---
title: "End-to-end Jordanian dialect speech-to-text self-supervised"
description: "The authors build an end-to-end transformer pipeline on Wav2Vec 2.0. 29 references. Among them: 5."
authors: ["Ali A. Safieh", "Ibrahim Abu Alhaol", "Rawan Ghnemat"]
year: 2022
venue: "arXiv:2609.24410 · Front. Robot. AI 9:1090012 (2022)"
source: "https://arxiv.org/abs/2609.24410"
doi: "10.3389/frobt.2022.1090012"
arxiv: "2609.24410"
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/safieh-2026-end-end-jordanian-dialect-speech/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/safieh-2026-end-end-jordanian-dialect-speech/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2022](https://md.research.elamur.ai/free/en/papers/2022/page/6/) / End-to-end Jordanian dialect speech-to-text self-supervised learning framework

# End-to-end Jordanian dialect speech-to-text self-supervised learning framework

**Authors:** Ali A. Safieh, Ibrahim Abu Alhaol and Rawan Ghnemat · **Year:** 2022 · **Venue:** arXiv:2609.24410 · Front. Robot. AI 9:1090012 (2022) · **Publisher:** [Frontiers Media](https://md.research.elamur.ai/free/en/publishers/frontiers-media/) · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://doi.org/10.3389/frobt.2022.1090012) · [PDF](https://arxiv.org/pdf/2609.24410) · [DOI 10.3389/frobt.2022.1090012](https://doi.org/10.3389/frobt.2022.1090012) · [arXiv 2609.24410](https://arxiv.org/abs/2609.24410) · [OpenAlex](https://openalex.org/W4312187715)

**Citations:** 5 citations (OpenAlex, 2026-10-14)

## Summary
The paper tackles automatic speech recognition for Jordanian Arabic, a spoken dialect with little transcribed audio and phonology that differs from Modern Standard Arabic and across urban, rural and bedouin varieties. The authors build an end-to-end transformer pipeline on Wav2Vec 2.0. They start from a cross-lingual self-supervised model pretrained on 53 languages and fine-tune it on dialect audio gathered from Jordanian videos. Transcripts for that audio came either from the Google speech engine or from manual work. They then run noisy student training. A teacher labels a new batch of unlabeled audio, and a larger student is trained on clean, pseudo-labeled and augmented data together. Each student becomes the next teacher. Over five generations, WER on a held-out set falls from 0.568 to 0.515, which the authors describe as roughly a 5-point absolute or 8% relative gain over the fine-tuned baseline. The work matters because it shows that unlabeled audio is useful after pretraining as well as before it. It also yields a Jordanian spoken dataset and a reusable recipe for low-resource languages.

## Key points
- The labeled training set has 12.5 h of speech. Unlabeled audio is split into five batches of 2.5 h, one batch per generation. All audio is resampled to 16 kHz, with clips averaging about 6 s.
- 20% of the data is held out as a validation set and is never used in any self-training round.
- WER/CER by generation: teacher 0.568/0.279, then 0.547/0.278, 0.537/0.274, 0.534/0.274, 0.523/0.279, and finally 0.515/0.264 at the fifth generation. WER improves at every step, but CER does not (it bumps up in generation four).
- Training ran on two NVIDIA Quadro RTX 8000 GPUs. The authors name compute cost as the main limit: each generation means a new, bigger model, so they could not go past five rounds.
- A small table illustrates the dialect split. The same word, for example "killed", "dog" or "how", is pronounced differently in Standard Arabic and in each Jordanian variety.

## Interesting ideas
- Each self-training round uses a fresh slice of unlabeled data instead of relabeling the same pool, so every generation's gain can be attributed to new material.
- Student batches deliberately mix clean gold, pseudo-labeled and augmented examples, so pseudo-labels are always anchored to trusted data.
- Rough labels from an external engine are used to speed up human annotation, so people correct drafts instead of transcribing from scratch.
- The paper reports a coarse metric (WER) and a finer one (CER) per generation, and the finer one shows a non-monotonic dip that the headline number hides.
- A model trained on many standard varieties is adapted to a dialect with only hours of data, treating dialect forms as variants of known units.

