
  • Residual orthographic variation may inflate WER and CER beyond purely acoustic recognition errors despite shared guidelines and normalization.The authors cite multi-reference evaluation as a future direction for this challenge.Source: Limitations · p. 10
  • Most segments were not independently labeled by multiple annotators, so the paper does not report full-corpus inter-annotator agreement.Future work proposes targeted double annotation of selected micro-dialect pairs to quantify boundary uncertainty.Source: Limitations · p. 10
  • YouTube source videos may disappear, and uncontrolled, uneven genre and recording-condition distributions can influence cross-variety statistics.The authors release standoff annotations and interpret duration, code-switching, and speaking-rate statistics descriptively in light of possible source-domain effects.Source: Limitations · pp. 9–10
  • At approximately 40 hours, CARDAMOM is intended for evaluation, few-shot use, and adaptation of existing models, not training an end-to-end ASR system from scratch.The authors characterize it as a low-resource diagnostic benchmark rather than a training-scale corpus.Source: Limitations · p. 10

A Related Work

Existing Arabic speech resources have broadened dialect coverage, but their country- or region-level labels leave localized speech variation underrepresented. CARDAMOM targets that gap with practical micro-dialect categories informed by linguistic work and native-speaker judgments, while treating its taxonomy as operational rather than definitive.

A Related Work

  • Arabic speech resources and shared tasks expanded dialect coverage, but speech supervision remains mostly organized by country or broad region rather than micro-dialect.This leaves a gap between fine-grained dialect modeling in Arabic NLP and available supervision for Arabic ASR.Source: A Related Work · p. 17
  • CARDAMOM complements broad-coverage resources with micro-dialect-level supervision, enabling evaluation of within-dialect variation and cross-micro-dialect generalization.The paper describes it as the first Arabic ASR resource designed explicitly around micro-dialect supervision across multiple localized varieties.Source: A Related Work · p. 18

B.1 Operational Status of the Taxonomy

  • The proposed categories are practical annotation and evaluation labels, not a definitive linguistic classification; finer subdivisions may be warranted.The authors note that they did not conduct the comprehensive linguistic analysis needed to establish theoretical boundaries, and give Jordanian Falahi as a category that can be subdivided further.Source: §B.1 · p. 19

B.2 Cross-Dialect Phonological Variation

  • Table 3 summarizes representative phonological differences across Egyptian, Jordanian, Palestinian, Saudi, Lebanese, and Mauritanian varieties relative to MSA.Its rows identify MSA phonemes, while the second column lists attested variety-specific realizations and lexical examples.Source: §B.2 · p. 19, Table 3 · p. 19

B.3 Country and Micro-dialect Coverage

  • The inventory also distinguishes Northern and Southern/Eastern Ḥassāniyya groupings in Mauritania, urban, Northern Fallahi, and Southern Badawi speech in Palestine, and five broad regional groups in Saudi Arabia.The authors characterize Saudi labels as high-level groupings rather than exhaustive representations of local varieties.Source: §B.3 · p. 20
  • The inventory adapts prior dialectological descriptions into localized categories intended to be recognizable to annotators, usable with collected speech, and useful for evaluation beyond country-level labels.It covers Egypt, Jordan, Lebanon, Mauritania, Palestine, and Saudi Arabia; the country groupings include Egyptian Cairene, Coastal, and Saidi, and Jordanian Madani, Falahi, and Badawi varieties.Source: §B.3 · pp. 19–20

B.4 Phonological Variation in Arabic: A Micro-Dialectal Analysis

This section examines phonological and regional variation across Arabic micro-dialects, focusing on /q/ reflexes, interdental realization, and velar palatalization. It shows how broad labels encompass distinct local patterns and motivates practical regional categories for Egyptian, Lebanese, and Hassaniya speech.

