Paper deep dive
Aslema at NADI 2026: Data Augmentation for Intent Recognition and Slot Filling
Tajwaar Shafiq, Hunzalah Hassan Bhatti, Firoj Alam, Shammur Absar Chowdhury
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 8/25/2026, 6:23:53 AM
Summary
This paper presents Aslema, a system for the NADI 2026 Shared Task 5, focusing on intent recognition and slot filling for Tunisian Derja spoken language understanding. The authors evaluate four omni LLMs (Qwen2.5-Omni, Qwen3-Omni, Gemma-4) in zero-shot and fine-tuned settings, finding that LoRA fine-tuning significantly outperforms zero-shot inference. They introduce a synthetic data augmentation pipeline using LLMs to generate culturally grounded utterances and VoxCPM for voice cloning. The final system, based on Qwen3-Omni-30B fine-tuned on a mix of real and synthetic data, achieved 1st place in slot filling and 4th in intent recognition.
Entities (15)
Relation Signals (12)
Fine-tuning → outperforms → Zero-shot inference
confidence 99% · Our results show that fine-tuning consistently outperforms zero-shot inference
Aslema → participatedin → NADI 2026 Shared Task 5
confidence 99% · We present Aslema, our system for NADI 2026 Shared Task 5
Aslema → targetslanguage → Tunisian Derja
confidence 99% · Tunisian Dialect (Derja) is low-resource... We participate in both subtasks
Aslema → usesdataset → SLURP-TN
confidence 99% · using SLURP-TN (Elleuch et al., 2026)
Aslema → usesmodel → Qwen3-Omni-30B
confidence 99% · Our final submitted system, based on Qwen3-Omni-30B
Aslema → usestechnique → Data Augmentation
confidence 99% · We further explore synthetic data augmentation
Aslema → usestechnique → LoRA
confidence 99% · LoRA-based fine-tuning (Hu et al., 2022)
Qwen3-Omni-30B → achievedrank → 4th in intent recognition
confidence 95% · and 4th among 8 teams in intent recognition (66.1% accuracy)
Qwen3-Omni-30B → achievedrank →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present Aslema, our system for NADI 2026 Shared Task 5, which consists of two subtasks: intent recognition and slot filling. We evaluate four omni LLMs in a zero-shot setting and compare them with fine-tuned models. Our results show that fine-tuning consistently outperforms zero-shot inference. We further explore synthetic data augmentation by using an LLM to generate culturally grounded Tunisian Derja utterances, followed by voice cloning to generate synthetic speech. Incorporating this synthetic data improves performance on both tasks. Our final submitted system, based on Qwen3-Omni-30B and trained with a mixture of original and synthetic data, achieves 86.8% intent accuracy and 34.7 WER on the devtest split. On the official test set it ranks 1st in slot filling (59.5 CoER) and 4th among 8 teams in intent recognition (66.1% accuracy). We release our experimental scripts and will soon share the synthetic dataset to support further research in this area.
Tags
Links
- Source: https://arxiv.org/abs/2608.18689v2
- Canonical: https://arxiv.org/abs/2608.18689v2
Trouble viewing inline? Open PDF directly →
Full Text
43,538 characters extracted from source content.
Expand or collapse full text
Aslema at NADI 2026: Data Augmentation for Intent Recognition and Slot Filling Tajwaar Shafiq * , Hunzalah Hassan Bhatti, Firoj Alam, Shammur Absar Chowdhury Qatar Computing Research Institute, HBKU, Qatar tajwaar.shafiq@alumni.utoronto.ca, hubh90945, fialam, shchowdhury@hbku.edu.qa Abstract We present Aslema, our system for NADI 2026 Shared Task 5, which consists of two subtasks: intent recognition and slot filling. We evalu- ate four omni LLMs in a zero-shot setting and compare them with fine-tuned models. Our results show that fine-tuning consistently out- performs zero-shot inference. We further ex- plore synthetic data augmentation by using an LLM to generate culturally grounded Tunisian Derja utterances, followed by voice cloning to generate synthetic speech. Incorporating this synthetic data improves performance on both tasks. Our final submitted system, based on Qwen3-Omni-30B and trained with a mixture of original and synthetic data, achieves 86.8% intent accuracy and 34.7 WER on the devtest split. On the official test set it ranks 1st in slot filling (59.5 CoER) and 4th among 8 teams in intent recognition (66.1% accuracy). We release our experimental scripts 1 and will soon share the synthetic dataset to support further research in this area. 1 Introduction Spoken dialogue interfaces are increasingly driven by large language models (LLMs), with recent au- dio LLMs processing speech directly and com- bining speech recognition with language under- standing (Chu et al., 2023; Tang et al., 2024; Xu et al., 2025a). However, their effectiveness for dialectal Arabic remains limited (Abdelali et al., 2024; Al-Khalifa et al., 2025; Bhatti and Alam, 2026). Tunisian Dialect (Derja) is low-resource, heavily code-switched with French and English, and substantially different from the Modern Stan- dard Arabic (MSA) that dominates Arabic training corpora (Talafha et al., 2024). The NADI shared- task series has increasingly focused on such di- alectal settings (Abdul-Mageed et al., 2024; Ta- * The contribution was made while the author was intern- ing at the Qatar Computing Research Institute. 