Aslema at NADI 2026: Augmentation through Fewshot for SLU

# Aslema at NADI 2026: Augmentation through Fewshot for SLU
Source: [https://arxiv.org/html/2608.18689](https://arxiv.org/html/2608.18689)
Tajwaar ShafiqThanks:˜The contribution was made while the author was interning at the Qatar Computing Research Institute\.Shammur Absar ChowdhuryFiroj AlamAffiliation:Qatar Computing Research Institute, HBKU, QatarEmail:[tajwaar\.shafiq@alumni\.utoronto\.ca,](mailto:)Affiliation:\{hubh90945, fialam, shchowdhury\}@hbku\.edu\.qa

###### Abstract

We presentAslema, our system for NADI 2026 Shared Task 5, which consists of two subtasks:intent recognitionandslot 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 ranks1st in slot filling\(59\.5 CoER\) and4th among 8 teams in intent recognition\(66\.1% accuracy\)\. We release our experimental scripts111[https://github\.com/hunzed/aslema\_nadi2026](https://github.com/hunzed/aslema_nadi2026)and will soon share the synthetic dataset to support further research in this area\.

## 1Introduction

Spoken dialogue interfaces are increasingly driven by large language models \(LLMs\), with recent*audio*LLMs processing speech directly and combining speech recognition with language understanding[11](https://arxiv.org/html/2608.18689#bib.bib11);[28](https://arxiv.org/html/2608.18689#bib.bib15);[32](https://arxiv.org/html/2608.18689#bib.bib12)\. However, their effectiveness for dialectal Arabic remains limited[1](https://arxiv.org/html/2608.18689#bib.bib35);[4](https://arxiv.org/html/2608.18689#bib.bib36);[8](https://arxiv.org/html/2608.18689#bib.bib37)\. Tunisian Dialect \(Derja\) is low\-resource, heavily code\-switched with French and English, and substantially different from the Modern Standard Arabic \(MSA\) that dominates Arabic training corpora[26](https://arxiv.org/html/2608.18689#bib.bib21)\. The NADI shared\-task series has increasingly focused on such dialectal settings[2](https://arxiv.org/html/2608.18689#bib.bib3);[27](https://arxiv.org/html/2608.18689#bib.bib1)\. Its 2026 edition[25](https://arxiv.org/html/2608.18689#bib.bib34)introduces Shared Task 5 on end\-to\-end Spoken Language Understanding \(SLU\) using SLURP\-TN[12](https://arxiv.org/html/2608.18689#bib.bib5), a Tunisian re\-recording of the SLURP benchmark[6](https://arxiv.org/html/2608.18689#bib.bib6), covering intent recognition and slot filling\.

We participate in both subtasks and study the effectiveness 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[16](https://arxiv.org/html/2608.18689#bib.bib25), and a fully fine\-tuned Whisper\-small[22](https://arxiv.org/html/2608.18689#bib.bib17)baseline\. We also investigate the effect of training data size and synthetic augmentation 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 approach 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 using 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∼\\sim3 hours of supervised speech substantially improves performance\. Synthetic data alone also provides clear gains over zero\-shot inference, while combining real and synthetic data achieves the strongest overall results\.

## 2Related Work

##### Spoken language understanding\.

SLU maps speech to structured semantics representations\. Historically, early SLU systems followed a cascaded design in which an automatic speech recognition \(ASR\) 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[29](https://arxiv.org/html/2608.18689#bib.bib30);[15](https://arxiv.org/html/2608.18689#bib.bib31);[18](https://arxiv.org/html/2608.18689#bib.bib32)\. This development has been closely accompanied by the release of increasingly challenging benchmarks[3](https://arxiv.org/html/2608.18689#bib.bib2)\. SLURP[6](https://arxiv.org/html/2608.18689#bib.bib6)introduced a single\-turn spoken assistant benchmark, while MASSIVE[13](https://arxiv.org/html/2608.18689#bib.bib7)and Speech\-MASSIVE[19](https://arxiv.org/html/2608.18689#bib.bib8)extended intent and slot annotation to multiple languages\. SLURP\-TN[12](https://arxiv.org/html/2608.18689#bib.bib5)further adapts this, re\-recording SLURP prompts in Tunisian Derja, complementing existing resources such as TARIC\-SLU[20](https://arxiv.org/html/2608.18689#bib.bib9)and TEDxTN[9](https://arxiv.org/html/2608.18689#bib.bib10)\.

