Arabic Dialect Identification

• Arabic Dialect Identification is the process of classifying text and speech into distinct dialect labels, employing methods from feature engineering to deep neural networks.
• It leverages diverse corpora—from tweets to broadcast speech—to achieve granular classification at country, regional, and city levels while addressing cross-domain challenges.
• Recent research critiques the single-label formulation and promotes multi-label approaches to more accurately capture overlapping dialect characteristics and enhance downstream applications.

Searching arXiv for recent and foundational work on Arabic Dialect Identification across text and speech.

Arabic Dialect Identification (ADI) is the task of automatically assigning an Arabic utterance or text fragment to a dialectal label, typically from a predefined inventory that may include Modern Standard Arabic (MSA), broad regional groupings, country-level dialects, city-level dialects, or, more recently, multiple simultaneously valid dialect labels. In practice, ADI spans both written and spoken modalities, and its difficulty derives from the close relatedness of Arabic varieties, pervasive code-switching with MSA, orthographic non-standardization in informal writing, inter-dialect overlap, and strong domain effects across broadcast speech, social media, and conversational content. Research on ADI has progressed through corpus creation, feature-engineered classifiers, deep acoustic and textual encoders, and growing scrutiny of the single-label formulation itself (Abdelali et al., 2020, Ali et al., 2015, Althobaiti, 2020, Keleg et al., 2023).

1. Problem Formulation and Dialect Granularity

ADI has been formulated at multiple levels of granularity. In broadcast speech, early work focused on five-way classification among Egyptian, Gulf, Levantine, North African, and MSA (Ali et al., 2015). Challenge settings later expanded to 17 spoken dialects in ADI-17 and to 19 dialects plus MSA in ADI-20 (Lin et al., 2020, Elleuch et al., 13 Nov 2025). In written social media, country-level identification has been studied on 18 dialects with QADI and 21 dialects in NADI Shared Task 1 (Abdelali et al., 2020, Beltagy et al., 2020). Other work uses sentence-level city dialect identification, as in MADAR-based systems operating over Beirut, Cairo, Doha, Rabat, and Tunis, alongside MSA (Alhafni et al., 2024).

The standard formalization in written ADI maps a text $T$ to a dialect label $D$, where $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$ (Althobaiti, 2020). In supervised multiclass settings, evaluation commonly relies on accuracy or macro-averaged $F_1$, with macro-$F_1$ defined as

$Macro\text{-}F1 = \frac{1}{N} \sum_{i=1}^N F1_i$

for $N$ classes (Abdelali et al., 2020, Beltagy et al., 2020).

A central development in recent research is the critique of the single-label assumption. Manual validation of false positives on QADI showed that out of 490 validated false positives, 325, approximately 66%, were also valid in the predicted dialect, implying that many nominal errors are artifacts of incomplete annotation rather than true mistakes (Keleg et al., 2023). Related multi-label analysis over 11 country-level dialects found that only 44% of sentences are valid in a single regional dialect, while 56% are valid in multiple regional dialects; at country level, only 25% are single-label (Keleg et al., 27 May 2025). This suggests that the classical closed-set single-label formulation is appropriate only for some ADI settings, especially where labels reflect speaker origin rather than sentence-level acceptability.

2. Corpora, Benchmarks, and Annotation Regimes

Dataset construction has strongly shaped the trajectory of ADI research. For written ADI, QADI introduced a large balanced tweet corpus covering 18 Arab countries, with 540,000 tweets from 2,525 users and approximately 30,000 tweets per country (Abdelali et al., 2020). It was collected automatically from 25 million Arabic-language tweets gathered through Twitter’s streaming API in March–April 2018. User profiles were filtered by explicit self-declaration of country using a gazetteer, likely MSA-dominant accounts were removed using a fastText dialect–MSA classifier, and accounts dominated by vulgar or offensive language were excluded (Abdelali et al., 2020). Intrinsic evaluation found tweet labels to be 91.5% accurate on average, with 87% inter-annotator agreement in countries sampled by two annotators (Abdelali et al., 2020).