1 https://github.com/hunzed/aslema_nadi2026 lafha et al., 2025). Its 2026 edition (Sullivan et al., 2026) introduces Shared Task 5 on end-to-end Spo- ken Language Understanding (SLU) using SLURP- TN (Elleuch et al., 2026), a Tunisian re-recording of the SLURP benchmark (Bastianelli et al., 2020), covering intent recognition and slot filling. We participate in both subtasks and study the ef- fectiveness of audio LLMs for dialectal SLU under different levels of supervision. Our experiments cover four instruction-tuned audio LLMs (3B–30B parameters), LoRA-based fine-tuning (Hu et al., 2022), and a fully fine-tuned Whisper-small (Rad- ford et al., 2023) baseline. We also investigate the effect of training data size and synthetic augmenta- tion using an LLM+TTS pipeline. To summarize, our main contributions are as follows. •We provide a systematic evaluation of four audio LLMs for intent recognition and slot filling under zero-shot and LoRA fine-tuning settings. • We develop an LLM+TTS augmentation ap- proach that generates culturally grounded Tunisian Derja utterances with slot annotations and converts them to speech using voice-cloned TTS, targeting underrepresented intents. •We evaluate different training configurations us- ing real, synthetic, and mixed real–synthetic data, and use the best-performing configuration to build our submitted system. Findings. Our results show that zero-shot audio LLMs perform poorly on both subtasks, while LoRA fine-tuning with∼3 hours of supervised speech substantially improves performance. Syn- thetic data alone also provides clear gains over zero- shot inference, while combining real and synthetic data achieves the strongest overall results. 2 Related Work Spoken language understanding. SLU maps speech to structured semantics representations. His- torically, early SLU systems followed a cascaded arXiv:2608.18689v2 [cs.CL] 23 Aug 2026 design in which an automatic speech recogni- tion (ASR) module produced a transcript that was then processed by a text-based NLU model, in comparison to recent and increasing adoptions to end-to-end architectures (Tur and De Mori, 2011; Ghannay et al., 2018; Laperrière et al., 2022). This development has been closely accompanied by the release of increasingly challenging bench- marks (Ahmed et al., 2026). SLURP (Bastianelli et al., 2020) introduced a single-turn spoken as- sistant benchmark, while MASSIVE (FitzGerald et al., 2023) and Speech-MASSIVE (Lee et al., 2024) extended intent and slot annotation to mul- tiple languages. SLURP-TN (Elleuch et al., 2026) further adapts this, re-recording SLURP prompts in Tunisian Derja, complementing existing resources such as TARIC-SLU (Mdhaffar et al., 2024) and TEDxTN (Bougares et al., 2025). Audio LLMs. Recent audio LLMs such as Qwen- Audio and its Omni successors (Chu et al., 2023; Xu et al., 2025a,b), SALMONN (Tang et al., 2024), SpeechGPT (Zhang et al., 2023) and Gemma- 4 (Gemma Team, 2026) combine a speech en- coder with an LLM to enable zero-shot SLU via prompting. Although these models achieve strong results on English benchmarks, performance re- mains noticeably weaker on dialectal and other low-resource speech (Yang et al., 2024; Wang et al., 2025; Bhatti et al., 2026; Chen et al., 2026), including Arabic dialects (Talafha et al., 2024; Alam et al., 2025). Parameter-efficient fine-tuning (PEFT), LoRA in particular (Hu et al., 2022), is the standard approach for adapting such models in resource-constrained settings, and previous NADI editions saw adapted speech models consistently outperform approaches in dialectal ASR tracks (Ta- lafha et al., 2025; Salhab et al., 2025). Synthetic data for speech tasks.Annotated speech data collection remains expensive, par- ticularly for low-resource dialects, pushing for a growing interest in synthetic data generation. LLM-driven data generation in the Self-Instruct paradigm (Wang et al., 2023) has been combined with TTS to construct paired speech-semantic data (Noroozi et al., 2024), including recent ef- forts targeting Arabic (Sheikh Ali et al., 2026). We adopt this approach to dialectal SLU, focus- ing on generating synthetic training samples to in- crease coverage of underrepresented intent classes in the original training set. On the speech side, we build on VoxCPM (Zhou et al., 2026), a recent open, tokenizer-free TTS model with coverage of 30 languages, including Arabic. Its zero-shot voice cloning capability allows a small number of dialec- tal reference clips to represent a variety of Arabic dialects, including those without dedicated TTS voices. 3 Task and Dataset 3.1 Task Overview NADI 2026 Shared Task 5 evaluates end-to-end SLU for Tunisian Arabic speech. The dataset con- sists of short spoken assistant commands, averag- ing about four seconds (3.7 s in training, 4.3-4.5 s in the evaluation splits). Subtask intent recognition requires assigning each utterance to a single intent label. The released training data contains 23 intent labels, while the blind test set follows the full 60-label SLURP in- tent label set. The primary evaluation metrics are accuracy and weighted-F 1 , as used by the official Codabench leaderboard. We additionally report macro-F 1 to better capture performance on under- represented intent classes. These metrics provide complementary views: accuracy and weighted-F 1 are influenced by frequent classes, whereas macro- F 1 gives equal weight to each intent class. Subtask slot filling requires a transcription with inline slot annotations (<label> value >). The primary evaluation metrics are concept error rate (CoER) and concept-value error rate (CVER), which measure errors in slot labels and their associ- ated values. We additionally report word error rate (WER) and character error rate (CER) on the lexi- cal content after removing the slot markup. These metrics capture complementary aspects of perfor- mance: a system may transcribe the Derja utterance correctly while assigning incorrect slot labels or boundaries. Reporting both distinguishes transcrip- tion from semantic annotation errors. All scores use the organizers’ SLURP-TN baseline evaluation toolkit (Elleuch et al., 2026). 