Audio LLMs\.Recent audio LLMs such as Qwen\-Audio and its Omni successors[11](https://arxiv.org/html/2608.18689#bib.bib11);[32](https://arxiv.org/html/2608.18689#bib.bib12);[33](https://arxiv.org/html/2608.18689#bib.bib13), SALMONN[28](https://arxiv.org/html/2608.18689#bib.bib15), SpeechGPT[35](https://arxiv.org/html/2608.18689#bib.bib16)and Gemma\-4[14](https://arxiv.org/html/2608.18689#bib.bib14)combine a speech encoder with an LLM to enable zero\-shot SLU via prompting\. Although these models achieve strong results on English benchmarks, performance remains noticeably weaker on dialectal and other low\-resource speech[34](https://arxiv.org/html/2608.18689#bib.bib19);[30](https://arxiv.org/html/2608.18689#bib.bib18);[7](https://arxiv.org/html/2608.18689#bib.bib24);[10](https://arxiv.org/html/2608.18689#bib.bib20), including Arabic dialects[26](https://arxiv.org/html/2608.18689#bib.bib21);[5](https://arxiv.org/html/2608.18689#bib.bib22)\. Parameter\-efficient fine\-tuning \(PEFT\), LoRA in particular[16](https://arxiv.org/html/2608.18689#bib.bib25), 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[27](https://arxiv.org/html/2608.18689#bib.bib1);[23](https://arxiv.org/html/2608.18689#bib.bib4)\.

Synthetic data for speech tasks\.Annotated speech data collection remains expensive, particularly for low\-resource dialects, pushing for a growing interest in synthetic data generation\. LLM\-driven data generation in the Self\-Instruct paradigm[31](https://arxiv.org/html/2608.18689#bib.bib28)has been combined with TTS to construct paired speech\-semantic data[21](https://arxiv.org/html/2608.18689#bib.bib29), including recent efforts targeting Arabic[24](https://arxiv.org/html/2608.18689#bib.bib23)\. We adopt this approach to dialectal SLU, focusing on generating synthetic training samples to increase coverage of underrepresented intent classes in the original training set\. On the speech side, we build on VoxCPM[37](https://arxiv.org/html/2608.18689#bib.bib33), 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 dialectal reference clips to represent a variety of Arabic dialects, including those without dedicated TTS voices\.

## 3Task and Dataset

### 3\.1Task Overview

NADI 2026 Shared Task 5evaluates end\-to\-end SLU for Tunisian Arabic speech\. The dataset consists of short spoken assistant commands, averaging about four seconds \(3\.7 s in training, 4\.3\-4\.5 s in the evaluation splits\)\.

Subtask intent recognitionrequires 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 intent label set\. The primary evaluation metrics are accuracy and weighted\-F1, as used by the official Codabench leaderboard\. We additionally report macro\-F1to better capture performance on underrepresented intent classes\. These metrics provide complementary views: accuracy and weighted\-F1are influenced by frequent classes, whereas macro\-F1gives equal weight to each intent class\.

Subtask slot fillingrequires 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 associated values\. We additionally report word error rate \(WER\) and character error rate \(CER\) on the lexical content after removing the slot markup\. These metrics capture complementary aspects of performance: a system may transcribe the Derja utterance correctly while assigning incorrect slot labels or boundaries\. Reporting both distinguishes transcription from semantic annotation errors\. All scores use the organizers’ SLURP\-TN baseline evaluation toolkit[12](https://arxiv.org/html/2608.18689#bib.bib5)\.

### 3\.2Dataset

#### 3\.2\.1SLURP\-TN Dataset\.

SLURP\-TN[12](https://arxiv.org/html/2608.18689#bib.bib5)contains Tunisian Derja re\-recordings of SLURP assistant commands, with intent and slot annotations transferred from the original dataset\. In Table[1](https://arxiv.org/html/2608.18689#S3.T1), we summarize the data splits\. The dataset is relatively small, with∼\\sim2\.8 hours of training speech, and exhibits substantial class imbalance\. The released training set contains 23 intent labels, although only 21 are observed in the training split\. Among these, six intents have fewer than 10 training examples, and three are absent from the devtest split\. In addition, the most frequent intent accounts for 18\.8% of the devtest utterances\.

