NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task

NADI series

• NADI began with text-based dialect identification in 2020 and expanded through additional language tasks before pivoting to speech in 2025.Dialect identification has persisted across editions as a means to aggregate varieties, route utterances to dialect-specific models, and audit resource coverage.Source: §1 · p. 2

Benchmark design

• NADI 2026 evaluates ASR, SDID, SLU, SLT, and TTS under deployment-oriented shifts, including low-bandwidth, mixed-dialect, code-switched, out-of-domain, and zero-shot conditions.The edition adds TTS, SLT, and SLU, spanning speech perception, understanding, translation, and generation.Source: §1 · pp. 2–3

Participation and analysis

• The task attracted 21 teams from at least 13 countries, with 48 test-phase submissions and 14 system-description papers; the paper analyzes adaptation, ensembling, and post-processing across systems.The analysis identifies which strategies generalize across subtasks and which do not.Source: §1 · pp. 2–3

Key findings

• 81.19% in-domain dialect-identification accuracy fell to 28.70% out of domain, while the best system reached 56.15%; code-switched ASR recipes finished within 0.12 WER.Arabic-centric foundation models appeared in five of the top six SDID entries and the top entry of one ASR subtask; system combinations helped almost everywhere.Source: §1 · p. 3

Evaluation data

• New evaluation sets cover low-bandwidth Casablanca speech, six-country and 18-city dialect identification with withheld labels, ten-country dialectal TTS, dialect-to-English speech translation, and four zero-shot SLU domains.The speech-translation evaluation data are from WhiteHouse; Casablanca is cited as Talafha et al. (2024).Source: §1 · p. 3

Literature Review

• Low-resource ASR uses transfer learning, self-supervised pretraining, hybrid HMM/self-supervised architectures, in-context learning, and multilingual encoders, addressing Arabic dialects’ limited transcription resources.The cited work contrasts these approaches with the greater transcribed-resource availability for MSA.Source: §2 · p. 3
• The NADI shared tasks have promoted research on dialectal and code-switched Arabic (Talafha et al., 2025).Source: §2 · p. 3
• Arabic dialect ASR research explores dialect-specific fine-tuning and joint dialect identification with recognition, while dialect identification has progressed from i-vector and x-vector to Transformer models.Standard-language ASR often performs poorly on dialectal speech, and weak cross-domain generalization remains a limitation (Sullivan et al., 2023; Abdullah et al., 2025).Source: §2 · p. 4
• Resource constraints challenge dialectal Arabic TTS and SLT, while SLU research has mainly focused on Tunisian Arabic.The review points to Appendix A for a more detailed account of related literature.Source: §2 · p. 4

3 NADI 2026

• The edition comprises five tasks, with additional subtasks for ASR and SLU; all subtasks except TTS use Codabench development and evaluation phases.The section introduces the breakdown of each task and subtask.Source: §3 · p. 4

3.1 Task 1: ASR

• The ASR task comprises robust low-bandwidth, unlabeled mixed-dialect, and Tunisian Arabic–English–French code-switched transcription subtasks.Robust ASR uses audio downsampled to 4 kHz and resampled to 16 kHz; mixed-dialect data covers six countries and over 15 subdialects, while code-switched evaluation includes about 1 hour 50 minutes and 2,123 segments. Rankings use WER and CER without hamza normalization.Source: §3.1 · p. 4

3.2 Task 2: Spoken Dialect Identification

• Spoken dialect identification classifies utterances among dialects, with this edition emphasizing out-of-domain evaluation across six countries and 18 city-level dialects.The possible test countries were undisclosed, except that they would be a subset of ADI-20’s 20 country labels; evaluation uses Accuracy and NIST LRE 2022 Average Cost.Source: §3.2 · p. 4

3.3 Task 3: Dialectal Text-to-Speech

• New to NADI, the dialectal TTS task synthesizes speech from text prompts for ten Arab countries and permits external training data.Evaluation sentences come from Bulbul, manually collected and verified by native speakers; quality uses UTMOSv2 and NISQA, while semantic performance uses WER and CER from a 7B-parameter Omnilingual ASR system.Source: §3.3 · pp. 4–5

3.4 Task 4: Spoken Language Translation

• Spoken language translation systems translate speech from eight Arabic dialects into English under dialectal variation, spontaneous speech, and real-world speaking conditions.The WhiteHouse evaluation data contains 6,788 validation and 6,780 test utterances; submissions are scored with case-sensitive BLEU and chrF, without punctuation.Source: §3.4 · p. 5

