---
title: "CTC-DID: CTC-Based Arabic dialect — Farooq 2026"
description: "The paper addresses spoken Arabic dialect identification (DID) under limited training data and real-time constraints."
authors: ["Muhammad Umar Farooq", "Oscar Saz"]
year: 2026
venue: "arXiv:2601.12199"
source: "https://arxiv.org/abs/2601.12199"
doi: "10.48550/arxiv.2601.12199"
arxiv: "2601.12199"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/farooq-2026-ctc-did-ctc-based-arabic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/farooq-2026-ctc-did-ctc-based-arabic/)

[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/3/) / CTC-DID: CTC-Based Arabic dialect identification for streaming applications

# CTC-DID: CTC-Based Arabic dialect identification for streaming applications

**Authors:** Muhammad Umar Farooq, Oscar Saz · **Year:** 2026 · **Venue:** arXiv:2601.12199 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2601.12199) · [PDF](https://arxiv.org/pdf/2601.12199) · [DOI 10.48550/arxiv.2601.12199](https://doi.org/10.48550/arxiv.2601.12199) · [arXiv 2601.12199](https://arxiv.org/abs/2601.12199) · [OpenAlex](https://openalex.org/W7125129939)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14) · 1 in our library, of the 140884 papers whose bibliographies we read

## Summary
The paper addresses spoken Arabic dialect identification (DID) under limited training data and real-time constraints. Most existing systems, such as x-vectors, ECAPA-TDNN and fine-tuned Whisper, give one label per utterance. That design hurts streaming use and short inputs. The authors instead treat DID as a tiny-vocabulary speech recognition problem. Each training utterance gets a "transcript" made of its dialect tag repeated once per estimated word. A self-supervised encoder (mHuBERT) is then trained with CTC loss. The word count comes either from an off-the-shelf ASR system or from a Language-Agnostic Heuristic (LAH): Silero VAD measures speech duration, and the count is derived from an assumed speaking rate. At inference, the decoded tag sequence is reduced to its most frequent tag.

Trained on only 10 or 50 hours per dialect from ADI-17, CTC-DID beats Whisper-medium and ECAPA-TDNN on the ADI-17 test set. It also beats them in zero-shot testing on Casablanca. It holds up better on short clips and adapts to chunked streaming with modest loss. The approach matters because it offers a small, cheap and streamable alternative to large Whisper models.

## Key points
- Training data is drawn from ADI-17, which has 3,033 hours across 17 Arabic dialects. Only 10 h or 50 h per dialect were used. Yemeni, Moroccan and Jordan data were topped up with speed-perturbed copies, and Jordan reached only 40 h. Validation used a balanced 30 minutes per dialect.
- LAH estimates the word count as 5 × the VAD speech duration in seconds. Results stay stable for any rate of at least 3 words per second, so the heuristic is not finely tuned.
- The output vocabulary contains only the dialect labels, the CTC blank symbol and an optional space token. Decoding is greedy.
- On zero-shot Casablanca, the mHuBERT CTC-DID model (119M parameters) beats Whisper-medium (769M) by 4.66 absolute F1.
- Streaming works by splitting audio into chunks and prepending left context to each one. Outputs for the context frames are discarded, and the encoder yields about 50 frames per second. With a total window above 4 s, F1 reaches 82.34, compared with 86.98 on full utterances.

