---
title: "MarsadLab at NADI Shared Task — paper by Biswas et al. (2025)"
description: "The authors entered NADI 2025, which covers two problems on the Casablanca speech corpus across eight… 61.6% accuracy. 15 references."
authors: ["Md. Rafiul Biswas", "Kais Attia", "Shimaa Ibrahim", "Mabrouka Bessghaier", "Wajdi Zaghouani"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks"
source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.103/"
doi: "10.18653/v1/2025.arabicnlp-sharedtasks.103"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/biswas-2025-marsadlab-nadi-shared-task-arabic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/biswas-2025-marsadlab-nadi-shared-task-arabic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/18/) / MarsadLab at NADI Shared Task: Arabic Dialect Identification and Speech Recognition using ECAPA-TDNN and Whisper

# MarsadLab at NADI Shared Task: Arabic Dialect Identification and Speech Recognition using ECAPA-TDNN and Whisper

**Authors:** [Md. Rafiul Biswas](https://md.research.elamur.ai/free/en/authors/md-rafiul-biswas/), [Kais Attia](https://md.research.elamur.ai/free/en/authors/kais-attia/), [Shimaa Ibrahim](https://md.research.elamur.ai/free/en/authors/shimaa-amer-ibrahim/), [Mabrouka Bessghaier](https://md.research.elamur.ai/free/en/authors/mabrouka-bessghaier/) and [Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.103/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.103.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.103](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.103) · [OpenAlex](https://openalex.org/W4416037644)

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

## Summary
The authors entered NADI 2025, which covers two problems on the Casablanca speech corpus across eight Arabic dialects: guessing a speaker's dialect from a single utterance (ADI) and transcribing dialectal speech (ASR). For ADI they started from an ECAPA-TDNN model pretrained on VoxLingua107 and placed a small classifier on top. The classifier includes a feature-wise attention gate. Training ran in two stages: first only the head, then the upper encoder layers as well, each part with its own learning rate. They also mixed focal and cross-entropy loss, used heavy audio augmentation, and averaged predictions over augmented copies at test time. This system came third, with 61.6% accuracy and a macro cost of 0.3068 (the winner scored 0.7983). For ASR they used Whisper Large-v3 zero-shot, attached to a frozen MARBERT and one trainable linear layer. It placed seventh, with a WER of about 105 and a CER of about 85. The study shows that a carefully adapted acoustic model handles dialect discrimination well, while an untuned general ASR model struggles with colloquial Arabic.

## Key points
- The data has 25,600 labelled 16 kHz clips, 3,200 per dialect, split roughly in half between train (12,900) and validation (12,700). There are also 6,268 unlabelled test clips. Clip length ranges from 1.04 to 15.12 s, with a mean of 4.25 s.
- ADI training settings:
  - Phase 1 trains only the head for 2,500 steps.
  - Phase 2 unfreezes the top encoder layers, using a learning rate of 1e-6 for the encoder and 5e-5 for the classifier.
  - The loss is 0.3·focal (γ = 2.5) + 0.7·CE, with label smoothing and cosine LR with warmup.
- Inference uses 5–10 augmented variants per clip. Their softmax outputs are averaged, and temperature scaling is applied for calibration.
- The cascade has only 6,152 trainable parameters, which is 0.0004% of the model. These are an 8×768 linear map over the MARBERT [CLS] embedding.
- In the error analysis, the main confusions were Moroccan with Algerian and Levantine with Palestinian. Whisper failed on code-switched Arabic–French speech, often dropped short function words, and garbled named entities.

## Interesting ideas
- **Sigmoid gate on frozen features.** A sigmoid gate learns a per-dimension weight for a frozen pretrained embedding before classification, giving a cheap and inspectable form of feature selection.
- **Staged unfreezing with split learning rates.** Train the new head first, then unfreeze only the top layers, giving the backbone a learning rate about 50× smaller. This adapts the model to a new domain without wiping out what it learned in pretraining.
- **Blended focal and cross-entropy loss.** Mixing focal loss into ordinary cross-entropy shifts attention toward hard, confusable examples without discarding the stable CE signal.
- **Test-time augmentation with calibration.** Run the model on several perturbed versions of the same input, average the probabilities, then calibrate with a temperature, yielding steadier and better-calibrated decisions.
- **Cascade information loss.** The authors attribute part of the cascade's weakness to transcription discarding the acoustic cues that dialect identification relies on, so early commitment in a pipeline loses information.