## For Mova

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

- **I1** (learn): Silver-data rounds could each pull a new unseen text batch for the witness parsers, which would make per-round improvement on gold traceable and keep the model from feeding on its own old outputs.
- **I2** (learn): When training gparse or the perceptron parser on agreed silver trees, a fixed share of gold treebank sentences and perturbed copies could be kept in every batch to limit drift.
- **I3** (learn): LLM teacher tiers could pre-draft UD trees or frame annotations for humans to correct, which cuts the cost of building new gold for dialects or new domains.
- **I4** (learn): Self-training evaluation could track both LAS and finer measures such as UAS or per-relation accuracy per round, so a hidden regression shows up even when the main score rises.
- **I5** (tree): For UD dialect conversion, a parser trained on standard English treebanks could be adapted with a small dialect set plus self-training, and variant surface forms could be mapped to shared lemmas in the en crate.

*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

- [Princess Sumaya University for Technology](https://md.research.elamur.ai/free/en/universities/princess-sumaya-university-for-technology/) — University, Amman, Jordan · 3 works · 53 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 213 works

🧭 Topics (2): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 1,484 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 233 works

## Cite

Cite as: Ali A. Safieh, Ibrahim Abu Alhaol and Rawan Ghnemat. 2022. End-to-end Jordanian dialect speech-to-text self-supervised learning framework. arXiv:2609.24410 · Front. Robot. AI 9:1090012 (2022).

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

```bibtex
@misc{safieh2022end,
  title = {{End-to-end Jordanian dialect speech-to-text self-supervised learning framework}},
  author = {Ali A. Safieh, Ibrahim Abu Alhaol and Rawan Ghnemat},
  year = {2022},
  doi = {10.3389/frobt.2022.1090012},
  eprint = {2609.24410},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.3389/frobt.2022.1090012}
}
```

- [← previous Building a Manually Annotated Hungarian Coreference Corpus: Workflow and Tools](https://md.research.elamur.ai/free/en/papers/vadasz-2022-building-manually-annotated-hungarian-coreference/)
- [next → Parameter and Data Efficient Continual Pre-training for Robustness to Dialectal Variance in Arabic](https://md.research.elamur.ai/free/en/papers/sarkar-2022-parameter-data-efficient-continual-pre/)
- [↑ back to papers of 2022](https://md.research.elamur.ai/free/en/papers/2022/page/6/)

