---
title: "Nuha-Speech: Building General-Purpose Arabic Speech-LLMs — Wang 2026"
description: "Arabic has little dedicated infrastructure for training and testing speech-LLMs. 25 references."
authors: ["Yingzhi Wang", "Reem Alhazzani", "Muhammad Alqurishi"]
year: 2026
venue: "arXiv:2609.11892"
source: "https://arxiv.org/abs/2609.11892"
doi: "10.48550/arxiv.2609.11892"
arxiv: "2609.11892"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/wang-2026-nuha-speech-building-general-purpose/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wang-2026-nuha-speech-building-general-purpose/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) / Nuha-Speech: Building General-Purpose Arabic Speech-LLMs

# Nuha-Speech: Building General-Purpose Arabic Speech-LLMs

**Authors:** [Yingzhi Wang](https://md.research.elamur.ai/free/en/authors/yingzhi-wang/), Reem Alhazzani and [Muhammad Alqurishi](https://md.research.elamur.ai/free/en/authors/muhammad-alqurishi/) · **Year:** 2026 · **Venue:** arXiv:2609.11892 · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2609.11892) · [PDF](https://arxiv.org/pdf/2609.11892) · [DOI 10.48550/arxiv.2609.11892](https://doi.org/10.48550/arxiv.2609.11892) · [arXiv 2609.11892](https://arxiv.org/abs/2609.11892) · [OpenAlex](https://openalex.org/W7212277118)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Arabic has little dedicated infrastructure for training and testing speech-LLMs. Existing Arabic-centric systems cover only a few tasks, and the large multilingual models are rarely checked on Arabic. The authors address this with three linked contributions.

- **Training corpus:** about 1.5M speech instruction samples, each formatted as a {speech, instruction, output} triple. They come mostly from public corpora, with LLM-generated QA and translations and TTS-synthesized paralinguistic data filling the gaps.
- **Fine-tuning:** three Qwen-Omni models (3B, 7B, 30B) trained with LoRA in two stages.
- **Benchmark:** seven tasks, each with its own metric.

All fine-tuned models improved on every task. The biggest jumps came on paralinguistic tasks such as dialect and age recognition, where the base models scored near chance. This suggests those skills were barely present in pretraining. The 30B model gained least because it was already strong. The work matters as a reproducible recipe for building a general-purpose speech assistant in a language with scarce speech resources.

## Key points
- **Corpus composition (1501K samples):**
  - ASR: about 790K clips from SADA, Common Voice, MASC and MGB-2, kept only if 1–10s long.
  - AST: CoVoST-v2 Ar→En, plus 150K ASR transcripts translated by Qwen3-32B.
  - SQA: 150K closed-ended and 150K open-ended pairs generated from MGB-2 transcripts.
  - DI: 150K samples from ADI-17.
  - SER: 20K ElevenLabs-synthesized clips.
  - AR: 10K real and 10K synthetic clips.
  - GR: 51K SADA clips.
  - Multi-SQA: 15K multi-turn samples.
- **Training recipe:**
  - Stage 1: ASR only, 790K samples, 3 epochs.
  - Stage 2: all tasks, with ASR cut to 200K.
  - Audio encoders frozen; LoRA (rank 8, alpha 32) on all linear layers.
  - MoE load-balancing coefficient of 1e-3.
- **SQA test set:** 1,237 utterances synthesized with XTTS-v2 using 58 voices and kept only after a Whisper-large-v3 CER filter. Answers are scored 0–3 by a LLaMA-3.3-70B-Instruct judge.
- **Gains on DI and AR (baseline → fine-tuned):**

  | Model | DI accuracy | AR accuracy |
  |---|---|---|
  | 3B | 9.70% → 57.83% | 41.75% → 83.00% |
  | 7B | 13.66% → 72.65% | 33.33% → 73.00% |
  | 30B | 16.78% → 76.02% | 42.67% → 85.75% |

  GR reached up to 99.50% (Nuha-Speech-7B).
- **Excluded baseline:** Qwen2-Audio-7B-Instruct, fine-tuned on the same data, lagged Qwen2.5-Omni-7B on every task and was dropped from the benchmark.