## For Mova

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

- **I1** (tree): gparse could gate transformer token features before arc and label scoring, and the gate values could be logged to show which feature dimensions drove an attachment.
- **I2** (learn): when fine-tuning gparse on silver treebanks agreed by witness parsers, this schedule could limit drift away from gold-trained behaviour.
- **I3** (tokens): the tagger could use this blend to sharpen rare or easily confused tags, such as particle vs adposition, without destabilizing common tags.
- **I4** (learn): the parser could be applied to harmless surface variants of a sentence (casing, punctuation), with agreement across variants used as a confidence signal for accepting silver trees.
- **I5** (frames): pragmatics could keep pointers to alternative parses or lemmas, rather than only the single best tree, so that irony or role decisions can recover cues the parser dropped.

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

## Institutions

- [Northwestern University in Qatar](https://md.research.elamur.ai/free/en/universities/northwestern-university-in-qatar/) — University, Doha, Qatar · 39 works · 148 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 493 works

🧭 Topics (2): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 519 works

## Cite

Cite as: Md. Rafiul Biswas, Kais Attia, Shimaa Ibrahim, Mabrouka Bessghaier and Wajdi Zaghouani. 2025. MarsadLab at NADI Shared Task: Arabic Dialect Identification and Speech Recognition using ECAPA-TDNN and Whisper. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

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

```bibtex
@inproceedings{biswas2025marsadlab,
  title = {{MarsadLab at NADI Shared Task: Arabic Dialect Identification and Speech Recognition using ECAPA-TDNN and Whisper}},
  author = {Md. Rafiul Biswas, Kais Attia, Shimaa Ibrahim, Mabrouka Bessghaier and Wajdi Zaghouani},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.103},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.103}
}
```

- [← previous MAPROC at AHaSIS Shared Task: Few-Shot and Sentence Transformer for Sentiment Analysis of Arabic Hotel Reviews](https://md.research.elamur.ai/free/en/papers/zarnoufi-2025-maproc-ahasis-shared-task-few/)
- [next → Martín Arista, Javier, ed., Sara Domínguez Barragán, Luisa Fidalgo Allo, Laura García Fernández, Yosra Hamdoun Bghiyel, Miguel Lacalle Palacios, Raquel Mateo Mendaza, Carmen Novo Urraca, Ana Elvira Ojanguren López, Esaúl Ruiz Narbona, Roberto Torre Alonso, Marta Tío Sáenz, and Raquel Vea Escarza. 2024. “Nerthusv5. Interface of textual, lexicographical and secondary sources of Old English.” www.nerthusproject.com.](https://md.research.elamur.ai/free/en/papers/tichy-2025-martin-arista-javier-ed-sara/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/18/)