3.5 Task 5: Spoken Language Understanding

• Spoken language understanding extracts intent and slot information directly from spontaneous, potentially code-switched Tunisian speech, beyond literal transcription.Evaluation uses a new one-hour set of 989 utterances; slot filling is assessed with CoER for semantic labels and CVER for concept/value pairs.Source: §3.5 · p. 5

4.1 Participating Teams

• 21 teams participated in NADI 2026, with the team list presented in Table 2.The paper reports a final total of 21 participating teams.Source: §4.1 · p. 5, Table 2 · p. 6

4.2 Baselines

• ASR and SLT used off-the-shelf whisper-large-v3, while SDID finetuned ECAPA-TDNN on an ADI20 subset.The baselines used whisper-large-v3 (Radford et al., 2023) and ECAPA-TDNN (Desplanques et al., 2020), respectively.Source: §4.2 · p. 5
• TTS samples used XTTS-v2 with one reference speaker and no special prompting; SLU used Whisper-based baselines for intent recognition and slot filling.Intent recognition used a whisper-small encoder with average pooling and classification layers; slot filling used the full whisper-small system with its tokenizer expanded for semantic tags.Source: §4.2 · pp. 5–6

5.1.1 Subtask 1.1: Robust ASR

• WIFAQ led robust ASR with 43.24% WER and 19.75% CER, beating RESONATE by 7.05% absolute WER; the leaderboard showed no single winning recipe.WIFAQ combined four w2v-BERT 2.0 CTC recognizers, fivegram KenLM decoding, and ROVER; RESONATE adapted one Arabic-specialized encoder-decoder model, while THAKAA used a dialect-aware ensemble. RESONATE and THAKAA differed by only 0.13% WER, despite a larger CER gap.Source: §5.1.1 · p. 6, Table 3 · p. 7

5.1.2 Subtask 1.2: Mixed-Dialect ASR

• SALESTEQ achieved the lowest mixed-dialect ASR WER at 44.22%, while NAMAA COMMUNITY had the best CER at 17.16%.NAMAA COMMUNITY used zero-shot inference with an Arabic-specific Cohere model; ITGAN combined LoRA-adapted Whisper models with ROVER, and KAND CA used full-model fine-tuning with additional data and replay. Table 4 summarizes the task results and system features.Source: §5.1.2 · pp. 6–7, Table 4 · p. 7

5.1.3 Subtask 1.3: Code-switched ASR

• The top three code-switched ASR systems spanned only 0.12% absolute WER, while AR_INDIA was a far outlier at 73.76%.Hamza normalization lowered the top five teams’ WER by roughly 0.30% without changing their ranking. ABJAD AI led on WER, ITGAN had the lowest CER, and all three top systems used model combination; supplementary data included synthetic speech and Tunisian resources.Source: §5.1.3 · p. 7

5.2 Task 2: Spoken Dialect Identification

• THAKAA reached 56.15% accuracy and SALESTEQ achieved the best Cavg at 8.64, improving substantially on the baseline’s 28.70% accuracy and 20.02 Cavg.In-domain validation baseline performance had been 81.19% accuracy and 10.60 Cavg. Multimodal approaches appeared to dominate; THAKAA weighted an Ara-BEST-RQ acoustic model most heavily and text branches at 30%.Source: §5.2 · pp. 7–8

5.3 Task 3: Text-to-Speech

• The single TTS submission achieved aggregate scores of 2.72 UTMOS, 3.57 NISQA-MOS, and 3.21 NISQA-TTS, with 32.48% WER and 11.40% CER.Scores varied by dialect: Morocco had the highest UTMOS (3.23), Sudan the highest NISQA-MOS (4.46) and NISQA-TTS (3.98), and Saudi Arabia the lowest WER (14.28%) and CER (4.87%). The XTTS-V2 baseline outperformed the submission on all five metrics averaged across dialects, but its near-invariant scores and MSA-leaning output led the authors to caution that automatic metrics may miss dialectal identity; they call for human evaluation.Source: §5.3 · p. 8

5.4 Task 4: Spoken Language Translation

• The ABJAD AI submission achieved an overall BLEU score of 14.19 and a chrF score of 34.11, with BLEU ranging from 27.46 for Jordan to 0.39 for the UAE.The team directly finetuned Whisper Large-v3 on the provided training data with additional data augmentation.Source: §5.4 · p. 8