3.2 Dataset 3.2.1 SLURP-TN Dataset. SLURP-TN (Elleuch et al., 2026) contains Tunisian Derja re-recordings of SLURP assistant commands, with intent and slot annotations transferred from the original dataset. In Table 1, we summarize the data splits. The dataset is relatively small, with∼2.8 hours of training speech, and exhibits substantial class imbalance. The released training set contains 23 intent labels, although only 21 are observed in SplitUtts. Hours IntentsSlot Train2,6772.7821 66.2% Dev5950.7419 65.4% Devtest8931.0618 62.4% Test9891.17– Table 1: Dataset statistics. Intents: distinct labels present; Slotted: utterances with a gold slot. Test gold labels are not public. the training split. Among these, six intents have fewer than 10 training examples, and three are ab- sent from the devtest split. In addition, the most frequent intent accounts for 18.8% of the devtest utterances. 3.2.2 Data augmentation. In Figure 1, we show an overview of our data aug- mentation pipeline. We augment the training set to increase the number of examples for underrepre- sented intents among the 23 training labels. Seed utterances. We first determine the number of synthetic examples for each intent based on its frequency in the training set, generating more ex- amples for intents with fewer training instances. Fewshot generation. We use Gemini 3.1 Pro for the most underrepresented intents and Gemini 3.6 Flash for the remaining intents. Following Self- Instruct (Wang et al., 2023; Noroozi et al., 2024), the models generate Tunisian Derja utterances with inline slot annotations using six few-shot exam- ples drawn exclusively from the training split. We generate examples in three ways: (i) creating new utterances for an intent, (i) paraphrasing existing examples, and (i) generating more challenging examples with the same intent and no slot values. They contribute 16,387, 1,510, and 3,040 of the 20,937 raw candidates respectively Filtering. Since generated utterances may contain duplicates, formatting errors, or dialectal inconsis- tencies, we apply rule-based approach to identify and remove such cases. This step reduces the set to 13,876 utterances. Finally, three LLMs indepen- dently evaluate each utterance for Derja natural- ness, intent consistency, and slot correctness using the same evaluation criteria. We keep samples ac- cepted by at least two of the three models, result- ing in 12,138 synthetic utterances. Appendix C provides further details on the generation strategy, validation checks, and LLMs used at each stage (Ta- ble 5); the corresponding generation and judging prompts are released with our experimental scripts. Speech generation and filtering. We synthesize speech for each utterance using VoxCPM (Zhou et al., 2026) under two settings: the original model Figure 1: Overview of the data augmentation pipeline. and a VoxCPM model fine-tuned with LoRA on the∼2.8 hours of SLURP-TN training speech. For voice cloning, both models use the same pool of 152 reference utterances with zero WER, selected by an LLM-based ASR check from all 2,330 train- ing clips of 2.5-10s. We then filter the generated speech using rule-based criteria for duration, signal level, clipping, voiced-frame activity, and speaking rate. We generate 23,300 total utterances (11,946 from the base model and 11,354 from the LoRA- fine-tuned model), of which 22,940 retained after filtering. Mixing these with original 2,677 results in an augmented training set of 25,617 utterances. 4 System Models. Weevaluatefourinstruction- tunedaudioLLMs:Qwen2.5-Omni-3B, Qwen2.5-Omni-7B(Xuetal.,2025a), Qwen3-Omni-30B-A3B-Instruct(Xuetal., 2025b), andgemma-4-E4B-it(Gemma Team, 2026), served through ms-swift (Zhao et al., 2025) with a vLLM backend (Kwon et al., 2023) on a single H200 GPU, with greedy decoding. We additionally trainwhisper-small(Radford et al., 2023) as a small-model baseline. LoRA fine-tuning. We fine-tune each model on the SLURP-TN training split using LoRA (Hu et al., 2022), while keeping the audio encoder and audio– text aligner frozen. We use the same prompts and output formats as in the zero-shot setting to en- sure a direct comparison. For each model, we train a single LoRA adapter jointly on data from both subtasks and evaluate it separately on intent recognition and slot filling. We train all mod- els for two epochs and use the final checkpoint, without selecting checkpoints based on held-out loss. Appendix D provides the full hyperparame- ters, model merging and inference setup, and the Whisper baseline, while Appendix E provides the subtask prompts. Final system. We select Qwen3-Omni-30B-A3B for our submitted system, and fine-tune it for two epochs on the combined real and synthetic data. ModelIntent↑Slot↓ Dev-test (23 labels) Qwen2.5-Omni-3B29.2125.2 Qwen2.5-Omni-7B42.3150.1 Qwen3-Omni-30B52.5131.0 Gemma-4-E4B-it53.197.5 Whisper-small FT67.481.4 Qwen2.5-Omni-3B FT80.557.0 Qwen2.5-Omni-7B FT81.351.9 Gemma-4-E4B-it FT80.449.1 Qwen3-Omni-30B FT82.947.7 Qwen3-Omni-30B FT (Mix)86.836.9 Qwen3-Omni-30B FT (Synth)75.662.2 Blind test set (60 labels) Our system (Mix)66.159.5 Table 2: Results for both subtasks across different splits. Intent recognition is evaluated using accuracy (higher is better), while slot filling is evaluated using concept error rate (lower is better). FT: fine-tuned. 