Elamur Research⌕
☰
home / papers / 2025 / Poem Meter Classification of Recited Arabic Poetry: Integrating High-Resource Systems for a Low-Resource Task

Poem Meter Classification of Recited Arabic Poetry: Integrating High-Resource Systems for a Low-Resource Task

Summary

The paper tackles automatic identification of the meter (the rhythmic pattern studied in Arabic prosody, Arud) of Arabic poetry when the input is recorded recitation rather than written text. Labelled recited verse is scarce, so a model trained directly on audio has little to learn from. The authors compare two Wav2Vec2-based designs. The first maps sound straight to a meter label. The second is a pipeline: a speech model fine-tuned on poetry transcribes the verse, and an existing text-based meter classifier, trained on large written corpora, labels the transcript. The direct audio model performed poorly because there was too little training data. The pipeline produced state-of-the-art results, with F1 of 80% on a private Baseline set and 87% on a newly released public Benchmark. An ablation with gold transcripts shows how much of the error comes from each stage. The work matters because it shows a low-resource task can be solved by joining two components that each benefit from abundant data, and because the public benchmark gives future research a shared test set.

Key points

  • Arabic classical poetry has 16 meters. Earlier acoustic work (Al-Talabani) covered only 3 meters and used 230 verses with an LSTM.
  • Both systems build on Wav2Vec2. The transcription step compares Wav2Vec-Base with Wav2Vec-FineTuned, and the text step uses classifiers such as the one from Abandah et al. Related textual work relies on the MetRec dataset (55,400 verses, 14 meters).
  • The best pipeline reaches F1 80% (Baseline) and 87% (Benchmark). The end-to-end audio classifier lagged clearly behind.
  • With perfect transcriptions, F1 rises to 87% on Baseline and 89% on Benchmark (accuracy 89% and 95%). Transcription errors therefore cost about 7 F1 points on Baseline and about 2 points on Benchmark.
  • Meters are still recognised robustly even when WER exceeds 10% and CER sits at 2–5%. Rare but similar meters (Mudhare vs. Mujtath) are confused, while frequent meters are almost always correct.

Interesting ideas

  • I1. Instead of training one model end to end on tiny data, route the task through an intermediate representation (text) that lets two separately data-rich systems be chained. — For Mova (learn): When a target task has few labels, Mova can decompose it into stages that each have rich training sources and connect them through explicit, inspectable intermediate structures such as UD trees and frames.
  • I2. An oracle ablation feeds gold intermediate output into the downstream stage, splitting total error into upstream and downstream shares. — For Mova (learn): Mova's evaluation could run frames and world building on gold UD trees as well as predicted ones, attributing each lost point to the parser or to the later stages.
  • I3. The downstream classifier tolerated quite noisy transcripts, which suggests that measuring robustness to upstream noise is as important as measuring upstream accuracy. — For Mova (frames): Feed deliberately degraded or alternative parses into crate prag and check whether the who-did-what-to-whom frames stay stable, which shows where repair effort actually pays off.
  • I4. Fine-tuning the speech model on poetry seemed to make it treat poetry less like prose, which points to domain adaptation of the upstream component. — For Mova (tree): gparse could be fine-tuned on in-genre text such as fables or school texts, then checked for fewer genre-specific attachment errors that propagate into later stages.
  • I5. The confusion analysis showed that errors cluster in rare classes that are close to more frequent neighbours. — For Mova (world): When assigning VerbNet classes or ontology types, inspect confusions between rare, similar classes and add targeted silver data or explicit discriminating rules for those pairs.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2504.12172 · arXiv 2504.12172

@misc{alshaibani2025poem,
  title = {{Poem Meter Classification of Recited Arabic Poetry: Integrating High-Resource Systems for a Low-Resource Task}},
  author = {Al-Shaibani, Maged S. and Alyafeai, Zaid and Ahmad, Irfan},
  year = {2025},
  doi = {10.48550/arxiv.2504.12172},
  eprint = {2504.12172},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2504.12172}
}