Table 1:Dataset statistics\.*Intents*: distinct labels present;*Slotted*: utterances with a gold slot\.*Test*gold labels are not public\.
#### 3\.2\.2Data augmentation\.

In Figure[1](https://arxiv.org/html/2608.18689#S3.F1), we show an overview of our data augmentation pipeline\. We augment the training set to increase the number of examples for underrepresented 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 examples 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[31](https://arxiv.org/html/2608.18689#bib.bib28);[21](https://arxiv.org/html/2608.18689#bib.bib29), the models generate Tunisian Derja utterances with inline slot annotations using six few\-shot examples drawn exclusively from the training split\. We generate examples in three ways:\(i\)creating new utterances for an intent,\(ii\)paraphrasing existing examples, and\(iii\)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 inconsistencies, we apply rule\-based approach to identify and remove such cases\. This step reduces the set to13,876utterances\. Finally, three LLMs independently evaluate each utterance for Derja naturalness, intent consistency, and slot correctness using the same evaluation criteria\. We keep samples accepted by at least two of the three models, resulting in12,138synthetic utterances\. Appendix[C](https://arxiv.org/html/2608.18689#A3)provides further details on the generation strategy, validation checks, and LLMs used at each stage \(Table[5](https://arxiv.org/html/2608.18689#A3.T5)\); the corresponding generation and judging prompts are released with our experimental scripts\.

![Refer to caption](https://arxiv.org/html/2608.18689v1/figures/pipeline_handmade_v3.png)Figure 1:Overview of the data augmentation pipeline\.Speech generation and filtering\.We synthesize speech for each utterance using VoxCPM[37](https://arxiv.org/html/2608.18689#bib.bib33)under two settings: the original model and a VoxCPM model fine\-tuned with LoRA on the∼\\sim2\.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 training 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 generate23,300total utterances \(11,946 from the base model and 11,354 from the LoRA\-fine\-tuned model\), of which22,940retained after filtering\. Mixing these with original 2,677 results in an augmented training set of25,617utterances\.

## 4System

##### Models\.

We evaluate four instruction\-tuned audio LLMs:Qwen2\.5\-Omni\-3B,Qwen2\.5\-Omni\-7B[32](https://arxiv.org/html/2608.18689#bib.bib12),Qwen3\-Omni\-30B\-A3B\-Instruct[33](https://arxiv.org/html/2608.18689#bib.bib13), andgemma\-4\-E4B\-it[14](https://arxiv.org/html/2608.18689#bib.bib14), served through ms\-swift[36](https://arxiv.org/html/2608.18689#bib.bib27)with a vLLM backend[17](https://arxiv.org/html/2608.18689#bib.bib26)on a single H200 GPU, with greedy decoding\. We additionally trainwhisper\-small[22](https://arxiv.org/html/2608.18689#bib.bib17)as a small\-model baseline\.

LoRA fine\-tuning\.We fine\-tune each model on the SLURP\-TN training split using LoRA[16](https://arxiv.org/html/2608.18689#bib.bib25), 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 ensure 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 models for two epochs and use the final checkpoint, without selecting checkpoints based on held\-out loss\. Appendix[D](https://arxiv.org/html/2608.18689#A4)provides the full hyperparameters, model merging and inference setup, and the Whisper baseline, while Appendix[E](https://arxiv.org/html/2608.18689#A5)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\.

## 5Results

In Table[2](https://arxiv.org/html/2608.18689#S5.T2), we report results on the dev\-test and official test sets\. We provide additional model\-level results in Table[3](https://arxiv.org/html/2608.18689#A1.T3), and error analyses in Appendix[B](https://arxiv.org/html/2608.18689#A2)\.

Official test results\.Our final system ranked1stin slot filling \(Subtask 5\.2\) with a CoER of 59\.5 and CVER of 94\.2, and ranked4thon intent recognition \(Subtask 5\.1\) with a 66\.1% accuracy, and 66\.9 weighted F1\-score\. For the intent recognition, 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\_quirkyintent for 56\.8% of test utterances, compared with 20\.8% on dev\-test \(18\.8% gold\)\. We therefore mapgeneral\_quirkyto the acceptedunknownlabel at inference time, making the abstention explicit\. This deterministic mapping requires no retraining and increases intent accuracy by 35\.7 points to66\.1%, ranking4th of 8 teams\.