5 Results In Table 2, we report results on the dev-test and offi- cial test sets. We provide additional model-level re- sults in Table 3, and error analyses in Appendix B. Official test results. Our final system ranked 1st in slot filling (Subtask 5.2) with a CoER of 59.5 and CVER of 94.2, and ranked 4th on intent recog- nition (Subtask 5.1) with a 66.1% accuracy, and 66.9 weighted F1-score. For the intent recogni- tion, the initial submission achieves only 30.4% accuracy, compared with 86.8% on dev-test. Our analysis shows that roughly 40% of the official test utterances correspond to intents outside the 23 labels available during training. Consequently, the model predicts the broadgeneral_quirkyin- tent for 56.8% of test utterances, compared with 20.8% on dev-test (18.8% gold). We therefore map general_quirkyto the acceptedunknownlabel at inference time, making the abstention explicit. This deterministic mapping requires no retraining and increases intent accuracy by 35.7 points to 66.1%, ranking 4th of 8 teams. Zero-shot vs. baseline models. As shown in Ta- ble 2, the four omni models show limited zero-shot performance, reaching 29.2–53.1% intent accuracy and 97.5–150.1 CoER. In comparison, Whisper- small, fully fine-tuned on the training set, achieves 67.4% intent accuracy and 81.4 CoER, outperform- ing all zero-shot omni models on both subtasks. This result shows that task-specific fine-tuning with fewer than three hours of Tunisian Derja speech is more effective than direct zero-shot inference with recent omni models. In Table 3, we provide additional analysis of their zero-shot behavior. Effect of LoRA fine-tuning. We fine-tune each omni model with LoRA on the training split for both subtasks. As shown in Table 2, fine-tuning consistently improves all four models. Intent ac- curacy increases from 29.2-53.1% to 80.4-82.9%, while slot CoER decreases from 97.5-150.1 to 47.7- 57.0. Qwen3-Omni-30B achieves the best perfor- mance after fine-tuning, with 82.9% intent accu- racy and 47.7 CoER. Additionally, fine-tuning re- duces the performance gap across model sizes. The spread in intent accuracy among the 3B–30B mod- els decreases from 23.9 points in the zero-shot setting to only 2.5 points after fine-tuning. This suggests that task-specific adaptation substantially reduces the advantage of larger models on the in- domain dev-test set. Effect of data augmentation. We further fine- tune the best-performing model, Qwen3-Omni- 30B-A3B, using two data settings. Mix combines 2,677 real and 22,940 synthetic utterances, while Synth uses only synthetic speech. As shown in Ta- ble 2, Mix improves both subtasks over fine-tuning on real speech, increasing intent accuracy by 3.9 points and reducing CoER by 10.8 points. In con- trast, Synth performs better than zero-shot infer- ence but remains below fine-tuning on real speech. These results show that synthetic speech is more effective when combined with real data than when used alone. 6 Conclusions and Future Work We participated in both subtasks of NADI 2026 Shared Task 5. Our experiments show that current audio LLMs have limited zero-shot performance on Tunisian Derja SLU, and increasing model scale alone does not overcome this limitation. Fine- tuning on fewer than three hours of real speech im- proves performance to 82.9% intent accuracy and 40.0 WER for slot filling. Augmenting the training data with synthetic speech further improves per- formance to 86.8% intent accuracy and 34.7 WER. Our final system ranks 1st in slot filling and 4th of eight teams in intent recognition. Future work will extend the pipeline to other Arabic dialects, incorporate human validation into synthetic-data filtering, and release a human-validated subset for further analysis and auditing. Limitations Our evaluation is primarily based on the SLURP- TN dev-test split, while only the final setup is eval- uated on the official blind test set. In addition, the training and dev-test splits cover 23 intent labels, whereas the official test set includes a broader 60 labels, making open-intent generalization particu- larly challenging. Finally, we evaluate synthetic augmentation only with Qwen3-Omni-30B-A3B due to computational constraints. Extending this analysis to smaller models and incorporating ad- ditional human validation of synthetic speech are promising directions for future work. References Ahmed Abdelali, Hamdy Mubarak, Shammur Absar Chowdhury, Maram Hasanain, Basel Mousi, Sabri Boughorbel, Samir Abdaljalil, Yassine El Kheir, Daniel Izham, Fahim Dalvi, Majd Hawasly, Nizi Nazar, Yousseif Elshahawy, Ahmed Ali, Nadir Dur- rani, Natasa Milic-Frayling, and Firoj Alam. 2024. LAraBench: Benchmarking Arabic AI with large language models. In Proceedings of the 18th Confer- ence of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 487–520, St. Julian’s, Malta. Association for Computational Linguistics. Muhammad Abdul-Mageed, Amr Keleg, AbdelRahim Elmadany, Chiyu Zhang, Injy Hamed, Walid Magdy, Houda Bouamor, and Nizar Habash. 2024. NADI 2024: The fifth nuanced Arabic dialect identifica- tion shared task. In Proceedings of The Second Ara- bic Natural Language Processing Conference, pages 709–728, Bangkok, Thailand. Association for Com- putational Linguistics. Syeda Faiza Ahmed, Zien Sheikh Ali, Hunzalah Hassan Bhatti, Firoj Alam, and Shammur Absar Chowdhury. 2026. Multi-turn conversational ai from text to mul- timodal interaction: Data, models, evaluation, and open challenges. arXiv preprint arXiv:2608.17605. Shahad Al-Khalifa, Nadir Durrani, Hend Al-Khalifa, and Firoj Alam. 2025. The landscape of Arabic large language models. Communications of the ACM, 68(10):54–61. Firoj Alam, Md Arid Hasan, and Shammur Absar Chowdhury. 