5.5.1 Subtask 5.1: Intent Recognition

• LYNX led intent recognition with 76.86% weighted F1 and 75.73% accuracy, ahead of CIS-OPENSLU at 71.54% and 72.50%, respectively.The ranking was consistent across both metrics. LYNX and CIS-OPENSLU used ASR-to-TF-IDF pipelines without external data, while ASLEMA used a LoRA-finetuned multimodal LLM and synthetic training data; LYNX also filtered unknown-category data using MARBERTv2 trained on MASSIVE.Source: §5.5.1 · pp. 8–9

5.5.2 Subtask 5.2: Slot Filling

• ASLEMA led slot filling with a CoER of 59.53% and a CVER of 94.20%; ZILA followed with a CoER of 72.78% and a CVER of 100.36%.ASLEMA was the only team to provide a system paper, describing LoRA finetuning of Qwen3-Omni-30B on the SLURP-TN training split, synthetic data generation, and Gemini-based filtering.Source: §5.5.2 · p. 9

New model backbones

• Cohere’s Transcribe Arabic model featured in leading systems across several subtasks, while Ara-BEST-RQ strengthened the top Subtask 2 team’s submission.Transcribe Arabic accounted for two top-performing systems in Subtask 1.1, the second-place system in 1.2, the top system in 1.3, four described Subtask 2 systems, and the top Subtask 5.1 system.Source: §6 · p. 9

Multimodal SDID

• Multimodal systems made up three of the top five SDID systems, suggesting the value of combining textual and spoken features.The discussion frames multimodal SDID as an evolution of earlier use of ASR transcripts and phone-based features.Source: §6 · p. 9

Ensembling

• Combined systems accounted for 10 of 41 valid test submissions, but their added training and inference compute motivates examining efficiency across subtasks.The passage notes model efficiency as a potential direction for low-compute scenarios.Source: §6 · p. 9

Is SDID Necessary?

• No teams integrated SDID into Mixed-Dialect ASR; future work could compare dialect-specific ASR pipelines with general models such as Transcribe Arabic or Whisper.The discussion suggests that general multidialect models and limited data for training dialect-specific ASR may help explain this choice.Source: §6 · p. 9

Lessons learned

• Future NADI editions should redesign TTS evaluation, make SDID less purely zero-shot through dialect coverage and discrete domains, and better support contrastive submissions.The discussion proposes evaluation platforms that better display variations among systems.Source: §6 · p. 9

• NADI 2026 expanded to eight subtasks, including Robust, Mixed-dialect, and Code-switched ASR and newly introduced TTS, SLT, and SLU.It drew 21 participating teams, tied for the series’ joint-largest number, and 14 system-description papers, the second-largest number.Source: §7 · p. 9
• New pretrained Arabic ASR models show strong DA performance, multimodal approaches matter for SDID, and ensemble-based approaches dominate.Source: §7 · p. 9
• Future work will examine low-compute performance across tasks, investigate SDID systems within ASR pipelines, and possibly add new dialects.Source: §7 · p. 9

Limitations

• NADI acknowledges limited dialectal reach—TTS covers ten dialects versus 28 varieties identified by ISO—and warns speech technologies can enable surveillance, fraud, or deception.The task made slight progress over the previous edition by adding code-switching and sub-dialectal variation.Source: Limitations · p. 10
• TTS evaluation was constrained by low participation and the absence of human evaluation, limiting conclusions about the subtask.Participation may have been discouraged by expectations about training data, limited dissemination, and the lack of a submission platform; automatic metrics lacked human validation across dialects.Source: Limitations · p. 10
• Collapsing hamza variants lowers the top five teams’ WER by 0.29 points, about 2% of measured word error, without changing leaderboard order.Because hamza is only one axis of orthographic variability, the 2% reduction is a lower bound and absolute error rates are less robust; multi-reference WER is a possible future approach.Source: Limitations · p. 10

Appendix Table Of Contents

• The appendices provide a comprehensive literature review, additional task information and result tables, and complete submission descriptions.Source: Appendix Table Of Contents · p. 19

A.1 Automatic Speech Recognition; A.2 Spoken Language and Dialect ID

• Arabic ASR research draws on multilingual and self-supervised methods, while spoken language identification remains vulnerable to cross-domain performance.Arabic varieties have limited transcribed data compared with MSA (Ali et al., 2017), and NADI 2026 targets code-switching across dialect, MSA, and other languages. Spoken language identification models may overfit training domains (Sullivan et al., 2023; Abdullah et al., 2025; Dey et al., 2023).Source: §A.1 · p. 19, §A.2 · p. 20

A.3 Text-to-Speech.