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\.Zero\-shot vs\. baseline models\.As shown in Table[2](https://arxiv.org/html/2608.18689#S5.T2), 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, outperforming 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](https://arxiv.org/html/2608.18689#A1.T3), 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](https://arxiv.org/html/2608.18689#S5.T2), fine\-tuning consistently improves all four models\. Intent accuracy 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 performance after fine\-tuning, with 82\.9% intent accuracy and 47\.7 CoER\. Additionally, fine\-tuning reduces the performance gap across model sizes\. The spread in intent accuracy among the 3B–30B models 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 Table[2](https://arxiv.org/html/2608.18689#S5.T2),*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 contrast,*Synth*performs better than zero\-shot inference 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\.

## 6Conclusions 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 improves performance to 82\.9% intent accuracy and 40\.0 WER for slot filling\. Augmenting the training data with synthetic speech further improves performance to 86\.8% intent accuracy and 34\.7 WER\. Our final system ranks1stin slot filling and4thof 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 evaluated 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 particularly challenging\. Finally, we evaluate synthetic augmentation only with Qwen3\-Omni\-30B\-A3B due to computational constraints\. Extending this analysis to smaller models and incorporating additional human validation of synthetic speech are promising directions for future work\.

## References

- Abdelaliet al\.\(2024\)A\. Abdelali, H\. Mubarak, S\. A\. Chowdhury, M\. Hasanain, B\. Mousi, S\. Boughorbel, S\. Abdaljalil, Y\. El Kheir, D\. Izham, F\. Dalvi, M\. Hawasly, N\. Nazar, Y\. Elshahawy, A\. Ali, N\. Durrani, N\. Milic\-Frayling, and F\. AlamLAraBench: benchmarking Arabic AI with large language models\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),Y\. Graham and M\. Purver \(Eds\.\),St\. Julian’s, Malta,pp\. 487–520\.External Links:[Link](https://aclanthology.org/2024.eacl-long.30/),[Document](https://dx.doi.org/10.18653/v1/2024.eacl-long.30)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1)\.
- Abdul\-Mageedet al\.\(2024\)M\. Abdul\-Mageed, A\. Keleg, A\. Elmadany, C\. Zhang, I\. Hamed, W\. Magdy, H\. Bouamor, and N\. HabashNADI 2024: the fifth nuanced Arabic dialect identification shared task\.InProceedings of The Second Arabic Natural Language Processing Conference,Bangkok, Thailand,pp\. 709–728\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.arabicnlp-1.79)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1)\.
- Ahmedet al\.\(2026\)S\. F\. Ahmed, Z\. S\. Ali, H\. H\. Bhatti, F\. Alam, and S\. A\. ChowdhuryMulti\-turn conversational ai from text to multimodal interaction: data, models, evaluation, and open challenges\.arXiv preprint arXiv:2608\.17605\.External Links:[Link](https://arxiv.org/abs/2608.17605)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Al\-Khalifaet al\.\(2025\)S\. Al\-Khalifa, N\. Durrani, H\. Al\-Khalifa, and F\. AlamThe landscape of Arabic large language models\.Communications of the ACM68\(10\),pp\. 54–61\.External Links:[Document](https://dx.doi.org/10.1145/3737453)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1)\.
- Alamet al\.\(2025\)F\. Alam, M\. A\. Hasan, and S\. A\. ChowdhurySpokenNativQA: multilingual everyday spoken queries for LLMs\.InProceedings of the 26th Interspeech Conference \(Interspeech 2025\),Rotterdam, The Netherlands,pp\. 2685–2689\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-2011)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Bastianelliet al\.\(2020\)E\. Bastianelli, A\. Vanzo, P\. Swietojanski, and V\. RieserSLURP: a spoken language understanding resource package\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 7252–7262\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.588)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Bhattiet al\.