## Interesting ideas
- An utterance-level label can be turned into a per-unit label sequence by repeating the tag and then taking a majority vote. This yields running predictions and allows mixed labels within one input. — For Mova (tree): Variety or register tags (e.g., US/UK spelling, informal web style) could be predicted per token and aggregated by vote. The UD dialect-conversion step could then choose rules per sentence and still handle mixed passages.
- The target length needs only a rough, cheap estimate, and the authors show that the method tolerates a wide range of values for it. Sweeping such a knob is a useful habit whenever supervision is approximate. — For Mova (learn): When silver data relies on crude heuristics, Mova could routinely run a sensitivity sweep over the heuristic's parameters. This would show that results do not hinge on a lucky constant.
- In chunked inference with left context, the model sees earlier material but outputs for that overlap are discarded. The kept outputs are concatenated, which gives near-full-context quality at a bounded cost. — For Mova (tree): gparse could parse long documents in windows that carry preceding sentences as context. It would keep predictions only for the new span, which keeps memory bounded.
- The authors report accuracy as a function of input length and as a grid of chunk size × context size. This shows where a method breaks down, not just its average score. — For Mova (answer): Reader and QA accuracy could be reported by text length and by the amount of preceding context given. This would reveal whether answers depend on long-range discourse.
- Testing zero-shot on a separate out-of-domain corpus (Casablanca) checks generalisation beyond the training distribution. — For Mova (learn): Each trained component could be scored on a held-out treebank or genre never seen in training, alongside the in-domain test.

*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/)

## Cite

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

```bibtex
@misc{farooq2026ctc,
  title = {{CTC-DID: CTC-Based Arabic dialect identification for streaming applications}},
  author = {Farooq, Muhammad Umar and Saz, Oscar},
  year = {2026},
  doi = {10.48550/arxiv.2601.12199},
  eprint = {2601.12199},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2601.12199}
}
```

- [← previous CS-YODAS: A Mined Dataset of In-the-Wild Code-Switched Speech](https://md.research.elamur.ai/free/en/papers/yan-2026-cs-yodas-mined-dataset-wild/)
- [next → CUHKSZ Simultaneous Speech Translation System for IWSLT 2026](https://md.research.elamur.ai/free/en/papers/yang-2026-cuhksz-simultaneous-speech-translation-system/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/)