## References

29 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Al-Fetyani, M., Al-Barham, M., Abandah, G., Alsharkawi, A., and Dawas, M. (2021). MASC: Massive Arabic Speech Corpus. IEEE Dataport. doi:10.21227/e1qb-jv46 [↗](https://doi.org/10.21227/e1qb-jv46)
2. Natural Language: Proceedings of a workshop held at pacific grove, California, Pacific Grove, California, February 19–22, 1991. [find](https://scholar.google.com/scholar?q=Natural+Language%3A+Proceedings+of+a+workshop+held+at+pacific+grove%2C+California%2C+Pacific+Grove%2C+California%2C+February+19%E2%80%9322%2C+1991%2E)
3. Kannan, A. (2022). Large-scale multilingual speech recognition with a streaming end-to-end model. [find](https://scholar.google.com/scholar?q=Kannan%2C+A%2E+%282022%29%2E+Large%2Dscale+multilingual+speech+recognition+with+a+streaming+end%2Dto%2Dend+model%2E)
4. Amodei, D., Ananthanarayanan, S., Anubhai, R., Bai, J., Battenberg, E., Case, C., et al. (2016). “[Deep speech 2: End-to-end speech recognition in English and Mandarin](https://md.research.elamur.ai/free/en/papers/amodei-2015-deep-speech-2-end-end/),” in International conference on machine learning, New York City, NY, June 19–24, 2016. 173–182. [↗](https://doi.org/10.48550/arxiv.1512.02595)
5. Karpagavalli, S., and Chandra, E. (2016). A review on automatic speech recognition architecture and approaches. Int. J. Signal Process. Image Process. Pattern Recognit. 9, 393–404. doi:10.14257/ijsip.2016.9. 4.34 [↗](https://doi.org/10.14257/ijsip.2016.9)
6. Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. (2020). [wav2vec 2.0: A framework for self-supervised learning of speech representations](https://md.research.elamur.ai/free/en/papers/baevski-2020-wav2vec-2-0-framework-self/). Adv. Neural Inf. Process. Syst. 33, 12449
7. Karray, F., Alemzadeh, M., Abou Saleh, J., and Arab, M. N. (2008). Human-computer interaction: Overview on state of the art. Int. J. smart Sens. intelligent Syst. 1, 137–159. doi:10.21307/ijssis-2017-283 [↗](https://doi.org/10.21307/ijssis-2017-283)
8. Braun, S., Neil, D., and Liu, S.-C. (2017). “A curriculum learning method for improved noise robustness in automatic speech recognition,” in 2017 25th European Signal Processing Conference (EUSIPCO) (IEEE), 548. [↗](https://doi.org/10.5281/zenodo.1159271)
9. Kessler, S., Thomas, B., and Karout, S. (2021). Continual-wav2vec2: An application of continual learning for self-supervised automatic speech recognition. arXiv preprint. arXiv:2107.13530. [↗](https://arxiv.org/abs/2107.13530)
10. Chaudhary, A. (2022). Self supervised representation learning in nlp. [find](https://scholar.google.com/scholar?q=Chaudhary%2C+A%2E+%282022%29%2E+Self+supervised+representation+learning+in+nlp%2E)
11. Conneau, A., Baevski, A., Collobert, R., Mohamed, A., and Auli, M. (2020). Unsupervised cross-lingual representation learning for speech recognition. Interspeech. arXiv preprint arXiv:2006.13979. [↗](https://arxiv.org/abs/2006.13979)
12. Lamere, P., Kwok, P., Gouvea, E., Raj, B., Singh, R., Walker, W., et al. (2003). “The cmu sphinx-4 speech recognition system,” in Ieee intl. conf. on acoustics, speech and signal processing (icassp 2003), Hong Kong, 1, 2. [find](https://scholar.google.com/scholar?q=Lamere%2C+P%2E%2C+Kwok%2C+P%2E%2C+Gouvea%2C+E%2E%2C+Raj%2C+B%2E%2C+Singh%2C+R%2E%2C+Walker%2C+W%2E%2C+et+al%2E+%282003%29%2E+%E2%80%9CThe+cmu+sphinx%2D4+speech+recognition+system%2C%E2%80%9D+in+Ieee+intl%2E+conf%2E+on+acoustics%2C+speech+and+signal+processing+%28icassp+20)
13. Fendji, J. L. K. E., Tala, D. C. M., Yenke, B. O., and Atemkeng, M. (2022). Automatic speech recognition using limited vocabulary: A survey. Appl. Artif. Intell. 36, 2095039. doi:10.1080/08839514.2022.2095039 [↗](https://doi.org/10.1080/08839514.2022.2095039)
14. MacKenzie, I. S., and Soukoreff, R. W. (2002). “A character-level error analysis technique for evaluating text entry methods,” in Proceedings of the second Nordic conference on Human-computer interaction, 243. [find](https://scholar.google.com/scholar?q=MacKenzie%2C+I%2E+S%2E%2C+and+Soukoreff%2C+R%2E+W%2E+%282002%29%2E+%E2%80%9CA+character%2Dlevel+error+analysis+technique+for+evaluating+text+entry+methods%2C%E2%80%9D+in+Proceedings+of+the+second+Nordic+conference+on+Human%2Dcomputer+interact)
15. Gondi, S. (2022). Wav2vec2. 0 on the edge: Performance evaluation. arXiv preprint arXiv:2202.05993. [↗](https://arxiv.org/abs/2202.05993)