NADI Shared Task 1 used 21,000 training tweets, 4,957 development tweets, and 5,000 test tweets for 21 country-level dialects, along with 10 million unlabeled tweet IDs for semi-supervised methods (Beltagy et al., 2020). MADAR-based sentence-level resources support city-level ADI and downstream dialect-aware normalization (Alhafni et al., 2024). The written-text survey documents a broad ecology of corpora, including AOC, MADAR, PADIC, Gumar, LICSD, and VarDial datasets, spanning token-, sentence-, tweet-, and document-level annotations (Althobaiti, 2020).

In speech, the 2015 broadcast corpus released train and test data for five-way ADI and became a standard resource for MGB-style evaluation (Ali et al., 2015). The MGB-3 challenge used 13,825 training utterances, 1,524 development utterances, and 1,492 test utterances, with pronounced domain mismatch between training and the matched development/test channels (Shon et al., 2017). ADI-17 later expanded coverage to 17 dialects from YouTube programs, with 3,000 hours of training data and approximately 57 hours verified and balanced for evaluation (Lin et al., 2020, Miao et al., 2019). ADI-20 extends ADI-17 to cover all Arabic-speaking countries’ dialects plus MSA, comprising 3,556 hours from 19 Arabic dialects in addition to MSA (Elleuch et al., 13 Nov 2025).

The following table summarizes several benchmark regimes that recur in the literature.

Resource / benchmark	Modality	Label space
QADI	Tweets	18 country-level dialects
NADI Shared Task 1	Tweets	21 country-level dialects
MGB-3 / ADI-5	Broadcast speech	EGY, GLF, LEV, NOR, MSA
ADI-17	Speech	17 country-level dialects
ADI-20	Speech	19 dialects + MSA
MADAR-based DID	Sentences	5 city dialects + MSA

Annotation strategy is itself controversial. QADI relies on user-level country self-identification and manual tweet validation (Abdelali et al., 2020). Much speech work labels by program origin or speaker identity (Ali et al., 2015, Sullivan et al., 2023). Recent multi-label work argues that such single labels encode provenance but not necessarily sentence-level validity, and that future datasets should elicit acceptability judgments from native speakers of each dialect separately (Keleg et al., 2023, Keleg et al., 27 May 2025).

3. Classical Methods and Feature-Engineered ADI

The earliest strong ADI systems relied on manually designed feature spaces and comparatively shallow classifiers. In Arabic broadcast speech, phonetic features, lexical features derived from ASR output, and acoustic i-vectors were combined with multi-class SVMs (Ali et al., 2015). In that work, five-way dialect discrimination achieved 52% accuracy, while the best fused system, combining senone-based and i-vector representations at score level, reached 60.2% accuracy (Ali et al., 2015). The same study also reported 100% accuracy for Arabic-versus-English identification and 100% accuracy for MSA-versus-dialectal Arabic (Ali et al., 2015). Confusion was concentrated between MSA and dialects, reflecting code-switching in broadcast speech (Ali et al., 2015).

In text ADI, the dominant classical feature sets have included word and character $n$-grams, skip-grams, dictionary-based cues, language-model scores, and TF-IDF representations (Althobaiti, 2020). The survey of written ADI reports that carefully engineered traditional models, especially Naive Bayes, SVMs, and ensembles, were often stronger than early deep learning systems on available corpora (Althobaiti, 2020). This suggests that lexical sparsity patterns and subword orthographic regularities remained highly informative under limited-resource conditions.

Kernel methods formed another influential branch. The UnibucKernel system, which ranked first in the 2018 VarDial ADI closed shared task, combined multiple string kernels on character $p$-grams from speech and phonetic transcripts with an RBF kernel over neural audio embeddings (Butnaru et al., 2018). Kernel Ridge Regression outperformed Kernel Discriminant Analysis in development experiments, and the top submitted system reached a macro-$F_1$ of 58.92%, ahead of the second-best score of 57.59%; post-competition use of improved audio embeddings raised macro-$D$0 to 62.28% (Butnaru et al., 2018). This line of work showed that shallow multiple-kernel learning remained highly competitive when acoustic and transcript-derived evidence were fused.