## References

16 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. A. Kulkarni and H. Aldarmaki, “Yet another model for Arabic dialect identification,” in ArabicNLP 2023, 2023, pp. 435–440. [find](https://scholar.google.com/scholar?q=A%2E+Kulkarni+and+H%2E+Aldarmaki%2C+%E2%80%9CYet+another+model+for+Arabic+dialect+identification%2C%E2%80%9D+in+ArabicNLP+2023%2C+2023%2C+pp%2E+435%E2%80%93440%2E)
2. M. Torgbi, A. Clayman, J. J. Speight, and H. Tayyar Madabushi, “Adapting whisper for regional dialects: Enhancing public services for vulnerable populations in the United Kingdom,” in Proceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, 2025, pp. 29–38. [find](https://scholar.google.com/scholar?q=M%2E+Torgbi%2C+A%2E+Clayman%2C+J%2E+J%2E+Speight%2C+and+H%2E+Tayyar+Madabushi%2C+%E2%80%9CAdapting+whisper+for+regional+dialects%3A+Enhancing+public+services+for+vulnerable+populations+in+the+United+Kingdom%2C%E2%80%9D+in+Proceedings+of+t)
3. A. Hanani and R. Naser, “Spoken arabic dialect recognition using x-vectors,” Natural Language Engineering, vol. 26, no. 6, p. 691–700, 2020. [find](https://scholar.google.com/scholar?q=A%2E+Hanani+and+R%2E+Naser%2C+%E2%80%9CSpoken+arabic+dialect+recognition+using+x%2Dvectors%2C%E2%80%9D+Natural+Language+Engineering%2C+vol%2E+26%2C+no%2E+6%2C+p%2E+691%E2%80%93700%2C+2020%2E)
4. L. Lonergan, M. Qian, N. Nı́ Chiaráin, C. Gobl, and A. Nı́ Chasaide, “Towards spoken dialect identification of irish,” in 2nd Annual Meeting of the ELRA/ISCA SIG on Under-resourced Languages (SIGUL 2023), 2023, pp. 63–67. [find](https://scholar.google.com/scholar?q=L%2E+Lonergan%2C+M%2E+Qian%2C+N%2E+N%C4%B1%CC%81+Chiara%CC%81in%2C+C%2E+Gobl%2C+and+A%2E+N%C4%B1%CC%81+Chasaide%2C+%E2%80%9CTowards+spoken+dialect+identification+of+irish%2C%E2%80%9D+in+2nd+Annual+Meeting+of+the+ELRA%2FISCA+SIG+on+Under%2Dresourced+Languages+%28SIGUL+2)
5. D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X-vectors: Robust dnn embeddings for speaker recognition,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 5329–5333. [find](https://scholar.google.com/scholar?q=D%2E+Snyder%2C+D%2E+Garcia%2DRomero%2C+G%2E+Sell%2C+D%2E+Povey%2C+and+S%2E+Khudanpur%2C+%E2%80%9CX%2Dvectors%3A+Robust+dnn+embeddings+for+speaker+recognition%2C%E2%80%9D+in+2018+IEEE+International+Conference+on+Acoustics%2C+Speech+and+Signal+Proc)
6. B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” in Interspeech 2020, 2020, pp. 3830–3834. [find](https://scholar.google.com/scholar?q=B%2E+Desplanques%2C+J%2E+Thienpondt%2C+and+K%2E+Demuynck%2C+%E2%80%9CEcapa%2Dtdnn%3A+Emphasized+channel+attention%2C+propagation+and+aggregation+in+tdnn+based+speaker+verification%2C%E2%80%9D+in+Interspeech+2020%2C+2020%2C+pp%2E+3830%E2%80%933834%2E)
7. A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in ICML, 2023. [find](https://scholar.google.com/scholar?q=A%2E+Radford%2C+J%2E+W%2E+Kim%2C+T%2E+Xu%2C+G%2E+Brockman%2C+C%2E+McLeavey%2C+and+I%2E+Sutskever%2C+%E2%80%9CRobust+speech+recognition+via+large%2Dscale+weak+supervision%2C%E2%80%9D+in+ICML%2C+2023%2E) [our card](https://md.research.elamur.ai/free/en/papers/radford-2022-robust-speech-recognition-large-scale/)
8. H. Elleuch, S. Mdhaffar, Y. Estève, and F. Bougares, “ADI-20: Arabic Dialect Identification dataset and models,” in Interspeech 2025, 2025, pp. 2775–2779. [find](https://scholar.google.com/scholar?q=H%2E+Elleuch%2C+S%2E+Mdhaffar%2C+Y%2E+Este%CC%80ve%2C+and+F%2E+Bougares%2C+%E2%80%9CADI%2D20%3A+Arabic+Dialect+Identification+dataset+and+models%2C%E2%80%9D+in+Interspeech+2025%2C+2025%2C+pp%2E+2775%E2%80%932779%2E) [our card](https://md.research.elamur.ai/free/en/papers/elleuch-2025-adi-20-arabic-dialect-identification/)