References

64 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abandah GA, Khedher MZ, Abdel-Majeed MR, et al (2020) Classifying and diacritizing arabic poems using deep recurrent neural networks. Journal of King Saud University-Computer and Information Sciences find
  2. Abdel-Hamid O, Mohamed Ar, Jiang H, et al (2012) Applying convolutional neural networks concepts to hybrid nn-hmm model for speech recognition. In: 2012 IEEE international conference on Acoustics, speech and signal processing (ICASSP), IEEE, pp 4277–4280 find
  3. Abdel-Hamid O, Mohamed Ar, Jiang H, et al (2014) Convolutional neural networks for speech recognition. IEEE/ACM Transactions on audio, speech, and language processing 22(10):1533–1545 find
  4. Abdelhamid AA, Alsayadi HA, Hegazy I, et al (2020) End-to-end arabic speech recognition: A review. https://www.researchgate.net/profile/Hamzah-Alsayadi/publication/344799361 End-to-End Arabic Speech Recognition A Review/links/5f90babfa6fdccfd7b74575d/ End-to-End-Arabic-Speech-Recognition-A-Review.pdf, accessed: 25-08-2021 ↗
  5. Agirrezabal M, Alegria I, Hulden M (2016) Machine learning for metrical analysis of english poetry. In: Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pp 772–781 find
  6. Ahmed A, Hifny Y, Shaalan K, et al (2019) End-to-end lexicon free arabic speech recognition using recurrent neural networks. In: Computational Linguistics, Speech And Image Processing For Arabic Language. World Scientific, p 231–248 find
  7. Ahmed AF, Mohamed R, Mostafa B, et al (2015) Authorship attribution in arabic poetry. In: 2015 10th international conference on intelligent systems: theories and applications (SITA), IEEE, pp 1–6 find
  8. Al-Anzi F, AbuZeina D (2018) Literature survey of arabic speech recognition. In: 2018 International Conference on Computing Sciences and Engineering (ICCSE), IEEE, pp 1–6 find
  9. Al-shaibani MS, Alyafeai Z, Ahmad I (2020) Meter classification of arabic poems using deep bidirectional recurrent neural networks. Pattern Recognition Letters 136:1–7 find
  10. Al-Shaibani MS, Alyafeai Z, Ahmad I (2020) Metrec: A dataset for meter classification of arabic poetry. Data in Brief 33:106,497 find
  11. Al-Talabani AK (2020) Automatic recognition of arabic poetry meter from speech signal using long short-term memory and support vector machine. ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY 8(1):50–54 find
  12. Algihab W, Alawwad N, Aldawish A, et al (2019) Arabic speech recognition with deep learning: A review. In: International Conference on Human-Computer Interaction, Springer, pp 15–31 find
  13. Ali A, Bell P, Glass J, et al (2016) The mgb-2 challenge: Arabic multi-dialect broadcast media recognition. In: 2016 IEEE Spoken Language Technology Workshop (SLT), IEEE, pp 279–284 find our card
  14. Ali A, Vogel S, Renals S (2017) Speech recognition challenge in the wild: Arabic mgb-3. In: 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), IEEE, pp 316–322 find our card
  15. Ali A, Shon S, Samih Y, et al (2019) The mgb-5 challenge: Recognition and dialect identification of dialectal arabic speech. In: 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), IEEE, pp 1026–1033 find
  16. Alsayadi HA, Abdelhamid AA, Hegazy I, et al (2021) Arabic speech recognition using end-to-end deep learning. IET Signal Processing 15(8):521– 534 find
  17. Alsharif O, Alshamaa D, Ghneim N (2013) Emotion classification in arabic poetry using machine learning. International Journal of Computer Applications 65(16) find
  18. Alyafeai Z, AlShaibani MS, Ahmad I (2020) A survey on transfer learning in natural language processing. arXiv preprint arXiv:200704239 find our card
  19. Alyafeai Z, Al-shaibani MS, Ghaleb M, et al (2022) Evaluating various tokenizers for arabic text classification. Neural Processing Letters pp 1–23 find our card
  20. Amodei D, Ananthanarayanan S, Anubhai R, et al (2016) Deep speech 2: End-to-end speech recognition in english and mandarin. In: International conference on machine learning, PMLR, pp 173–182 find our card
  21. Ardila R, Branson M, Davis K, et al (2019) Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:191206670 find our card
  22. Baevski A, Zhou H, Mohamed A, et al (2020) wav2vec 2.0: A framework for self-supervised learning of speech representations. arXiv preprint arXiv:200611477 find our card
  23. Bakheet M (2021) Improving speech recognition for arabic language using low amounts of labeled data. https://www.diva-portal.org/smash/get/diva2:1567188/FULLTEXT01.pdf, accessed: 25-08-2021 ↗
  24. Brown TB, Mann B, Ryder N, et al (2020) Language models are few-shot learners. arXiv preprint arXiv:200514165 find
  25. Dahl GE, Yu D, Deng L, et al (2011) Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition. IEEE Transactions on audio, speech, and language processing 20(1):30–42 find