## Interesting ideas
- An LLM writes a wide pool of instruction phrasings for each task, so the model learns the task rather than one fixed prompt; this helps any system that must cope with varied requests.
- The training runs as a curriculum: first the basic mapping (speech→text) alone, then all tasks together with the basic data subsampled to keep its skill while rebalancing the mix.
- Synthetic test audio is checked by a round trip through ASR and discarded when the transcript drifts too far (CER filter), a cheap automatic quality gate for generated data.
- In closed-set tasks, any answer outside the allowed label set counts as wrong, so free-form drift cannot inflate accuracy.
- Measuring the before/after gain per task reveals which abilities pretraining neglected, since the largest jumps mark the thinnest prior coverage.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (answer): Generate many rewordings of each school-test or fable question and confirm the reader returns the same answer, which tests understanding rather than template matching.
- **I2** (learn): Train gparse on plain UD trees first, then jointly with frame or role objectives on a reduced tree subset, so parsing quality holds while the other heads catch up.
- **I3** (back): Run the text → graph → text compressor and keep only silver items whose regenerated text stays close to the source, giving a self-check on both generation and parsing.
- **I4** (answer): Score multiple-choice and numeric word problems so that off-format or out-of-range outputs are errors, keeping evaluation honest about format compliance.
- **I5** (learn): Track per-stage deltas after each training round to find which layer (morphology, attachments, roles) the data underserves and target new silver data there.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · [🐪 Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Cite