2025. SpokenNativQA: Multilingual everyday spoken queries for LLMs. In Proceed- ings of the 26th Interspeech Conference (Interspeech 2025), pages 2685–2689, Rotterdam, The Nether- lands. ISCA. Emanuele Bastianelli, Andrea Vanzo, Pawel Swieto- janski, and Verena Rieser. 2020. SLURP: A spo- ken language understanding resource package. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7252–7262. Association for Computational Linguistics. Hunzalah Hassan Bhatti and Firoj Alam. 2026. Beyond MCQ: An open-ended Arabic cultural QA bench- mark with dialect variants. In Proceedings of the Fifteenth Language Resources and Evaluation Con- ference, pages 5215–5231, Palma de Mallorca, Spain. ELRA Language Resource Association. Hunzalah Hassan Bhatti, Firoj Alam, and Shammur Ab- sar Chowdhury. 2026. Multi-task instruction tuning via data scheduling for low-resource Arabic Speech- LLMs. arXiv preprint arXiv:2601.12494. Fethi Bougares, Salima Mdhaffar, Haroun Elleuch, and Yannick Estève. 2025. TEDxTN: A three-way speech translation corpus for code-switched Tunisian Ara- bic - English. In Proceedings of the Third Arabic Natural Language Processing Conference (Arabic- NLP), pages 278–287, Suzhou, China. Association for Computational Linguistics. Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou Li. 2026. VoiceBench: Benchmarking LLM-based voice assistants. Transac- tions of the Association for Computational Linguis- tics, 14:378–398. Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shil- iang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-Audio: Advancing universal audio understanding via unified large-scale audio- language models. arXiv preprint arXiv:2311.07919. Haroun Elleuch, Salima Mdhaffar, Yannick Estève, and Fethi Bougares. 2026. SLURP-TN: Resource for Tunisian dialect spoken language understanding. arXiv preprint arXiv:2603.21940. Jack FitzGerald, Christopher Hench, Charith Peris, Scott Mackie, Kay Rottmann, Ana Sanchez, Aaron Nash, Liam Urbach, Vishesh Kakarala, Richa Singh, Swetha Ranganath, Laurie Crist, Misha Britan, Wouter Leeuwis, Gokhan Tur, and Prem Natara- jan. 2023.MASSIVE: A 1M-example multilin- gual natural language understanding dataset with 51 typologically-diverse languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4277–4302, Toronto, Canada. Association for Computational Linguistics. Gemma Team. 2026. Gemma 4 technical report. Tech- nical report, Google DeepMind. ArXiv:2607.02770. Sahar Ghannay, Antoine Caubrière, Yannick Estève, Nathalie Camelin, Edwin Simonnet, Antoine Laurent, and Emmanuel Morin. 2018. End-to-end named en- tity and semantic concept extraction from speech. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 692–699, Athens, Greece. IEEE. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR). Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gon- zalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serv- ing with PagedAttention. In Proceedings of the 29th Symposium on Operating Systems Principles (SOSP), pages 611–626, New York, NY, USA. Association for Computing Machinery. Gaëlle Laperrière, Valentin Pelloin, Antoine Caubrière, Salima Mdhaffar, Nathalie Camelin, Sahar Ghannay, Bassam Jabaian, and Yannick Estève. 2022. The spoken language understanding MEDIA benchmark dataset in the era of deep learning: data updates, training and evaluation tools. In Proceedings of the Thirteenth Language Resources and Evaluation Con- ference, pages 1595–1602, Marseille, France. Euro- pean Language Resources Association. Beomseok Lee, Ioan Calapodescu, Marco Gaido, Mat- teo Negri, and Laurent Besacier. 2024. Speech- MASSIVE: A multilingual speech dataset for SLU and beyond. In Interspeech 2024, pages 817–821. ISCA. Salima Mdhaffar, Fethi Bougares, Renato De Mori, Salah Zaiem, Mirco Ravanelli, and Yannick Estève. 2024. TARIC-SLU: A Tunisian benchmark dataset for spoken language understanding. In Proceedings of the 2024 Joint International Conference on Compu- tational Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), pages 15606–15616, Torino, Italy. ELRA and ICCL. Vahid Noroozi, Zhehuai Chen, Somshubra Majumdar, Steve Huang, Jagadeesh Balam, and Boris Ginsburg. 2024. Instruction data generation and unsupervised adaptation for speech language models. In Inter- speech 2024, pages 4049–4053. ISCA. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak su- pervision. In Proceedings of the 40th International Conference on Machine Learning (ICML), volume 202 of Proceedings of Machine Learning Research, pages 28492–28518. PMLR. Mahmoud Salhab,Shameed Sait,Mohammad Abusheikh, and Hasan Abusheikh. 2025. Munsit at NADI 2025 shared task 2: Pushing the boundaries of multidialectal Arabic ASR with weakly supervised pretraining and continual supervised fine-tuning. In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 734– 739, Suzhou, China. Association for Computational Linguistics. Zien Sheikh Ali, Hunzalah Hassan Bhatti, Ra- bindra Nath Nandi, Shammur Absar Chowdhury, and Firoj Alam. 2026. MENASpeechBank: A ref- erence voice bank with persona-conditioned multi- turn conversations for AudioLLMs. arXiv preprint arXiv:2602.07036. Peter Sullivan, Bashar Talafha, Ahmed Ashraf, Fethi Bougares, Haroun Elleuch, Chiyu Zhang, Abdel- Rahim Elmadany, Youssef Mohamed, Salima Mdhaf- far, Yannick Estève, Mohamed Elhoseiny, Hamzah Luqman, Nizar Habash, and Muhammad Abdul- Mageed. 2026. NADI-2026: The second multidi- alectal Arabic speech processing shared task. In Proceedings of the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026), Budapest, Hungary. Association for Computational Linguistics. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Mo- hamedou Cheikh Tourad, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, Hour Mohamed, Fakhraddin Al- wajih, Abdelrahman Mohamed, Abdellah El Mekki, El Moatez Billah Nagoudi, Benelhadj Djelloul Mama Saadia, Hamzah A. Alsayadi, Walid Al-Dhabyani, and 8 others. 2024. Casablanca: Data and models for multidialectal Arabic speech recognition. In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21745–21758, Miami, Florida, USA. Association for Computational Linguistics. Bashar Talafha, Hawau Olamide Toyin, Peter Sulli- van, AbdelRahim A. Elmadany, Abdurrahman Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshe- hhi, Hanan Aldarmaki, Mustafa Jarrar, Nizar Habash, and Muhammad Abdul-Mageed. 2025. NADI 2025: The first multidialectal Arabic speech processing shared task. In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 720–733, Suzhou, China. Association for Computational Linguistics. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2024. SALMONN: Towards generic hearing abilities for large language models. In The Twelfth In- ternational Conference on Learning Representations (ICLR). Gokhan Tur and Renato De Mori. 2011. Spoken Lan- guage Understanding: Systems for Extracting Seman- tic Information from Speech. John Wiley & Sons, Chichester, UK. Bin Wang, Xunlong Zou, Geyu Lin, Shuo Sun, Zhuo- han Liu, Wenyu Zhang, Zhengyuan Liu, AiTi Aw, and Nancy F. Chen. 2025. AudioBench: A univer- sal benchmark for audio large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Com- putational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), pages 4297–4316, Albuquerque, New Mexico. Association for Compu- tational Linguistics. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484–13508, Toronto, Canada. Association for Computational Linguistics. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025a. Qwen2.5-Omni technical report. Technical report, Alibaba Group. ArXiv:2503.20215. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, Yuanjun Lv, Yongqi Wang, Dake Guo, He Wang, Linhan Ma, Pei Zhang, Xinyu Zhang, Hongkun Hao, Zishan Guo, and 19 others. 2025b. Qwen3-Omni technical report. Technical report, Al- ibaba Group. ArXiv:2509.17765. Qian Yang, Jin Xu, Wenrui Liu, Yunfei Chu, Ziyue Jiang, Xiaohuan Zhou, Yichong Leng, Yuanjun Lv, Zhou Zhao, Chang Zhou, and Jingren Zhou. 2024. AIR-Bench: Benchmarking large audio-language models via generative comprehension. In Proceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 1979–1998, Bangkok, Thailand. Associ- ation for Computational Linguistics. Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. 2023. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757–15773, Sin- gapore. Association for Computational Linguistics. Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. 2025. SWIFT: A scalable lightweight infras- tructure for fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 29733–29735. AAAI Press. Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Ren- jie Yu, Jiancheng Gui, Jiaheng Wu, Ziyang Wang, Xudong Shen, Runchuan Ye, Zhisheng Zhang, Ji- uyang Zhou, Bingsong Bai, Weiyue Sun, Mengyuan Deng, Qundong Shi, Zhiyong Wu, and Zhiyuan Liu. 2026. VoxCPM2 technical report. arXiv preprint arXiv:2606.06928. A Additional Results In Table 3, we report detailed results for intent recognition and slot filling under both zero-shot and fine-tuned settings. Fine-tuning improves all omni models across the evaluation metrics, and Qwen3- Omni-30B-A3B achieves the strongest overall per- formance. In Table 4, we further analyze the effect of synthetic data and training duration on Qwen3-Omni-30B-A3B. Combining real and syn- thetic speech performs better than using either source alone, while increasing training from 2 to 2.5 epochs provides only marginal gains. Intent Recognition ModelAcc↑M-F 1 ↑W-F 1 ↑ Dev-test (23 labels) Qwen2.5-Omni-3B29.222.634.4 Qwen2.5-Omni-7B42.329.345.3 Qwen3-Omni-30B52.538.957.6 Gemma-4-E4B-it53.146.059.2 Whisper-small FT67.447.770.4 Qwen2.5-Omni-3B FT80.554.979.7 Qwen2.5-Omni-7B FT81.357.080.6 Gemma-4-E4B-it FT80.456.579.8 Qwen3-Omni-30B FT82.958.082.4 Qwen3-Omni-30B FT (Mix)86.861.986.3 Qwen3-Omni-30B FT (Synth)75.656.979.0 Blind test set (60 labels) Our system (Mix)66.1–66.9 Slot Filling ModelWER↓CER↓CoER↓CVER↓ Dev-test (23 labels) Qwen2.5-Omni-3B121.9120.1125.2134.7 Qwen2.5-Omni-7B104.7102.5150.1159.6 Qwen3-Omni-30B87.879.6131.0144.7 Gemma-4-E4B-it68.340.697.5102.1 Whisper-small FT56.225.181.4107.2 Qwen2.5-Omni-3B FT47.318.957.097.5 Qwen2.5-Omni-7B FT45.318.051.991.6 Gemma-4-E4B-it FT41.416.449.189.1 Qwen3-Omni-30B FT40.015.647.784.7 Qwen3-Omni-30B FT (Mix)34.712.536.973.6 Qwen3-Omni-30B FT (Synth)60.635.062.2103.2 Blind test set (60 labels) Our system (Mix)–59.594.2 Table 3: Detailed results for both subtasks across dif- ferent splits. Intent recognition is evaluated using ac- curacy, macro-F 1 and weighted-F 1 (higher is better), while slot filling is evaluated using WER, CER, CoER and CVER (lower is better). Mix combines real and synthetic training data, while Synth uses synthetic data only. FT: fine-tuned. Train data Ep. Acc WER CER CoER CVER Dev-test split Real only2 82.940.015.647.784.7 Mix286.834.712.536.973.6 Mix2.5 87.234.812.336.275.0 Synth only 2.5 75.660.635.062.2103.2 Validation split Mix286.132.010.834.266.0 Mix2.5 85.731.610.835.567.8 Table 4: Effect of synthetic augmentation on Qwen3- Omni-30B-A3B. Mix combines real and synthetic train- ing data, while Synth uses synthetic data only. B Error Analysis Zero-shot slot-filling