• Arabic TTS spans multilingual, MSA-specific, and dialectal systems, but each faces limitations in producing authentic colloquial varieties.Multilingual systems lack explicit regional-dialect conditioning, MSA systems do not generalize well to colloquial speech, and dialectal TTS remains under-resourced because paired speech-text corpora are scarce.Source: §A.3 · p. 20

A.4 Spoken Language Translation

• Arabic spoken language translation remains under-resourced, and broad multilingual models may fail to preserve dialect-specific characteristics.Resources include the multi-dialectal WhiteHouse corpus, Tunisian–English resources including 160 hours for the IWSLT task and TEDxTN’s approximately 25 hours, and Egyptian ArzEnST’s approximately 12 hours and 6,216 sentences; open challenges include dialect variation, code-switching, spontaneous speech, and limited parallel supervision.Source: §A.4 · pp. 20–21

A.5 Spoken Language Understanding

• Arabic SLU remains comparatively under-resourced, particularly for dialectal varieties.Speech-MASSIVE’s Arabic subset has only 115 training utterances; in Tunisian SLU, w2v-BERT 2.0 achieved the best overall performance, semantically enhanced SENSE improved performance further, and SLURP-TN added speech-level intent and slot annotations.Source: §A.5 · pp. 21–22

B.1 TTS Data Overview

• The TTS text was normalized to NFKC, with diacritics, Tatweel, emojis, parenthetical metadata, HTML tags, and non-Arabic characters removed.Source: §B.1 · p. 22
• The pipeline reduced repeated-character runs, removed online laughter and non-standard punctuation, and excluded sentences with English words, URLs, user tags, long gibberish words, or fewer than five words.Source: §B.1 · p. 22
• The benchmark contains 4,001 samples across 10 target dialects, sourced from public datasets, web data, and private team resources.Source: §B.1 · p. 22
• Country-wise evaluation-sample distribution is detailed in Table B.1.Source: §B.1 · p. 22

B.2 SLT Data Overview

• The SLT subtask’s duration and word counts are overviewed in Table B.2.Source: §B.2 · p. 22

B.3 SLU Data Overview

• The evaluation material comprises 989 utterances from 12 speakers across five regional backgrounds, with an exactly balanced speaker gender distribution.Male speakers contributed 523 utterances (52.88%) and female speakers 466 (47.12%); Greater Tunis, Medenine, Kebili, Kairouan, and Sfax are represented.Source: §B.3 · p. 22
• The SLU evaluation set combines three returning domains with four unseen domains, testing generalization to new speakers and utterances as well as transfer to new application areas.Emails, Weather, and General are returning domains; Calendar, Question Answering, Music, and Cooking are absent from the provided SLURP-TN training and development sets.Source: §B.3 · p. 22

C Baseline Systems

• For ASR, the baseline uses multilingual whisper-Large-v3, which was trained without distinguishing Arabic varieties and may favor MSA through subtitle training.The passage describes the model as trained on noisylabeled internet audio and notes the possible MSA preference as an assumption.Source: C Baseline Systems · p. 22
• For spoken dialect identification, the baseline adapts an ECAPA-TDNN system originally trained on LinguaVox107 by fine-tuning it on 10 hours per dialect of ADI-20.The approach is based on the NADI 2025 ADI subtask baseline (Talafha et al., 2025), with modifications; the original system was trained by Valk and Alumäe (2021).Source: C Baseline Systems · p. 22
• For TTS, the baseline uses XTTS-v2 with “Suad Qasim” as the single reference speaker across dialects.XTTS is described as a Coqui cross-lingual TTS model that synthesizes speech from text using a short reference clip to reproduce target-speaker voice characteristics; it was added for contextualization after the original evaluation period.Source: C Baseline Systems · pp. 22–24
• For spoken language translation, the baseline uses whisper-large-v3 to translate the considered Arabic dialects into English.Whisper is described as an end-to-end multitask speech model that natively performs language identification, transcription, and speech translation into English.Source: C Baseline Systems · p. 24
• For SLU, whisper-small underlies both baselines: intent recognition adds masked average pooling and a linear classifier, while slot filling fine-tunes the model to label slots in its transcription output.Both baselines use SLURP-TN training data.Source: C Baseline Systems · p. 24

D Additional Results

• Tables of results are provided for the remaining subtasks: code-switched ASR (Table D.1), SDID (D.2), TTS (D.3), SLT (D.4), and intent classification (D.5).Source: D Additional Results · p. 24