\(2026\)H\. H\. Bhatti, F\. Alam, and S\. A\. ChowdhuryMulti\-task instruction tuning via data scheduling for low\-resource Arabic SpeechLLMs\.arXiv preprint arXiv:2601\.12494\.External Links:[Link](https://arxiv.org/abs/2601.12494)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Bhatti and Alam \(2026\)H\. H\. Bhatti and F\. AlamBeyond MCQ: an open\-ended Arabic cultural QA benchmark with dialect variants\.InProceedings of the Fifteenth Language Resources and Evaluation Conference,S\. Piperidis, N\. Bel, H\. van den Heuvel, N\. Ide, S\. Krek, and A\. Toral \(Eds\.\),Palma de Mallorca, Spain,pp\. 5215–5231\.External Links:[Link](https://aclanthology.org/2026.lrec-1.408/),[Document](https://dx.doi.org/10.63317/2smjp2wega4e)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1)\.
- Bougareset al\.\(2025\)F\. Bougares, S\. Mdhaffar, H\. Elleuch, and Y\. EstèveTEDxTN: a three\-way speech translation corpus for code\-switched Tunisian Arabic \- English\.InProceedings of the Third Arabic Natural Language Processing Conference \(ArabicNLP\),Suzhou, China,pp\. 278–287\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.arabicnlp-main.22)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2026\)Y\. Chen, X\. Yue, C\. Zhang, X\. Gao, R\. T\. Tan, and H\. LiVoiceBench: benchmarking LLM\-based voice assistants\.Transactions of the Association for Computational Linguistics14,pp\. 378–398\.External Links:[Document](https://dx.doi.org/10.1162/tacl.a.628)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Chuet al\.\(2023\)Y\. Chu, J\. Xu, X\. Zhou, Q\. Yang, S\. Zhang, Z\. Yan, C\. Zhou, and J\. ZhouQwen\-Audio: advancing universal audio understanding via unified large\-scale audio\-language models\.arXiv preprint arXiv:2311\.07919\.External Links:[Link](https://arxiv.org/abs/2311.07919)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Elleuchet al\.\(2026\)H\. Elleuch, S\. Mdhaffar, Y\. Estève, and F\. BougaresSLURP\-TN: resource for Tunisian dialect spoken language understanding\.arXiv preprint arXiv:2603\.21940\.External Links:[Link](https://arxiv.org/abs/2603.21940)Cited by:[§E\.2](https://arxiv.org/html/2608.18689#A5.SS2.p1.1),[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2608.18689#S3.SS1.p3.1),[§3\.2\.1](https://arxiv.org/html/2608.18689#S3.SS2.SSS1.p1.1)\.
- FitzGeraldet al\.\(2023\)J\. FitzGerald, C\. Hench, C\. Peris, S\. Mackie, K\. Rottmann, A\. Sanchez, A\. Nash, L\. Urbach, V\. Kakarala, R\. Singh, S\. Ranganath, L\. Crist, M\. Britan, W\. Leeuwis, G\. Tur, and P\. NatarajanMASSIVE: a 1M\-example multilingual natural language understanding dataset with 51 typologically\-diverse languages\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 4277–4302\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.235)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Gemma Team \(2026\)Gemma TeamGemma 4 technical report\.Technical reportGoogle DeepMind\.Note:arXiv:2607\.02770External Links:[Link](https://arxiv.org/abs/2607.02770)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1),[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Ghannayet al\.\(2018\)S\. Ghannay, A\. Caubrière, Y\. Estève, N\. Camelin, E\. Simonnet, A\. Laurent, and E\. MorinEnd\-to\-end named entity and semantic concept extraction from speech\.In2018 IEEE Spoken Language Technology Workshop \(SLT\),Athens, Greece,pp\. 692–699\.External Links:[Document](https://dx.doi.org/10.1109/SLT.2018.8639513)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p2.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1),[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p2.1)\.