These classical systems established several persistent observations: acoustic and linguistic cues are complementary; confusion is most frequent among geographically proximate dialects; and dialect identification is substantially harder than coarse language identification because the classes are closely related (Ali et al., 2015, Shon et al., 2017, Butnaru et al., 2018).

4. Neural Architectures for Spoken ADI

Neural spoken ADI evolved from i-vector post-processing to end-to-end encoders and self-supervised representations. A landmark transitional system is the MIT-QCRI submission to MGB-3, which combined acoustic and linguistic features with Siamese neural networks, recursive whitening, interpolated i-vector dialect modeling, and linear fusion (Shon et al., 2017). The Siamese network learned a transformed i-vector space by minimizing

$D$1

where $D$2 indicated same or different dialect, and $D$3 was cosine distance after transformation (Shon et al., 2017). The best primary system achieved 75.0% accuracy on the official 10-hour test set (Shon et al., 2017). UTD-CRSS, in the same challenge, fused multiple front ends including MFCC, BNF, and unsupervised bottleneck features with Gaussian back-end and semi-supervised GAN classifiers, reaching 76.94% accuracy in the submitted contrastive system and 79.76% after post-evaluation correction (Bulut et al., 2017).

For ADI-17, deep end-to-end architectures became dominant. A transformer-based system operating on 80-dimensional log-Mel filterbank features used self-attention with downsampling, 4 encoder layers, and score-level fusion with a CNN baseline (Lin et al., 2020). Downsampling increased transformer test accuracy from 76.0% to 82.5%, and fusion with CNN raised overall accuracy to 86.3% on ADI17 (Lin et al., 2020). A CLSTM architecture combining convolutional front-end layers, an LSTM layer, and TDNN back-end improved over a DNN x-vector baseline; with time-scale modification and traditional augmentation, it achieved 93.06% test accuracy and 2.09 test EER, ranking second in the MGB-5 ADI challenge (Miao et al., 2019). A separate prosodic line used intonation pattern embeddings mined from pitch contours, achieving 81.25% test accuracy and 81.56% weighted $D$4 on VarDial 17 using only intonation-based representations (Alvarez et al., 2020).

More recent systems rely on self-supervised speech models and discriminative back-ends. Fusion of ResNet and ECAPA-TDNN architectures with MFCC and UniSpeech-SAT features achieved 84.7% on ADI-5 and 96.9% on ADI-17, outperforming previously published results on both datasets (Kulkarni et al., 2023). Parameter-efficient adaptation of Whisper explored residual adapters and input reprogramming; encoder fine-tuning reached 95.01% test accuracy on ADI-17, while Adapter-256 achieved 93.15% using only 2.5% of trainable parameters (Radhakrishnan et al., 2023). ADI-20 then benchmarked ECAPA-TDNN and Whisper encoder variants over a 20-way dialect space; Whisper-large with layer freezing and augmentation reached 95.82% weighted $D$5 on ADI-17 test and 94.83% on ADI-20 test (Elleuch et al., 13 Nov 2025).

A notable new direction is CTC-DID, which reframes spoken dialect identification as limited-vocabulary ASR using repeated dialect tags and CTC loss: $D$6 (Farooq et al., 18 Jan 2026). On limited data, the SSL-based CTC-DID model outperformed ECAPA-TDNN and Whisper-base, achieved 86.98 weighted $D$7 with fine-tuned SSL on ADI-17 (10h regime), and was more robust to short utterances and streaming inference (Farooq et al., 18 Jan 2026). This suggests that frame-level dialect token emission may alleviate the long-context dependence of utterance-level pooling systems.