B.4 Phonological Variation in Arabic: A Micro-Dialectal Analysis

  • The reflex of /q/ varies across micro-dialects, including glottalization, [g], [k], [tS], and retention, with some varieties showing mixed patterns.Cairene, Port-Saidi, and Jordanian and Palestinian Madani varieties glottalize /q/; Saidi and Hassaniya use [g], North Rural Falahi uses [k], and South Rural Badawi uses [tS].Source: §B.4 · p. 21
  • Interdentals are retained in several Saudi, Jordanian, Palestinian, and Hassaniya varieties, but merge with stops or sibilants in Egyptian, Hijazi, Madani, and Lebanese varieties.B’albaki shows variable retention and merger, while Hassaniya preservation is also subject to regional and lexical variation.Source: §B.4 · p. 21
  • Conditional /k/ palatalization to [tS] occurs in Jordanian Falahi and Badawi and Palestinian North Rural Falahi, but is absent in the corresponding examples from several other varieties.The Palestinian examples suggest the process may be both micro-dialect-specific and lexically conditioned; South Rural Badawi retains [k] in one cited form.Source: §B.4 · p. 21
  • Hassaniya examples show /q/ realized as [g], variation involving ghayn, de-emphasis of /sQ/, initial hamza reduction, and a Western/Gabla emphatic reflex.The passage frames these patterns as part of a regional continuum with shared and localized phonological features.Source: §B.4 · p. 21

B.5 Egyptian Micro Dialects

  • For Egyptian sampling, CARDAMOM uses Cairene, Coastal/Delta urban, and Saidi categories as practical annotation groups, not replacements for atlas-based classifications.The atlas-based account contrasts Cairo/central Delta /j/→[g], /q/→[ʔ] with Sinai Bedouin /j/→[dZ] or [Z], /q/→[g], and describes interdental and morphological isoglosses.Source: §B.5 · p. 22

B.7.2 Southern

  • CARDAMOM’s explicit Lebanese micro-dialect split covers Southern, Chamali, and Bikaa varieties to help address thin computational coverage of Lebanese.The passage identifies Curras+Baladi as the only morphologically annotated Lebanese corpus to date, alongside Syrian-side Nâbra.Source: §B.7.2 · p. 25

B.8 Mauritanian Micro-dialects

  • CARDAMOM treats Hassaniya variation through practical Western, Northern, and Southern/Eastern geo-social categories rather than rigid dialect boundaries.The categories simplify four broad literature sub-dialects; the passages characterize variation as more evident in pronunciation, lexicon, and usage than in deep structure, and note preservation of interdental contrasts.Source: §B.8 · p. 25, §B.8.3 · p. 26

B.8.3 Southern

  • Hassaniya /q/ may surface as [g] or alternate with /G/, while syllable shortening and consonant clustering distinguish it from MSA transcription norms.Contact also contributes marginal /v/, particularly in borrowed words; MSA-normalized transcription may erase features relevant to micro-dialect modeling.Source: §B.8.3 · p. 26

B.9 Palestinian Micro-dialects

  • Palestinian speech forms a continuum with mixed local and social features, analyzed here as urban, northern rural, and southern rural dialects.The three-part scheme reflects recurring co-occurrence across phonological, morphosyntactic, and lexical features rather than a rigid geographic taxonomy.Source: §B.9 · pp. 26–27

B.9.1 Madani – Urban Palestinian

  • Urban Palestinian Arabic commonly realizes /q/ as a glottal stop and /th, dh/ as /t, d/, while displacement helped produce a leveled Ramallah urban koine.The Ramallah koine is also described as the “white” dialect and as similar to the Amman urban koine.Source: §B.9.1 · p. 27

B.9.2 Fal¯ah. ¯ı - Northern Rural Palestinian

  • Northern rural Palestinian speech features a chain shift from /q/ to [k] and /k/ to [ch], with realizations and the word for “now” varying by subregion.Ramallah more often affricates /k/, Qalqilya favors [k] for /q/, and some northernmost speakers retain emphatic [q].Source: §B.9.2 · p. 27

B.9.3 Fal¯ah. ¯ı - Southern Rural Palestinian

  • Southern rural Palestinian often realizes /q/ as [g], contrasting with northern [k] and urban glottal-stop patterns, and its profile reflects contact with neighboring Bedouin dialects.The variety spans rural and tribal communities south of Jerusalem, into the Hebron highlands, Naqab, and Gaza.Source: §B.9.3 · p. 27