9. A. Prasad, A. Carofilis, G. Vanderreydt, D. Khalil, S. Madikeri, P. Motlicek, and C. Schuepbach, “Fine-tuning self-supervised models for language identification using orthonormal constraint,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 11 921–11 925. [find](https://scholar.google.com/scholar?q=A%2E+Prasad%2C+A%2E+Carofilis%2C+G%2E+Vanderreydt%2C+D%2E+Khalil%2C+S%2E+Madikeri%2C+P%2E+Motlicek%2C+and+C%2E+Schuepbach%2C+%E2%80%9CFine%2Dtuning+self%2Dsupervised+models+for+language+identification+using+orthonormal+constraint%2C%E2%80%9D+in+ICASS)
10. Z. Fan, M. Li, S. Zhou, and B. Xu, “Exploring wav2vec 2.0 on speaker verification and language identification,” in Interspeech 2021, 2021, pp. 1509–1513. [find](https://scholar.google.com/scholar?q=Z%2E+Fan%2C+M%2E+Li%2C+S%2E+Zhou%2C+and+B%2E+Xu%2C+%E2%80%9CExploring+wav2vec+2%2E0+on+speaker+verification+and+language+identification%2C%E2%80%9D+in+Interspeech+2021%2C+2021%2C+pp%2E+1509%E2%80%931513%2E)
11. A. Angra, H. Muralikrishna, D. A. Dinesh, and V. Thenkanidiyoor, “Exploring aggregated wav2vec 2.0 features and dual-stream tdnn for efficient spoken dialect identification,” IEEE Access, vol. 13, pp. 3115–3129, 2025. [find](https://scholar.google.com/scholar?q=A%2E+Angra%2C+H%2E+Muralikrishna%2C+D%2E+A%2E+Dinesh%2C+and+V%2E+Thenkanidiyoor%2C+%E2%80%9CExploring+aggregated+wav2vec+2%2E0+features+and+dual%2Dstream+tdnn+for+efficient+spoken+dialect+identification%2C%E2%80%9D+IEEE+Access%2C+vol%2E+13%2C+pp%2E)
12. A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” in ICML, 2006, p. 369–376. [find](https://scholar.google.com/scholar?q=A%2E+Graves%2C+S%2E+Ferna%CC%81ndez%2C+F%2E+Gomez%2C+and+J%2E+Schmidhuber%2C+%E2%80%9CConnectionist+temporal+classification%3A+labelling+unsegmented+sequence+data+with+recurrent+neural+networks%2C%E2%80%9D+in+ICML%2C+2006%2C+p%2E+369%E2%80%93376%2E)
13. S. Team, “Silero vad: pre-trained enterprise-grade voice activity detector (vad), number detector and language classifier,” [https://github.com/snakers4/silero-vad](https://github.com/snakers4/silero-vad), 2024. [↗](https://github.com/snakers4/silero-vad)
14. S. Shon, A. Ali, Y. Samih, H. Mubarak, and J. Glass, “Adi17: A fine-grained arabic dialect identification dataset,” in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 8244– 8248. [find](https://scholar.google.com/scholar?q=S%2E+Shon%2C+A%2E+Ali%2C+Y%2E+Samih%2C+H%2E+Mubarak%2C+and+J%2E+Glass%2C+%E2%80%9CAdi17%3A+A+fine%2Dgrained+arabic+dialect+identification+dataset%2C%E2%80%9D+in+ICASSP+2020+%2D+2020+IEEE+International+Conference+on+Acoustics%2C+Speech+and+Signal+) [our card](https://md.research.elamur.ai/free/en/papers/salameh-2018-fine-grained-arabic-dialect-identification/)
15. B. Talafha, K. Kadaoui, S. M. Magdy, M. Habiboullah, C. M. Chafei, A. O. El-Shangiti, H. Zayed, M. cheikh tourad, R. Alhamouri, R. Assi, A. Alraeesi, H. Mohamed, F. Alwajih, A. Mohamed, A. E. 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-Mageed, “Casablanca: Data and models for multidialectal arabic speech recognition,” in ACL 2024, 2024. [find](https://scholar.google.com/scholar?q=B%2E+Talafha%2C+K%2E+Kadaoui%2C+S%2E+M%2E+Magdy%2C+M%2E+Habiboullah%2C+C%2E+M%2E+Chafei%2C+A%2E+O%2E+El%2DShangiti%2C+H%2E+Zayed%2C+M%2E+cheikh+tourad%2C+R%2E+Alhamouri%2C+R%2E+Assi%2C+A%2E+Alraeesi%2C+H%2E+Mohamed%2C+F%2E+Alwajih%2C+A%2E+Mohamed%2C+A%2E+E%2E+Mekki%2C+E)
16. M. U. Farooq and O. Saz, “Sage: Spliced-audio generated data for enhancing foundational models in low-resource arabic-english code-switched speech recognition,” in MLSP, 2025. [find](https://scholar.google.com/scholar?q=M%2E+U%2E+Farooq+and+O%2E+Saz%2C+%E2%80%9CSage%3A+Spliced%2Daudio+generated+data+for+enhancing+foundational+models+in+low%2Dresource+arabic%2Denglish+code%2Dswitched+speech+recognition%2C%E2%80%9D+in+MLSP%2C+2025%2E)