5. Neural Architectures for Written ADI

Written ADI moved from feature engineering to pretrained LLMs, but progress has remained sensitive to preprocessing, class imbalance, and domain adaptation. In QADI-based tweet classification across 18 country labels, surface features, static embeddings, and transformer encoders were compared. AraBERT achieved the best macro-averaged $D$8 of 60.6%, ahead of mBERT at 58.9%, SVM with combined character–word $D$9-gram features at 57.2%, and Mazajak static embeddings at 39.8% (Abdelali et al., 2020). The same work reported that training/testing on MADAR yielded much lower macro-$$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$0, approximately 29%, for tweet-level ADI, highlighting the mismatch between sentence-domain benchmarks and social-media dialect identification (Abdelali et al., 2020).

In NADI Shared Task 1, a semi-supervised BERT-based system used AraBERT with cleaning, Farasa segmentation, upsampling, and domain-adaptive masked LLM fine-tuning on 2 million unlabeled tweets (Beltagy et al., 2020). On the development set, performance improved from 18.6 macro-$$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$1 for the baseline AraBERT model to 24.43 after MLM domain-adaptive fine-tuning, and the best test score was 23.09 macro-$$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$2, ranking fourth in the shared task (Beltagy et al., 2020). The gains were incremental but systematic: cleaning, augmentation, and domain-adaptive pretraining each contributed measurable improvement (Beltagy et al., 2020).

The written-ADI survey documents that deep architectures in text, including CNNs, LSTMs, BiLSTMs, and hybrid models, often lagged behind strong classical baselines on modest-sized corpora (Althobaiti, 2020). A plausible implication is that fine-grained dialect distinctions in text require either very large in-domain corpora or annotation schemes that better reflect dialect overlap than classical single-label datasets permit.

This tension has become explicit in multi-label written ADI. LAHJATBERT constructs multi-label pseudo-annotations for single-label training data using GPT-4o, binary dialect acceptability classifiers, and ALDi-guided aggregation, then trains a MARBERT-based multi-label classifier with curriculum learning (Mekky et al., 12 Feb 2026). On the MLADI leaderboard, the best-performing system reached a macro $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$3 of 0.69, compared to 0.55 for the strongest previously reported system (Mekky et al., 12 Feb 2026). The main technical claim is that repurposing single-label data for multi-label ADI is difficult chiefly because negative samples are unreliable, since many supposed negatives are acceptable in multiple dialects (Mekky et al., 12 Feb 2026).

6. Domain Shift, Robustness, and Evaluation Pathologies

Domain mismatch is a recurring difficulty in both speech and text ADI. MGB-3 explicitly involved mismatch between training and matched development/test domains, motivating recursive whitening and interpolated dialect modeling in i-vector space (Shon et al., 2017). In tweet ADI, differences between curated corpora such as MADAR and in-the-wild Twitter data substantially alter achievable performance (Abdelali et al., 2020). For written models using AraBERT, domain-adaptive MLM on unlabeled tweets provided the largest development gain among tested interventions (Beltagy et al., 2020).

A dedicated robustness study using self-supervised spoken ADI models confirmed that domain shift is a major challenge (Sullivan et al., 2023). On ADI-17, HuBERT-17 achieved 92.12 macro-$$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$4 in-domain, but transfer to ADI-5 dropped to 80.36, and heavily shifted YouTube Dramas conditions could reduce performance below 10 macro-$$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$5 without adaptation (Sullivan et al., 2023). Self-training improved domain-shifted performance, but the authors conclude that it may be insufficient for realistic conditions (Sullivan et al., 2023). Their human analysis also found that surrogate country labels matched actual utterance dialect in only approximately 25% of examined cases (Sullivan et al., 2023).

Voice conversion has emerged as a direct response to cross-domain degradation and speaker bias. Training MMS wav2vec2 ADI models on both natural and voice-converted speech increased cross-domain average accuracy from 60.22% to 80.73% on the MADIS-5 benchmark, a relative improvement of +34.07%, while also improving in-domain accuracy from 75.94% to 85.32% (Abdullah et al., 30 May 2025). A controlled biased-versus-unbiased VC experiment showed that when target speakers were tied to dialect labels, performance collapsed toward chance, supporting the claim that conventional ADI datasets allow models to exploit speaker–dialect correlations (Abdullah et al., 30 May 2025).