## References

15 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Eman Aboelela and Omar Mansour. 2025. A review of speech recognition and application to arabic speech recognition. In Future of Information and Communication Conference, pages 13–31. Springer. [find](https://scholar.google.com/scholar?q=Eman+Aboelela+and+Omar+Mansour%2E+2025%2E+A+review+of+speech+recognition+and+application+to+arabic+speech+recognition%2E+In+Future+of+Information+and+Communication+Conference%2C+pages+13%E2%80%9331%2E+Springer%2E)
2. Ahmed Ali, Peter Bell, James Glass, Yacine Messaoui, Hamdy Mubarak, Steve Renals, and Yifan Zhang. 2019. [The mgb-2 challenge: Arabic multi-dialect broadcast media recognition](https://md.research.elamur.ai/free/en/papers/ali-2016-mgb-2-challenge-arabic-multi/). Preprint, arXiv:1609.05625. [↗](https://arxiv.org/abs/1609.05625)
3. Ahmed Ali, Stephan Vogel, and Steve Renals. 2017. [Speech recognition challenge in the wild: Arabic mgb-3](https://md.research.elamur.ai/free/en/papers/ali-2017-speech-recognition-challenge-wild-arabic/). Preprint, arXiv:1709.07276. [↗](https://arxiv.org/abs/1709.07276)
4. Maha J. Althobaiti. 2020. [Automatic arabic dialect identification systems for written texts: A survey](https://md.research.elamur.ai/free/en/papers/althobaiti-2020-automatic-arabic-dialect-identification-systems/). Preprint, arXiv:2009.12622. [↗](https://arxiv.org/abs/2009.12622)
5. Fadi Biadsy, Julia Bell Hirschberg, and Nizar Y Habash. 2009. [Spoken arabic dialect identification using phonotactic modeling](https://md.research.elamur.ai/free/en/papers/biadsy-2009-spoken-arabic-dialect-identification-phonotactic/).
6. Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. 2020a. ECAPA-TDNN: Emphasized Channel Attention, propagation and aggregation in TDNN based speaker verification. In Interspeech 2020, pages 3830–3834. [find](https://scholar.google.com/scholar?q=Brecht+Desplanques%2C+Jenthe+Thienpondt%2C+and+Kris+Demuynck%2E+2020a%2E+ECAPA%2DTDNN%3A+Emphasized+Channel+Attention%2C+propagation+and+aggregation+in+TDNN+based+speaker+verification%2E+In+Interspeech+2020%2C+pages+38)
7. Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. 2020b. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. In Interspeech 2020, pages 3830–3834. [find](https://scholar.google.com/scholar?q=Brecht+Desplanques%2C+Jenthe+Thienpondt%2C+and+Kris+Demuynck%2E+2020b%2E+Ecapa%2Dtdnn%3A+Emphasized+channel+attention%2C+propagation+and+aggregation+in+tdnn+based+speaker+verification%2E+In+Interspeech+2020%2C+pages+38)
8. Ashifur Rahman, Md Mohsin Kabir, Muhammad Firoz Mridha, Mohammed Alatiyyah, Haifa F Alhasson, and Shuaa S Alharbi. 2024. Arabic speech recognition: Advancement and challenges. IEEE Access, 12:39689–39716. [find](https://scholar.google.com/scholar?q=Ashifur+Rahman%2C+Md+Mohsin+Kabir%2C+Muhammad+Firoz+Mridha%2C+Mohammed+Alatiyyah%2C+Haifa+F+Alhasson%2C+and+Shuaa+S+Alharbi%2E+2024%2E+Arabic+speech+recognition%3A+Advancement+and+challenges%2E+IEEE+Access%2C+12%3A39689%E2%80%9339)
9. Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan, Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, Ju-Chieh Chou, Sung-Lin Yeh, Szu-Wei Fu, Chien-Feng Liao, Elena Rastorgueva, François Grondin, William Aris, Hwidong Na, Yan Gao, and 2 others. 2021a. SpeechBrain: A general-purpose speech toolkit. Preprint, arXiv:2106.04624. ArXiv:2106.04624. [↗](https://arxiv.org/abs/2106.04624)
10. Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan, Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, Ju-Chieh Chou, Sung-Lin Yeh, Szu-Wei Fu, Chien-Feng Liao, Elena Rastorgueva, François Grondin, William Aris, Hwidong Na, Yan Gao, and 2 others. 2021b. Speechbrain: A general-purpose speech toolkit. Preprint, arXiv:2106.04624. [↗](https://arxiv.org/abs/2106.04624)
11. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, and 1 others. 2024. [Casablanca: Data and models for multidialectal arabic speech recognition](https://md.research.elamur.ai/free/en/papers/talafha-2024-casablanca-data-models-multidialectal-arabic/). arXiv preprint arXiv:2410.04527. [↗](https://arxiv.org/abs/2410.04527)
12. Bashar Talafha, Hawau Olamide Toyin, Peter Sullivan, AbdelRahim Elmadany, Abdurrahman Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshehhi, Hanan Aldarmaki, Mustafa Jarar, Nizar Habash, and Muhammad Abdul-Mageed. 2025. [Nadi 2025: The first multidialectal arabic speech processing shared task](https://md.research.elamur.ai/free/en/papers/talafha-2025-nadi-2025-first-multidialectal-arabic/). In The Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou. Association for Computational Linguistics.
13. Raphael Tang, Karun Kumar, Gefei Yang, Akshat Pandey, Yajie Mao, Vladislav Belyaev, Madhuri Em-madi, Craig Murray, Ferhan Ture, and Jimmy Lin. 2022. Speechnet: Weakly supervised, end-to-end speech recognition at industrial scale. arXiv preprint arXiv:2211.11740. [↗](https://arxiv.org/abs/2211.11740)
14. Jörgen Valk and Tanel Alumäe. 2021. VoxLingua107: a dataset for spoken language recognition. In Proc. IEEE SLT Workshop. [find](https://scholar.google.com/scholar?q=J%C3%B6rgen+Valk+and+Tanel+Alum%C3%A4e%2E+2021%2E+VoxLingua107%3A+a+dataset+for+spoken+language+recognition%2E+In+Proc%2E+IEEE+SLT+Workshop%2E)
15. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. [Scaling speech technology to 1,000+ languages](https://md.research.elamur.ai/free/en/papers/pratap-2023-scaling-speech-technology-1-000/). Preprint, arXiv:2305.13516. [↗](https://arxiv.org/abs/2305.13516)

*Updated 2 November 2026*