Cite as: Yingzhi Wang, Reem Alhazzani and Muhammad Alqurishi. 2026. Nuha-Speech: Building General-Purpose Arabic Speech-LLMs. arXiv:2609.11892.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{wang2026nuha,
  title = {{Nuha-Speech: Building General-Purpose Arabic Speech-LLMs}},
  author = {Yingzhi Wang, Reem Alhazzani and Muhammad Alqurishi},
  year = {2026},
  doi = {10.48550/arxiv.2609.11892},
  eprint = {2609.11892},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.11892}
}
```

- [← previous NRD: A Hybrid Disentanglement Framework for Mitigating Interference in Multilingual Machine Translation](https://md.research.elamur.ai/free/en/papers/zhang-2026-nrd-hybrid-disentanglement-framework-mitigating/)
- [next → OasisSimp: An Open-source Asian-English Sentence Simplification Dataset](https://md.research.elamur.ai/free/en/papers/liu-2026-oasissimp-open-source-asian-english/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/)

## References

25 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Mohammad Al-Fetyani, Muhammad Al-Barham, Gheith Abandah, Adham Alsharkawi, and Maha Dawas. 2023. Masc: Massive arabic speech corpus. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 1006–1013. IEEE. [find](https://scholar.google.com/scholar?q=Mohammad+Al%2DFetyani%2C+Muhammad+Al%2DBarham%2C+Gheith+Abandah%2C+Adham+Alsharkawi%2C+and+Maha+Dawas%2E+2023%2E+Masc%3A+Massive+arabic+speech+corpus%2E+In+2022+IEEE+Spoken+Language+Technology+Workshop+%28SLT%29%2C+pages+1006%E2%80%93)
2. Sadeen Alharbi, Areeb Alowisheq, Zoltán Tüske, Kareem Darwish, Abdullah Alrajeh, Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma Ibrahim, Raghad Aloraini, Raneem Alnajim, and 1 others. 2024. Sada: Saudi audio dataset for arabic. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10286–10290. IEEE. [find](https://scholar.google.com/scholar?q=Sadeen+Alharbi%2C+Areeb+Alowisheq%2C+Zolt%C3%A1n+T%C3%BCske%2C+Kareem+Darwish%2C+Abdullah+Alrajeh%2C+Abdulmajeed+Alrowithi%2C+Aljawharah+Bin+Tamran%2C+Asma+Ibrahim%2C+Raghad+Aloraini%2C+Raneem+Alnajim%2C+and+1+others%2E+2024%2E+Sada%3A+)
3. Ahmed Ali, Peter Bell, James Glass, Yacine Messaoui, Hamdy Mubarak, Steve Renals, and Yifan Zhang. 2016. The mgb-2 challenge: Arabic multi-dialect broadcast media recognition. In 2016 IEEE Spoken Language Technology Workshop (SLT), pages 279– 284. IEEE. [our card](https://md.research.elamur.ai/free/en/papers/ali-2016-mgb-2-challenge-arabic-multi/)
4. Alhanoof Althnian, Norah A Alzahrani, Shaykhah Z Alsubaie, Eman Albilali, Ahmed Abdelali, Nouf M Alotaibi, M Saiful Bari, Yazeed Alnumay, Abdulhamed Alothaimen, Maryam Saif, and 1 others. 2025. Araeval: An arabic multi-task evaluation suite for large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 33025–33049. [find](https://scholar.google.com/scholar?q=Alhanoof+Althnian%2C+Norah+A+Alzahrani%2C+Shaykhah+Z+Alsubaie%2C+Eman+Albilali%2C+Ahmed+Abdelali%2C+Nouf+M+Alotaibi%2C+M+Saiful+Bari%2C+Yazeed+Alnumay%2C+Abdulhamed+Alothaimen%2C+Maryam+Saif%2C+and+1+others%2E+2025%2E+Araeva)
5. Sara Althubaiti, Vasista Sai Lodagala, Tjad Clark, Yousseif Ahmed Elshahawy, Daniel Izham, Abdullah Alrajeh, Aljawahrah Bin Tamran, and Ahmed Ali. 2025. Octopus: Towards building the arabic speech llm suite. In Proceedings of The Third Arabic Natural Language Processing Conference, pages 425–435. [find](https://scholar.google.com/scholar?q=Sara+Althubaiti%2C+Vasista+Sai+Lodagala%2C+Tjad+Clark%2C+Yousseif+Ahmed+Elshahawy%2C+Daniel+Izham%2C+Abdullah+Alrajeh%2C+Aljawahrah+Bin+Tamran%2C+and+Ahmed+Ali%2E+2025%2E+Octopus%3A+Towards+building+the+arabic+speech+llm)
6. Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. 2019. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670. [↗](https://arxiv.org/abs/1912.06670) [our card](https://md.research.elamur.ai/free/en/papers/ardila-2019-common-voice-massively-multilingual-speech/)
7. Giuseppe Attanasio, Beatrice Savoldi, Dennis Fucci, and Dirk Hovy. 2024. Twists, humps, and pebbles: Multilingual speech recognition models exhibit gender performance gaps. arXiv preprint arXiv:2402.17954. [↗](https://arxiv.org/abs/2402.17954)
8. Edresson Casanova, Kelly Davis, Eren Gölge, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, and 1 others. 2024. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904. [↗](https://arxiv.org/abs/2406.04904)
9. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, and 1 others. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759. [↗](https://arxiv.org/abs/2407.10759) [our card](https://md.research.elamur.ai/free/en/papers/chu-2024-qwen2-audio-technical-report/)
10. Nilaksh Das, Saket Dingliwal, Srikanth Ronanki, Rohit Paturi, Zhaocheng Huang, Prashant Mathur, Jie Yuan, Dhanush Bekal, Xing Niu, Sai Muralidhar Jayanthi, and 1 others. 2024. Speechverse: A large-scale generalizable audio language model. arXiv preprint arXiv:2405.08295. [↗](https://arxiv.org/abs/2405.08295)