errors. Zero-shot models often fail to follow the required slot markup for- mat. Fewer than 11% of their outputs use the ex- pected<label> value >structure, with missing markup as the dominant error. Models also oc- casionally generate alternative XML-style forms or over-generate slot content. After fine-tuning, however, the markup rate closely matches the gold distribution, showing that task-specific adaptation largely resolves these formatting errors. Effect of intent-label granularity. Some zero- shot intent errors reflect confusion among seman- tically related labels rather than complete misun- derstanding of the utterance. Mapping intents to coarser action labels improves zero-shot accuracy by about 7-8 points for the strongest models, while the same mapping yields only a small gain after fine-tuning. This suggests that fine-tuning helps models distinguish closely related intent labels more reliably. Effect of data augmentation. We also examine how synthetic augmentation changes errors on the dev-test set. Fine-tuning on real speech corrects a large portion of the zero-shot errors, and adding synthetic speech provides further gains on both subtasks. Overall, augmentation improves intent accuracy by 3.9 points and reduces CoER by 10.8 points (Table 4). The gains are strongest for less frequent intents and slot types. Common intents show only small improvements, whereas rarer intents gain substan- tially more (Table 6). We observe the same trend for slot filling, where utterances containing less frequent slot types benefit considerably more from augmentation than those containing only common slot types. These results suggest that synthetic data primarily improves coverage of underrepresented labels. C Data Augmentation Pipeline Deterministic filtering.We first apply determin- istic filters to the generated utterances. We retain candidates that follow the annotation format, align with the annotated text, use valid task labels, con- tain sufficient Arabic-script content, and fall within a 2-25 token range. We also remove character 3- gram near-duplicates of real training utterances and previously accepted synthetic examples. These fil- ters reduce malformed, out-of-domain, and repeti- tive generations before LLM-based evaluation. Examples from the filtering pipeline. Figure 1 summarizes the complete augmentation process. Below, we illustrate three representative outcomes: rejection during text validation, rejection after speech synthesis, and acceptance into the final train- ing set. (a) Rejected during text validation INTENT email_addcontact TEXT nh . ebb nzid Mehdi lel carnet d adresses GLOSS “I want to add Mehdi to the address book.” Rejected: the Arabic-script ratio falls below the required threshold because most of the utterance is written in French. (b) Rejected after speech synthesis INTENT general_greet TEXT joma mabruka GLOSS “Blessed Friday.” Text accepted; audio rejected: the generated speech is unusually long for the short utterance and fails the speaking-rate filter. (c) Accepted into the training set INTENT alarm_query TEXT fammaši alarme mbarmja mt ̄ a el-q ̄ ayla GLOSS “Is there an alarm set for the early afternoon?” Accepted: the text passes all validation and judging stages, and both synthesized renditions satisfy the acoustic filters. LLM roles. We use different LLMs for comple- mentary stages of the pipeline, as summarized in Table 5. Gemini 3.6 Flash generates the bulk of the synthetic data, while Gemini 3.1 Pro focuses on lower-resource intents and provides reference ASR checks. We use all three models as indepen- dent judges and retain an utterance when at least two judges accept it. This majority-vote stage re- tains 12,138 of the 13,876 candidates that reach LLM-based validation. GenerationJudgingRef. ModelRareBulkpanelASR Gemini 3.6 Flash ×✓ × stage output–16,63785.4%– Gemini 3.1 Pro✓ ×✓ stage output4,300–77.6%2,330 Gemini 2.5 Pro × ×✓ × stage output–74.6%– Table 5: LLM roles in the augmentation pipeline. The judging percentages indicate the share of candidates accepted by each model. A 2-of-3 majority vote retains 12,138 of the 13,876 judged candidates. Effect on label coverage. Synthetic augmenta- tion also makes the training distribution more bal- anced across intents. As shown in Table 6, the Train %F 1 Intent labelDev.Real MixReal Mix∆ general_quirky16815.96.275.1 80.2+5.2 weather_query15218.86.089.0 89.6+0.6 news_query122 18.76.288.7 89.2+0.4 email_query1178.15.489.9 95.7+5.7 email_sendemail1136.74.888.0 92.7+4.8 alarm_set416.85.386.4 90.2+3.9 alarm_query34 4.98.283.1 90.9+7.8 takeaway_query324.57.180.6 84.4+3.8 email_querycontact262.45.780.8 88.0+7.2 takeaway_order225.06.572.7 80.0+7.3 alarm_remove212.97.475.0 83.7+8.7 general_joke16 2.07.971.0 78.8+7.8 email_addcontact121.18.264.0 71.4+7.4 Macro, 13 labels80.3 85.8+5.4 Table 6: Effect of augmentation across intent labels, sorted by dev-test support. Dev. denotes the number of gold dev-test utterances for each label. largest F 1 gains occur for less frequent labels, while already frequent intents such asweather_query andnews_querychange only slightly. Across the 13 intent labels with sufficient dev-test instances, macro-F 1 improves from 80.3 to 85.8. This pattern suggests that augmentation primarily improves cov- erage of underrepresented intents rather than fur- ther emphasizing already common classes. D Hyperparameters We use LoRA with rank 16 andα = 32, applied only to the attention projections of the language backbone. We set the learning rate to10 −4 and the effective batch size to 8, while keeping the au- dio encoder and