E.1 Subtask 1.1: Robust Dialectal ASR

• For robust dialectal ASR, Wifaq combined four w2v-bert-2.0 CTC recognizers with beam search and word-level ROVER, while Resonate fine-tuned Cohere with denoising, voice conversion, and audio filtering.Thakaa instead selected ROVER committees from seven candidates, including Cohere, FastConformer, Whisper, and Qwen3-ASR systems.Source: §E.1 · pp. 24–25

E.2 Subtask 1.2: Mixed dialect ASR

• For mixed-dialect ASR, Namaa Community used Cohere zero-shot without task-specific fine-tuning, whereas Itgan combined five Whisper recognizers with ROVER.KAND CA used sequential full-model fine-tuning with replayed Task 1.1 and Casablanca clips, then a single model with chunked decoding.Source: §E.2 · p. 27

E.3 Subtask 1.3: Code-switched Tunisian ASR

• For code-switched Tunisian ASR, Abjad AI fine-tuned Cohere using synthesized Tunisian speech, French and English speech, and noise, then averaged five checkpoints.Its training preserved FP32 master weights under bfloat16 autocasting, and its final system averaged checkpoint parameters in FP32.Source: §E.3 · p. 27
• Itgan’s code-switched ASR used Whisper LoRA adapters and SpecAugment; increasing dropout to 0.15 shrank the generalization gap by a factor of 3.5.Itgan combined adapters by weight-space concatenation, then used their model in a three-system word-level ROVER vote.Source: §E.3 · pp. 27–28
• For code-switched Tunisian ASR, Thakaa used a ten-system ROVER ensemble, while KAND CA reported a fresh two-epoch schedule over 70,595 clips as its decisive training step.Thakaa added about 6,900 labeled Tunisian utterances; KAND CA found that warm continuation and further epochs regressed.Source: §E.3 · p. 28

E.4 Subtask 2: Spoken Dialect ID

• In SDID, Rank 1 Thakaa combined an Ara-BEST-RQ audio component with two text pipelines, each pairing Cohere and Whisper transcription branches.Its training used 265 hours from eight ADI and ASR datasets; Rank 3 Lynx used transcript TF-IDF features and a Cohere acoustic feature extractor with an LSTM classifier.Source: §E.4 · p. 28
• In SDID, AI Elites fine-tuned Cohere for speech-only classification and used Casablanca and MADIS-5 for out-of-domain checkpoint selection.KAND CA combined audio classifiers on original and speech-enhanced audio with a text-route classifier, while Resonate used WavLM with attentive pooling and ADI17/ADI20 data.Source: §E.4 · pp. 28–29

E.5 Task 4 SLT

• Abjad AI (Rank 2) fine-tuned whisper-large-v3 directly on augmented, upsampled Task 4 data, which reached approximately 2.5 hours.Augmentations included Gaussian noise, time stretching, pitch shifting, and temporal shifting; direct fine-tuning outperformed intermediate fine-tuning on development data.Source: §E.5 · p. 29
• Abjad AI reran its system after discovering training-data overlap with the development set, minimizing that overlap in the reported results.Source: §E.5 · p. 29

E.6 Task 5.1 Intent Classification

• Lynx (Rank 1) transcribed speech with cohere-transcribe-arabic-07-2026, then classified intents using fused word- and character-level TF-IDF n-grams and one-rest logistic regression.MARBERTv2, trained for intent classification on MASSIVE, was used to label low-confidence utterances as unknown.Source: §E.6 · p. 30
• CIS-OpenSLU (Rank 2) used Faster-Whisper transcripts and a linear SVM on word- and character-level TF-IDF features, with low-confidence predictions assigned to an unknown class.Training and validation data were combined into 3,272 labeled utterances, without external labeled data or synthetic augmentation.Source: §E.6 · p. 30
• Aslema (Rank 4) used Qwen3-Omni-30B-A3B-Instruct with joint LoRA fine-tuning for intent recognition and slot filling, augmenting roughly 2.7K official examples with synthetic speech–text data.Synthetic Tunisian Arabic examples were generated with Gemini, quality-filtered, and synthesized using VoxCPM, yielding approximately 23K additional pairs.Source: §E.6 · p. 30

E.7 Task 5.2 Slot Filling

• Aslema (Rank 1) jointly fine-tuned Qwen3-Omni-30B-A3B-Instruct on official and synthetic data to predict slot labels and values directly from speech (Shafiq et al., 2026).The approximately 2.7K official examples were supplemented with approximately 23K synthetic speech–text pairs generated from Gemini-annotated Tunisian Arabic utterances and synthesized using VoxCPM (Zhou et al., 2026).Source: §E.7 · p. 30