Robustness questions also intersect with MSA. Training spoken models with mismatched MSA data can hurt transfer because models may rely on channel cues rather than linguistic content (Sullivan et al., 2023). In ADI-20, MSA was added explicitly as a class with 68 hours, and MSA remained among the more confusable categories, especially with North African dialects (Elleuch et al., 13 Nov 2025). More broadly, recurring confusions between geographically close dialects, or between MSA and mixed dialectal segments, should not always be interpreted as simple model error; several studies tie these patterns to genuine linguistic overlap, code-switching, or annotation incompleteness (Ali et al., 2015, Abdelali et al., 2020, Keleg et al., 2023).

7. Linguistic Overlap, Multi-Label ADI, and Downstream Use

The strongest conceptual challenge to standard ADI is that dialect membership may be non-exclusive at the sentence level. Manual analysis of MarBERT predictions on the QADI test set showed that approximately 66% of validated errors were not true errors (Keleg et al., 2023). The same work formalized an expected maximal single-label accuracy: $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$6 where $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$7 denotes the percentage of sentences valid in $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$8 dialects (Keleg et al., 2023). This provides a dataset-dependent upper bound under incomplete single-label annotation.

A broader empirical re-evaluation of assumptions about Arabic dialects found that 56% of sampled dialectal sentences were valid in more than one regional dialect and 75% in more than one country dialect (Keleg et al., 27 May 2025). Sentence length correlated only weakly with dialect ambiguity, with $$D \in \{\text{MSA}, \text{Dialect}_1, \ldots, \text{Dialect}_N\}$$9, whereas ALDi correlated more strongly, with $F_1$0 (Keleg et al., 27 May 2025). The same study found low recall for curated lists of distinctive dialectal lexical cues, indicating that lexical cue bootstrapping alone is a noisy mechanism for dataset creation (Keleg et al., 27 May 2025). These results reinforce the argument that ADI should often be modeled as a multi-label acceptability problem rather than exclusive provenance classification.

Multi-label thinking also connects ADI to downstream applications. In CODAfication, sentence-level DID is used to condition normalization as

$F_1$1

where $F_1$2 is the raw dialectal input, $F_1$3 the normalized CODA form, and $F_1$4 the dialect label (Alhafni et al., 2024). Using predicted dialect information via control tokens improved performance across all dialects: AraT5 baseline achieved $F_1$5, while AraT5 with city token reached $F_1$6, and AraT5 with dialectal phrase token reached $F_1$7, with statistically significant gains at $F_1$8 (Alhafni et al., 2024). This shows that even imperfect ADI can be operationally useful when incorporated as conditioning rather than as an end in itself.

The written-ADI survey already characterized ADI as a first step for machine translation, multilingual text-to-speech synthesis, and cross-language text generation (Althobaiti, 2020). QADI explicitly anticipates utility for machine translation, POS tagging, author profiling, and user geolocation (Abdelali et al., 2020). In speech pipelines, robust ADI is described as necessary for large-scale data collection for ASR across Arabic varieties (Sullivan et al., 2023, Abdullah et al., 30 May 2025). This suggests that the significance of ADI lies at least as much in data routing, normalization, and model specialization as in standalone classification accuracy.

In current research, ADI is therefore best understood not as a single settled task but as a family of related inference problems whose appropriate formulation depends on modality, annotation target, and downstream objective. Provenance classification remains useful in speaker- or user-level settings. Sentence-level acceptability increasingly motivates multi-label formulations. Cross-domain deployment demands robustness to speaker bias and channel shift. Across all of these settings, the field has moved from coarse regional broadcast classification to country-level and city-level modeling, from feature engineering to pretrained encoders, and from treating dialect labels as fixed ground truth to questioning what those labels actually mean (Ali et al., 2015, Shon et al., 2017, Abdelali et al., 2020, Keleg et al., 2023, Abdullah et al., 30 May 2025, Mekky et al., 12 Feb 2026).