audio-text aligner frozen. After fine-tuning, we merge the LoRA adapter into the base model and use greedy decoding with the same vLLM serving setup. We fully fine-tune Whisper- small for two epochs and prepend a[INTENT]or [SLOT]task marker during decoding. We train all fine-tuned systems for two epochs, as extending mixed-data training to 2.5 epochs provides no con- sistent improvement (Table 4). E Prompts We use separate prompts for intent recognition and slot filling. For the official intent-recognition test set, we extend the intent prompt to support the broader label inventory. E.1 Intent Recognition We use the following prompt for all SLURP-TN training and dev-test experiments reported in Ta- bles 3 and 4. System prompt You are a spoken language understanding system for Tunisian Arabic (Tunisian dialect; code-switching with French/English words is common). Task: listen to the audio utterance and classify the speaker’s INTENT. Choose exactly one label from the fixed inventory below - do not invent new labels, do not translate, do not explain. Valid intent labels (23): Emails, addcontact, alarm_query, alarm_remove, alarm_set, email_addcontact, email_query, email_querycontact, email_sendemail, general_greet, general_joke, general_quirky, greet, joke, news_query, query, querycontact, quirky, sendemail, set, takeaway_order, takeaway_query, weather_query OUTPUT FORMAT (valid single-line JSON, no markdown or extra text): "intent": "<one label copied exactly from the list above>" User turn <audio> Listen to the utterance and identify its intent. Respond ONLY with a single-line JSON object: "intent": "<label>" E.2 Slot Filling We use the organizers’ reference prompt for slot filling (Elleuch et al., 2026). This keeps the output format consistent across zero-shot and fine-tuned models and allows us to apply the same evaluation pipeline to all systems. System prompt You are an automatic speech recognition and spoken language understanding system for Tunisian Arabic (Tunisian dialect, written in Arabic script; code-switching with French/English words is common). Task: listen to the audio and output ONE line that is the exact spoken transcription, with semantic slots marked inline using this scheme: <label> slot value > A slot opens with its label in angle brackets and closes with a lone ’>’. Words outside any <label> ... > span are left as plain transcription. Valid slot labels: <alarm_type>, <app_name>, <artist_name>, <business>, <business_name>, <business_type>, <date>, <device_type>, <drink_type>, <email_address>, <email_folder>, <event_name>, <food_type>, <general_frequency>, <house_place>, <ingredient>, <joke_type>, <list_name>, <meal_type>, <media_type>, <movie_name>, <news_topic>, <order_name>, <order_type>, <person>, <personal>, <personal_info>, <place_name>, <relation>, <time>, <time_zone>, <timeofday>, <transport_type>, <weather_descriptor> Output only the annotated transcription line: no translation, no explanation, no surrounding quotes. User turn <audio> Transcribe the audio with inline semantic slots as instructed. E.3 Official Test Set For the official intent-recognition test set, we mod- ify only the prompt and keep the model unchanged. We expand the label inventory from the 23 released training labels to the full 60-label SLURP inven- tory and allowunknownwhen the utterance does not match any available label. We also identify the six scenarios covered by the training data to discourage the model from mapping unseen intents to familiar in-domain labels. System prompt You are a spoken language understanding system for Tunisian Arabic (Tunisian dialect; code-switching with French/English words is common). Task: listen to the audio utterance and classify the speaker’s INTENT. Choose exactly one label from the fixed inventory below - do not invent new labels, do not translate, do not explain. Valid intent labels (60): alarm_query, alarm_remove, alarm_set, audio_volume_down, audio_volume_mute, audio_volume_other, audio_volume_up, calendar_query, calendar_remove, calendar_set, cooking_query, cooking_recipe, datetime_convert, datetime_query, email_addcontact, email_query, email_querycontact, email_sendemail, general_greet, general_joke, general_quirky, iot_cleaning, iot_coffee, iot_hue_lightchange, iot_hue_lightdim, iot_hue_lightoff, iot_hue_lighton, iot_hue_lightup, iot_wemo_off, iot_wemo_on, lists_createoradd, lists_query, lists_remove, music_dislikeness, music_likeness, music_query, music_settings, news_query, play_audiobook, play_game, play_music, play_podcasts, play_radio, qa_currency, qa_definition, qa_factoid, qa_maths, qa_stock, recommendation_events, recommendation_locations, recommendation_movies, social_post, social_query, takeaway_order, takeaway_query, transport_query, transport_taxi, transport_ticket, transport_traffic, weather_query If the utterance does not fit ANY label above, answer exactly: unknown IMPORTANT: this test set covers the FULL inventory above, which is much broader than the six scenarios (alarm, email, general, news, takeaway, weather) you may be most familiar with. Many utterances are about music, calendars, lists, IoT/smart-home devices, transport, cooking, social media, general question-answering or audio volume. Classify what you actually hear. Do NOT use general_quirky as a catch-all: reserve it for genuinely nonsensical or unanswerable chit-chat. If an utterance has a clear topic that is not in the list, answer unknown instead of general_quirky. OUTPUT FORMAT (valid single-line JSON, no markdown or extra text): "intent": "<one label copied exactly from the list above, or unknown>"