B.10 Saudi Micro-dialects

  • Saudi dialect research distinguishes Najdi, Hijazi, Eastern, Northern, and Southern regions; Najdi includes four regional varieties, and some Najdi speech shifts /k/ to [ts].Hijazi itself includes urban and tribal varieties, with common /D/→[z] and /T/→[t] realizations.Source: §B.10 · p. 28, §B.10.1 · p. 28, §B.10.2 · p. 28

B.10.3 Eastern; B.10.4 Southern; B.10.5 Northern

  • Saudi Eastern speech is described as slower with distinct vowel qualities, Southern speech as fast and clipped with some /k/→[sh] forms, and Northern speech as rising in intonation with some /k/→[ts].Southern varieties include Asiri, Tihami, Najrani, and Jizani; Northern dialects are associated with Bedouin heritage and traditional speech patterns.Source: §B.10.3 · p. 28, §B.10.4 · p. 29, §B.10.5 · p. 29

C.1 Extended Corpus Statistics

  • Mauritanian segments show the highest micro-dialect overlap, while Jordan also has strong overlap; Saudi and Palestinian labels overlap much less.In Mauritania, overlap spans Northern, Southern, and Western labels; in Jordan, it is especially strong between Badawi–Falahi and Falahi–Madani.Source: §C.1 · p. 29

D Speaker-Controlled Overlap Analysis

To examine whether apparent label overlap reflects recurring speakers or channels, the analysis clusters speaker embeddings from single-label segments and reports speaker counts for dialects and multi-label intersections.

D Speaker-Controlled Overlap Analysis

  • 19,770 estimated speakers were obtained by clustering embeddings from 24,402 single-label segments, with multi-label segments excluded from the clustering.The exclusion was intended to avoid conflating speaker overlap with label overlap.Source: D Speaker-Controlled Overlap Analysis · p. 29

E Annotation Schema

  • Country-level totals count each segment once, while per-dialect rows count label memberships and can exceed the total when segments have multiple labels.Figure 2 shows the segment annotation schema, including transcription, code-switching, dialect and micro-dialect labels, speaker metadata, source video ID, and timestamp information.Source: Table 4 · p. 30, E Annotation Schema · p. 30

D Speaker-Controlled Overlap Analysis

  • Table 5 presents primary micro-dialect intersections by region, with columns for country, two dialects, and overlap count.Its caption frames the intersections as reflecting the corpus's most fluid linguistic boundaries.Source: D Speaker-Controlled Overlap Analysis · p. 30, Table 5 · p. 30
  • Table 7's multi-label intersection speaker counts are not independent cross-dialect evidence, but are informative alongside the single-label counts.Table 6 reports unique speaker counts per micro-dialect; Table 7 reports counts within the multi-label intersections.Source: D Speaker-Controlled Overlap Analysis · p. 30

E.1 Annotators per Micro-dialect

This block describes the annotator counts and multi-label examples, then documents code-switching, split construction, text preprocessing, and ASR adaptation setup.

E.1 Annotators per Micro-dialect

  • Table 8’s micro-dialect annotator counts are not mutually exclusive because some annotators covered multiple varieties within their country.The table reports per-variety counts alongside country-level totals.Source: §E.1 · p. 30, Table 8 · p. 32

F Representative Multi-label Dialect Segments

  • Table 7 counts estimated distinct speakers within multi-label intersections, omitting intersections with fewer than two estimated speakers.Multi-label assignments can preserve shared features across varieties rather than indicate annotation ambiguity.Source: Table 7 · p. 31, F Representative Multi-label Dialect… · p. 31

G Code-switching Examples

  • Table 10 illustrates insertional or intra-sentential Arabic–English/French switches and larger English phrases or discourse units alongside Arabic.Examples include English words or phrases within an otherwise Arabic utterance, as well as larger English units.Source: G Code-switching Examples · p. 31

H Train/Development/Test Splits

  • The country-wise split assigns each utterance to one partition and balances both country utterance counts and micro-dialect label memberships across train, development, and test.Multi-label utterances remain single examples assigned to one partition while contributing to the balance of each associated label.Source: H Train/Development/Test Splits · pp. 31–32