16. Mošner, L., Wu, M., Raju, A., Parthasarathi, K., Kumatani, K., Sundaram, S., et al. (2019). “Improving noise robustness of automatic speech recognition via parallel data and teacher-student learning,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brighton, United Kingdom, May 12–17, 2019, 6475–6479. doi:10.1109/ICASSP [↗](https://doi.org/10.1109/icassp)
17. Gondi, S., and Pratap, V. (2021). Performance evaluation of offline speech recognition on edge devices. Electronics 10, 2697. doi:10.3390/electronics10212697 [↗](https://doi.org/10.3390/electronics10212697)
18. Halabi, N. (2016). Modern standard Arabic phonetics for speech synthesis. Ph.D. thesis. University of Southampton. [find](https://scholar.google.com/scholar?q=Halabi%2C+N%2E+%282016%29%2E+Modern+standard+Arabic+phonetics+for+speech+synthesis%2E+Ph%2ED%2E+thesis%2E+University+of+Southampton%2E)
19. Nakayama, H., Kubo, T., Kamura, J., Taniguchi, Y., and Liang, X. (2018). doccano: Text annotation tool for human. Available at: [https://github.com/doccano/](https://github.com/doccano/) doccano. [↗](https://github.com/doccano/)
20. Han, W., Chan, C.-F., Choy, C.-S., and Pun, K.-P. (2006). “An efficient mfcc extraction method in speech recognition,” in 2006 IEEE International Symposium on Circuits and Systems (ISCAS) (IEEE). [find](https://scholar.google.com/scholar?q=Han%2C+W%2E%2C+Chan%2C+C%2E%2DF%2E%2C+Choy%2C+C%2E%2DS%2E%2C+and+Pun%2C+K%2E%2DP%2E+%282006%29%2E+%E2%80%9CAn+efficient+mfcc+extraction+method+in+speech+recognition%2C%E2%80%9D+in+2006+IEEE+International+Symposium+on+Circuits+and+Systems+%28ISCAS%29+%28IEEE%29%2E)
21. Park, D. S., Zhang, Y., Jia, Y., Han, W., Chiu, C.-C., Li, B., et al. (2020). Improved noisy student training for automatic speech recognition. arXiv preprint arXiv:2005.09629. [↗](https://arxiv.org/abs/2005.09629)
22. Jelinek, F., Merialdo, B., Roukos, S., and Strauss, M. (1991). “A dynamic language model for speech recognition,” in Speech and Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., et al. (2011). “The kaldi speech recognition toolkit,” in IEEE 2011 workshop on automatic speech recognition and understanding (IEEE Signal Processing Society), CONF. [find](https://scholar.google.com/scholar?q=Jelinek%2C+F%2E%2C+Merialdo%2C+B%2E%2C+Roukos%2C+S%2E%2C+and+Strauss%2C+M%2E+%281991%29%2E+%E2%80%9CA+dynamic+language+model+for+speech+recognition%2C%E2%80%9D+in+Speech+and+Povey%2C+D%2E%2C+Ghoshal%2C+A%2E%2C+Boulianne%2C+G%2E%2C+Burget%2C+L%2E%2C+Glembek%2C+O%2E%2C+Goel%2C+N%2E)
23. Yalta, N. (2020). Robot audition framework using deep learning techniques. Ph.D. thesis. doi:10.13140/RG.2.2.24456 [↗](https://doi.org/10.13140/rg.2.2.24456)
24. Yi, C., Wang, J., Cheng, N., Zhou, S., and Xu, B. (2020). Applying wav2vec2. 0 to speech recognition in various low-resource languages. arXiv preprint arXiv:2012.12121. [↗](https://arxiv.org/abs/2012.12121)
25. Sakarna, A. K. (1999). Phonological aspects of 9abady Arabic: A bedouin Jordanian dialect. The University of Wisconsin-Madison. [find](https://scholar.google.com/scholar?q=Sakarna%2C+A%2E+K%2E+%281999%29%2E+Phonological+aspects+of+9abady+Arabic%3A+A+bedouin+Jordanian+dialect%2E+The+University+of+Wisconsin%2DMadison%2E)
26. Schlüter, J., and Grill, T. (2015). Exploring data augmentation for improved singing voice detection with neural networks. ISMIR, 121. [↗](https://doi.org/10.5072/zenodo.243314)
27. Zhai, X., Oliver, A., Kolesnikov, A., and Beyer, L. (2019). “S4l: Self-supervised semi-supervised learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, South Korea, October 27–November 2, 2019, 1476–1485. wiseman (2022). py-webrtcvad. [https://github.com/wiseman/xpy-webrtcvad](https://github.com/wiseman/xpy-webrtcvad). [↗](https://github.com/wiseman/xpy-webrtcvad)
28. Xie, Q., Luong, M.-T., Hovy, E., and Le, Q. V. (2020). “Self-training with noisy student improves imagenet classification,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Seattle, WA, June 13–19, 2020. 10687–10698. [find](https://scholar.google.com/scholar?q=Xie%2C+Q%2E%2C+Luong%2C+M%2E%2DT%2E%2C+Hovy%2C+E%2E%2C+and+Le%2C+Q%2E+V%2E+%282020%29%2E+%E2%80%9CSelf%2Dtraining+with+noisy+student+improves+imagenet+classification%2C%E2%80%9D+in+Proceedings+of+the+IEEE%2FCVF+conference+on+computer+vision+and+pattern+rec)
29. Zhang, Y., Qin, J., Park, D. S., Han, W., Chiu, C.-C., Pang, R., et al. (2020). Pushing the limits of semi-supervised learning for automatic speech recognition. arXiv preprint arXiv:2010.10504. [↗](https://arxiv.org/abs/2010.10504)

*Updated 24 October 2026*