- Kwonet al\.\(2023\)W\. Kwon, Z\. Li, S\. Zhuang, Y\. Sheng, L\. Zheng, C\. H\. Yu, J\. E\. Gonzalez, H\. Zhang, and I\. StoicaEfficient memory management for large language model serving with PagedAttention\.InProceedings of the 29th Symposium on Operating Systems Principles \(SOSP\),New York, NY, USA,pp\. 611–626\.External Links:[Document](https://dx.doi.org/10.1145/3600006.3613165)Cited by:[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Laperrièreet al\.\(2022\)G\. Laperrière, V\. Pelloin, A\. Caubrière, S\. Mdhaffar, N\. Camelin, S\. Ghannay, B\. Jabaian, and Y\. EstèveThe spoken language understanding MEDIA benchmark dataset in the era of deep learning: data updates, training and evaluation tools\.InProceedings of the Thirteenth Language Resources and Evaluation Conference,Marseille, France,pp\. 1595–1602\.External Links:[Document](https://dx.doi.org/10.63317/25nrh2p76hnh)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Leeet al\.\(2024\)B\. Lee, I\. Calapodescu, M\. Gaido, M\. Negri, and L\. BesacierSpeech\-MASSIVE: a multilingual speech dataset for SLU and beyond\.InInterspeech 2024,pp\. 817–821\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-957)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Mdhaffaret al\.\(2024\)S\. Mdhaffar, F\. Bougares, R\. De Mori, S\. Zaiem, M\. Ravanelli, and Y\. EstèveTARIC\-SLU: a Tunisian benchmark dataset for spoken language understanding\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),Torino, Italy,pp\. 15606–15616\.External Links:[Document](https://dx.doi.org/10.63317/4evgpxxgwme5)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Norooziet al\.\(2024\)V\. Noroozi, Z\. Chen, S\. Majumdar, S\. Huang, J\. Balam, and B\. GinsburgInstruction data generation and unsupervised adaptation for speech language models\.InInterspeech 2024,pp\. 4049–4053\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-1575)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p3.1),[§3\.2\.2](https://arxiv.org/html/2608.18689#S3.SS2.SSS2.p3.1)\.
- Radfordet al\.\(2023\)A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. SutskeverRobust speech recognition via large\-scale weak supervision\.InProceedings of the 40th International Conference on Machine Learning \(ICML\),Proceedings of Machine Learning Research, Vol\.202,pp\. 28492–28518\.External Links:[Link](https://proceedings.mlr.press/v202/radford23a.html)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p2.1),[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Salhabet al\.\(2025\)M\. Salhab, S\. Sait, M\. Abusheikh, and H\. AbusheikhMunsit at NADI 2025 shared task 2: pushing the boundaries of multidialectal Arabic ASR with weakly supervised pretraining and continual supervised fine\-tuning\.InProceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks,Suzhou, China,pp\. 734–739\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.100)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Sheikh Aliet al\.\(2026\)Z\. Sheikh Ali, H\. H\. Bhatti, R\. N\. Nandi, S\. A\. Chowdhury, and F\. AlamMENASpeechBank: a reference voice bank with persona\-conditioned multi\-turn conversations for AudioLLMs\.arXiv preprint arXiv:2602\.07036\.External Links:[Link](https://arxiv.org/abs/2602.07036)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p3.1)\.
- Sullivanet al\.\(2026\)P\. Sullivan, B\. Talafha, A\. Ashraf, F\. Bougares, H\. Elleuch, C\. Zhang, A\. Elmadany, Y\. Mohamed, S\. Mdhaffar, Y\. Estève, M\. Elhoseiny, H\. Luqman, N\. Habash, and M\. Abdul\-MageedNADI\-2026: the second multidialectal Arabic speech processing shared task\.InProceedings of the Fourth Arabic Natural Language Processing Conference \(ArabicNLP 2026\),Budapest, Hungary\.Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1)\.
- Talafhaet al\.\(2024\)B\. Talafha, K\. Kadaoui, S\. M\. Magdy, M\. Habiboullah, C\. M\. Chafei, A\. O\. El\-Shangiti, H\. Zayed, M\. C\. Tourad, R\. Alhamouri, R\. Assi, A\. Alraeesi, H\. Mohamed, F\. Alwajih, A\. Mohamed, A\. El Mekki, E\. M\. B\. Nagoudi, B\. D\. M\. Saadia, H\. A\. Alsayadi, W\. Al\-Dhabyani, S\. Shatnawi, Y\. Ech\-chammakhy, A\. Makouar, Y\. Berrachedi, M\. Jarrar, S\. Shehata, I\. Berrada, and M\. Abdul\-MageedCasablanca: data and models for multidialectal Arabic speech recognition\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 21745–21758\.Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Talafhaet al\.\(2025\)B\. Talafha, H\. O\. Toyin, P\. Sullivan, A\. A\. Elmadany, A\. Juma, A\. Djanibekov, C\. Zhang, H\. Alshehhi, H\. Aldarmaki, M\. Jarrar, N\. Habash, and M\. Abdul\-MageedNADI 2025: the first multidialectal Arabic speech processing shared task\.InProceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks,Suzhou, China,pp\. 