I Text Preprocessing

  • WER and CER preprocessing removes punctuation, Arabic diacritics, hamzas, and maddas, converts Eastern Arabic numerals to Western numerals, and preserves Latin characters.Latin text is retained because it is valid lexical material in code-switched references and predictions.Source: I Text Preprocessing · p. 32

J ASR Adaptation Details

  • Table 12 reports aggregate WER/CER for all four ASR systems, comparing zero-shot and adapted settings, as summarized in Section 7.2.Source: J ASR Adaptation Details · p. 32

J.1 Checkpoints, Inference, and Fine-Tuning Setup

  • All four systems use released checkpoints for zero-shot evaluation, while adaptation uses only training data, development validation, and no test-time adaptation.No external language model, rescoring, or metadata is used.Source: §J.1 · p. 32, J ASR Adaptation Details · p. 32

J.2 Per-Micro-Dialect Confidence Intervals

This block describes per-micro-dialect confidence intervals and the identification tasks and systems used to evaluate audio-based micro-dialect prediction.

J.2 Per-Micro-Dialect Confidence Intervals

  • Per-micro-dialect WER intervals quantify test-sample uncertainty, not training-seed variance; broad difficulty ordering is stable, but narrow rankings are uncertain.Differences on the smallest subsets fall within overlapping intervals and should not be treated as conclusive.Source: §J.2 · pp. 32–33

K.1 Tasks

  • The identification experiments define country-dialect, global 21-way micro-dialect, and country-conditioned micro-dialect classification tasks.Country-conditioned prediction uses the gold country label; all tasks use the standard CARDAMOM splits.Source: §K.1 · p. 33

K.2 Systems

  • The systems include zero-shot Qwen3-Omni, LoRA fine-tuned Qwen3-Omni trained jointly on all three tasks, and a Whisper-large-v3 classifier with three jointly trained heads.The Qwen3-Omni fine-tuning uses 8.65M trainable adapter parameters.Source: §K.2 · pp. 33–34

K.3 Results

The identification results show that Whisper-v3 outperforms fine-tuned Qwen3-Omni across the three tasks, while fine-tuning substantially improves Qwen3-Omni and raises hierarchical consistency.

K.3 Results

  • 58.46 points on global micro-dialect identification, 47.72 on country identification, and 43.65 on country-conditioned identification are Qwen3-Omni’s accuracy gains after LoRA fine-tuning.Hierarchical consistency rises from 33.67% zero-shot to 95.48% after fine-tuning.Source: §K.3 · p. 34
  • 6.39, 7.91, and 3.83 accuracy points separate Whisper-v3 from fine-tuned Qwen3-Omni on country, global micro-dialect, and country-conditioned identification, respectively.The reported 95% confidence intervals are [5.73, 7.05], [7.07, 8.76], and [3.12, 4.54].Source: §K.3 · p. 34
  • 20.31% to 93.64% is Qwen3-Omni’s country-conditioned accuracy change for Saudi Arabia after LoRA fine-tuning; Mauritania is the only country without improvement over zero-shot.The passage relates Mauritania’s exception to very limited web-scale representation of Hassaniya speech.Source: §K.3 · p. 34

L Dialectness Result

Dialectness scores vary across micro-dialects and multi-label intersections, but high transcript dialectness does not always correspond to high ASR difficulty.

L Dialectness Result

  • 8.5% is the MSA dialectness score, multi-label intersections reach 97.1%, and Mauritanian varieties score 36.9–53.4%.The passage notes that Mauritanian scores are consistent with MSA-proximate broadcast sources.Source: L Dialectness Result · p. 34, Figure 3 · p. 36
  • 97.1% for Egypt-Alexandrian–Port Saidi and 88.9% for Jordan-Badawi–Madani are high dialectness scores for multi-label intersections.The paper suggests these labels capture strongly dialectal utterances shared across neighboring or socially related varieties.Source: L Dialectness Result · p. 34
  • Jordan-Madani and Egypt-Cairene score highly on dialectness yet remain comparatively easy for ASR to recognize.The paper suggests capital varieties may be better represented in pretrained web-scale speech and text data.Source: L Dialectness Result · p. 35
Loading 2609.34481v1…