11. Yexing Du, Youcheng Pan, Ziyang Ma, Bo Yang, Yifan Yang, Keqi Deng, Xie Chen, Yang Xiang, Ming Liu, and Bing Qin. 2025. Making llms better many-to-many speech-to-text translators with curriculum learning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12466–12478. [find](https://scholar.google.com/scholar?q=Yexing+Du%2C+Youcheng+Pan%2C+Ziyang+Ma%2C+Bo+Yang%2C+Yifan+Yang%2C+Keqi+Deng%2C+Xie+Chen%2C+Yang+Xiang%2C+Ming+Liu%2C+and+Bing+Qin%2E+2025%2E+Making+llms+better+many%2Dto%2Dmany+speech%2Dto%2Dtext+translators+with+curriculum+learn)
12. Fahim Faisal, Sharlina Keshava, Md Mahfuz Ibn Alam, and Antonios Anastasopoulos. 2021. Sd-qa: Spoken dialectal question answering for the real world. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3296–3315. [find](https://scholar.google.com/scholar?q=Fahim+Faisal%2C+Sharlina+Keshava%2C+Md+Mahfuz+Ibn+Alam%2C+and+Antonios+Anastasopoulos%2E+2021%2E+Sd%2Dqa%3A+Spoken+dialectal+question+answering+for+the+real+world%2E+In+Findings+of+the+Association+for+Computational+L)
13. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3. [our card](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/)
14. Chien-yu Huang, Ke-Han Lu, Shih-Heng Wang, Chi-Yuan Hsiao, Chun-Yi Kuan, Haibin Wu, Siddhant Arora, Kai-Wei Chang, Jiatong Shi, Yifan Peng, and 1 others. 2024. Dynamic-superb: Towards a dynamic, collaborative, and comprehensive instruction-tuning benchmark for speech. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 12136–12140. IEEE. [find](https://scholar.google.com/scholar?q=Chien%2Dyu+Huang%2C+Ke%2DHan+Lu%2C+Shih%2DHeng+Wang%2C+Chi%2DYuan+Hsiao%2C+Chun%2DYi+Kuan%2C+Haibin+Wu%2C+Siddhant+Arora%2C+Kai%2DWei+Chang%2C+Jiatong+Shi%2C+Yifan+Peng%2C+and+1+others%2E+2024%2E+Dynamic%2Dsuperb%3A+Towards+a+dynamic%2C+colla)
15. Zhaozhi Qian, Faroq Altam, Muhammad Alqurishi, and Riad Souissi. 2024. Cameleval: Advancing culturally aligned arabic language models and benchmarks. arXiv preprint arXiv:2409.12623. [↗](https://arxiv.org/abs/2409.12623) [our card](https://md.research.elamur.ai/free/en/papers/qian-2024-cameleval-advancing-culturally-aligned-arabic/)
16. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR. [our card](https://md.research.elamur.ai/free/en/papers/radford-2022-robust-speech-recognition-large-scale/)
17. Suwon Shon, Ahmed M. Ali, Younes Samih, Hamdy Mubarak, and James R. Glass. 2020. Adi17: A fine-grained arabic dialect identification dataset. ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8244–8248. [our card](https://md.research.elamur.ai/free/en/papers/salameh-2018-fine-grained-arabic-dialect-identification/)
18. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2023. Salmonn: Towards generic hearing abilities for large language models. arXiv preprint arXiv:2310.13289. [↗](https://arxiv.org/abs/2310.13289) [our card](https://md.research.elamur.ai/free/en/papers/tang-2023-salmonn-towards-generic-hearing-abilities/)
19. Bin Wang, Xunlong Zou, Geyu Lin, Shuo Sun, Zhuohan Liu, Wenyu Zhang, Zhengyuan Liu, AiTi Aw, and Nancy Chen. 2025a. Audiobench: A universal benchmark for audio large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4297–4316. [find](https://scholar.google.com/scholar?q=Bin+Wang%2C+Xunlong+Zou%2C+Geyu+Lin%2C+Shuo+Sun%2C+Zhuohan+Liu%2C+Wenyu+Zhang%2C+Zhengyuan+Liu%2C+AiTi+Aw%2C+and+Nancy+Chen%2E+2025a%2E+Audiobench%3A+A+universal+benchmark+for+audio+large+language+models%2E+In+Proceedings+of)
20. Changhan Wang, Anne Wu, Jiatao Gu, and Juan Pino. 2021. Covost 2 and massively multilingual speech translation. In Interspeech, volume 2021, pages 2247–2251. [find](https://scholar.google.com/scholar?q=Changhan+Wang%2C+Anne+Wu%2C+Jiatao+Gu%2C+and+Juan+Pino%2E+2021%2E+Covost+2+and+massively+multilingual+speech+translation%2E+In+Interspeech%2C+volume+2021%2C+pages+2247%E2%80%932251%2E)
21. Dingdong Wang, Jincenzi Wu, Junan Li, Dongchao Yang, Xueyuan Chen, Tianhua Zhang, and Helen Meng. 2025b. Mmsu: A massive multi-task spoken language understanding and reasoning benchmark. arXiv preprint arXiv:2506.04779. [↗](https://arxiv.org/abs/2506.04779)
22. Yingzhi Wang, Anas Alhmoud, and Muhammad Alqurishi. 2024. Open universal arabic asr leaderboard. arXiv preprint arXiv:2412.13788. [↗](https://arxiv.org/abs/2412.13788) [our card](https://md.research.elamur.ai/free/en/papers/wang-2024-open-universal-arabic-asr-leaderboard/)
23. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, and 1 others. 2025a. Qwen2. 5-omni technical report. arXiv preprint arXiv:2503.20215. [↗](https://arxiv.org/abs/2503.20215) [our card](https://md.research.elamur.ai/free/en/papers/xu-2025-qwen2-5-omni-technical/)
24. 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. Preprint, arXiv:2509.17765. Technical Report. [↗](https://arxiv.org/abs/2509.17765) [our card](https://md.research.elamur.ai/free/en/papers/xu-2025-qwen3-omni-technical-report/)
25. Qian Yang, Jin Xu, Wenrui Liu, Yunfei Chu, Ziyue Jiang, Xiaohuan Zhou, Yichong Leng, Yuanjun Lv, Zhou Zhao, Chang Zhou, and 1 others. 2024. Air-bench: Benchmarking large audio-language models via generative comprehension. arXiv preprint arXiv:2402.07729. [↗](https://arxiv.org/abs/2402.07729)