720–733\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.99)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Tanget al\.\(2024\)C\. Tang, W\. Yu, G\. Sun, X\. Chen, T\. Tan, W\. Li, L\. Lu, Z\. Ma, and C\. ZhangSALMONN: towards generic hearing abilities for large language models\.InThe Twelfth International Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2310.13289)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Tur and De Mori \(2011\)G\. Tur and R\. De MoriSpoken language understanding: systems for extracting semantic information from speech\.John Wiley & Sons,Chichester, UK\.Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p1.1)\.
- Wanget al\.\(2025\)B\. Wang, X\. Zou, G\. Lin, S\. Sun, Z\. Liu, W\. Zhang, Z\. Liu, A\. Aw, and N\. F\. ChenAudioBench: a universal benchmark for audio large language models\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),Albuquerque, New Mexico,pp\. 4297–4316\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.218)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Wanget al\.\(2023\)Y\. Wang, Y\. Kordi, S\. Mishra, A\. Liu, N\. A\. Smith, D\. Khashabi, and H\. HajishirziSelf\-instruct: aligning language models with self\-generated instructions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 13484–13508\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.754)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p3.1),[§3\.2\.2](https://arxiv.org/html/2608.18689#S3.SS2.SSS2.p3.1)\.
- Xuet al\.\(2025a\)J\. Xu, Z\. Guo, J\. He, H\. Hu, T\. He, S\. Bai, K\. Chen, J\. Wang, Y\. Fan, K\. Dang, B\. Zhang, X\. Wang, Y\. Chu, and J\. LinQwen2\.5\-Omni technical report\.Technical reportAlibaba Group\.Note:arXiv:2503\.20215External Links:[Link](https://arxiv.org/abs/2503.20215)Cited by:[§1](https://arxiv.org/html/2608.18689#S1.p1.1),[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1),[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Xuet al\.\(2025b\)J\. Xu, Z\. Guo, H\. Hu, Y\. Chu, X\. Wang, J\. He, Y\. Wang, X\. Shi, T\. He, X\. Zhu, Y\. Lv, Y\. Wang, D\. Guo, H\. Wang, L\. Ma, P\. Zhang, X\. Zhang, H\. Hao, Z\. Guo, B\. Yang, B\. Zhang, Z\. Ma, X\. Wei, S\. Bai, K\. Chen, X\. Liu, P\. Wang, M\. Yang, D\. Liu, X\. Ren, B\. Zheng, R\. Men, F\. Zhou, B\. Yu, J\. Yang, L\. Yu, J\. Zhou, and J\. LinQwen3\-Omni technical report\.Technical reportAlibaba Group\.Note:arXiv:2509\.17765External Links:[Link](https://arxiv.org/abs/2509.17765)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1),[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Yanget al\.\(2024\)Q\. Yang, J\. Xu, W\. Liu, Y\. Chu, Z\. Jiang, X\. Zhou, Y\. Leng, Y\. Lv, Z\. Zhao, C\. Zhou, and J\. ZhouAIR\-Bench: benchmarking large audio\-language models via generative comprehension\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 1979–1998\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.109)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Zhanget al\.\(2023\)D\. Zhang, S\. Li, X\. Zhang, J\. Zhan, P\. Wang, Y\. Zhou, and X\. QiuSpeechGPT: empowering large language models with intrinsic cross\-modal conversational abilities\.InFindings of the Association for Computational Linguistics: EMNLP 2023,Singapore,pp\. 15757–15773\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.1055)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p2.1)\.
- Zhaoet al\.\(2025\)Y\. Zhao, J\. Huang, J\. Hu, X\. Wang, Y\. Mao, D\. Zhang, Z\. Jiang, Z\. Wu, B\. Ai, A\. Wang, W\. Zhou, and Y\. ChenSWIFT: a scalable lightweight infrastructure for fine\-tuning\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 29733–29735\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v39i28.35383)Cited by:[§4](https://arxiv.org/html/2608.18689#S4.SS0.SSS0.Px1.p1.1)\.
- Zhouet al\.\(2026\)Y\. Zhou, G\. Zeng, X\. Liu, X\. Li, R\. Yu, J\. Gui, J\. Wu, Z\. Wang, X\. Shen, R\. Ye, Z\. Zhang, J\. Zhou, B\. Bai, W\. Sun, M\. Deng, Q\. Shi, Z\. Wu, and Z\. LiuVoxCPM2 technical report\.arXiv preprint arXiv:2606\.06928\.External Links:[Link](https://arxiv.org/abs/2606.06928)Cited by:[§2](https://arxiv.org/html/2608.18689#S2.SS0.SSS0.Px1.p3.1),[§3\.2\.2](https://arxiv.org/html/2608.18689#S3.SS2.SSS2.p5.1)\.

## Appendix AAdditional Results

In Table[3](https://arxiv.org/html/2608.18689#A1.T3), 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 performance\. In Table[4](https://arxiv.org/html/2608.18689#A1.T4), we further analyze the effect of synthetic data and training duration on Qwen3\-Omni\-30B\-A3B\. Combining real and synthetic speech performs better than using either source alone, while increasing training from 2 to 2\.5 epochs provides only marginal gains\.

Table 3:Detailed results for both subtasks across different splits\. Intent recognition is evaluated using accuracy, macro\-F1and weighted\-F1\(higher is better\), 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\.Table 4:Effect of synthetic augmentation on Qwen3\-Omni\-30B\-A3B\.*Mix*combines real and synthetic training data, while*Synth*uses synthetic data only\.
## Appendix BError Analysis

##### Zero\-shot slot\-filling errors\.

Zero\-shot models often fail to follow the required slot markup format\. Fewer than 11% of their outputs use the expected<label\> value \>structure, with missing markup as the dominant error\. Models also occasionally 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 semantically related labels rather than complete misunderstanding 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 \(Table[4](https://arxiv.org/html/2608.18689#A1.T4)\)\.

The gains are strongest for less frequent intents and slot types\. Common intents show only small improvements, whereas rarer intents gain substantially more \(Table[6](https://arxiv.org/html/2608.18689#A3.T6)\)\. 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\.

## Appendix CData Augmentation Pipeline

##### Deterministic filtering\.

We first apply deterministic filters to the generated utterances\. We retain candidates that follow the annotation format, align with the annotated text, use valid task labels, contain 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 filters reduce malformed, out\-of\-domain, and repetitive generations before LLM\-based evaluation\.

##### Examples from the filtering pipeline\.

Figure[1](https://arxiv.org/html/2608.18689#S3.F1)summarizes the complete augmentation process\. Below, we illustrate three representative outcomes: rejection during text validation, rejection after speech synthesis, and acceptance into the final training set\.

\(a\) Rejected during text validation

intentemail\_addcontact

textnḥebb 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

intentgeneral\_greet

textjom\{a 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

intentalarm\_query

textfammaši alarme mbarmja mtā\{el\-qāyla

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 complementary stages of the pipeline, as summarized in Table[5](https://arxiv.org/html/2608.18689#A3.T5)\. 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 independent judges and retain an utterance when at least two judges accept it\. This majority\-vote stage retains 12,138 of the 13,876 candidates that reach LLM\-based validation\.

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 augmentation also makes the training distribution more balanced across intents\. As shown in Table[6](https://arxiv.org/html/2608.18689#A3.T6), the largest F1gains occur for less frequent labels, while already frequent intents such asweather\_queryandnews\_querychange only slightly\. Across the 13 intent labels with sufficient dev\-test instances, macro\-F1improves from 80\.3 to 85\.8\. This pattern suggests that augmentation primarily improves coverage of underrepresented intents rather than further emphasizing already common classes\.

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\.

## Appendix DHyperparameters

We use LoRA with rank 16 andα=32\\alpha=32, applied only to the attention projections of the language backbone\. We set the learning rate to10−410^\{\-4\}and the effective batch size to 8, while keeping the audio 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 consistent improvement \(Table[4](https://arxiv.org/html/2608.18689#A1.T4)\)\.

## Appendix EPrompts

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\.1Intent Recognition

We use the following prompt for all SLURP\-TN training and dev\-test experiments reported in Tables[3](https://arxiv.org/html/2608.18689#A1.T3)and[4](https://arxiv.org/html/2608.18689#A1.T4)\.

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\.2Slot Filling

We use the organizers’ reference prompt for slot filling[12](https://arxiv.org/html/2608.18689#bib.bib5)\. 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\.3Official Test Set

For the official intent\-recognition test set, we modify 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 inventory 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\>"\}