  26. De Mori R, Lam L (1986) Plan refinement in a knowledge-based system for automatic speech recognition. In: ICASSP’86. IEEE International Conference on Acoustics, Speech, and Signal Processing, IEEE, pp 1217–1220 find
  27. Devlin J, Chang MW, Lee K, et al (2018) Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:181004805 find
  28. Dhouib A, Othman A, El Ghoul O, et al (2022) Arabic automatic speech recognition: A systematic literature review. Applied Sciences 12(17):8898 find
  29. Elnagar A, Yagi SM, Nassif AB, et al (2021) Systematic literature review of dialectal arabic: identification and detection. IEEE Access 9:31,010– 31,042 find
  30. Graves A, Fernández S, Gomez F, et al (2006) Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: Proceedings of the 23rd international conference on Machine learning, pp 369–376 find
  31. Haider T (2021) Metrical tagging in the wild: Building and annotating poetry corpora with rhythmic features. arXiv preprint arXiv:210208858 find
  32. Halabi N (2016) Modern standard arabic phonetics for speech synthesis. PhD thesis, University of Southampton find
  33. Hamidi S, Razzazi F, Ghaemmaghami MP (2009) Automatic meter classification in persian poetries using support vector machines. In: 2009 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), IEEE, pp 563–567 find
  34. Hannun A, Case C, Casper J, et al (2014) Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:14125567 find our card
  35. Heafield K (2011) Kenlm: Faster and smaller language model queries. In: Proceedings of the sixth workshop on statistical machine translation, pp 187–197 find our card
  36. Jitsi (????) jitsi/jiwer: Evaluate your speech-to-text system with similarity measures such as word error rate (wer). https://github.com/jitsi/jiwer, (Accessed on 02/16/2023) ↗
  37. Kuhn R, De Mori R (1992) Corrections to ä cache-based language model for speech recognition.̈ IEEE Transactions on Pattern Analysis & Machine Intelligence 14(06):691–692 find
  38. Lee KF (1988) Large-vocabulary speaker-independent continuous speech recognition: The SPHINX system. Carnegie Mellon University find
  39. Liu X, Zhang F, Hou Z, et al (2021) Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering find
  40. Manna HS, Arifin Z (2021) Metrics in arabic poetry. nternational Journal of West Asian Studies find
  41. Morris AC, Maier V, Green P (2004) From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. In: Eighth International Conference on Spoken Language Processing find
  42. Mubarak H, Hussein A, Chowdhury SA, et al (2021) Qasr: Qcri aljazeera speech resource–a large scale annotated arabic speech corpus. arXiv preprint arXiv:210613000 find
  43. Nassif AB, Shahin I, Attili I, et al (2019) Speech recognition using deep neural networks: A systematic review. IEEE access 7:19,143–19,165 find
  44. Pan SJ, Yang Q (2009) A survey on transfer learning. IEEE Transactions on knowledge and data engineering 22(10):1345–1359 find
  45. Paoli B (2001) Meters and formulas: The case of ancient arabic poetry. Belgian journal of linguistics 15(1):113–136 find
  46. Povey D, Ghoshal A, Boulianne G, et al (2011) The kaldi speech recognition toolkit. In: IEEE 2011 workshop on automatic speech recognition and understanding, IEEE Signal Processing Society, CONF find
  47. Rabiner L, Levinson S, Rosenberg A, et al (1979) Speaker-independent recognition of isolated words using clustering techniques. IEEE Transactions on Acoustics, Speech, and Signal Processing 27(4):336–349 find
  48. Radford A, Wu J, Child R, et al (2019) Language models are unsupervised multitask learners. OpenAI blog 1(8):9 find
  49. Radford A, Kim JW, Xu T, et al (2022) Robust speech recognition via large-scale weak supervision. Tech. rep., Tech. Rep., Technical report, OpenAI find our card
  50. Radha V, Vimala C (2012) A review on speech recognition challenges and approaches. doaj org 2(1):1–7 find
  51. Rajan R, Raju AA (2018) Poetic meter classification using acoustic cues. In: 2018 International Conference on Signal Processing and Communications (SPCOM), IEEE, pp 31–35 find
  52. Rajan R, Kumar AV, Babu BP (2020) Poetic meter classification using i-vector-mtf fusion. In: INTERSPEECH, pp 145–149 find
  53. Rajan R, Chandrika Reghunath L, Varghese LT (2022) Pomet: a corpus for poetic meter classification. Language Resources and Evaluation 56(4):1131–1152 find
  54. Saleh AZAK, Elshafei M (2012) Arabic poetry meter identification system and method. US Patent 8,219,386 find
  55. Schneider S, Baevski A, Collobert R, et al (2019) wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:190405862 find
  56. Soliman AB (2021) bakrianoo/sinai-voice-ar-stt · hugging face. https://huggingface.co/bakrianoo/sinai-voice-ar-stt, (Accessed on 12/01/2021) ↗
  57. tsurumeso (2022) tsurumeso/vocal-remover: Vocal remover using deep neural networks. https://github.com/tsurumeso/vocal-remover, (Accessed on 10/08/2022) ↗
  58. Vaessen N, van Leeuwen DA (2022) Training speaker recognition systems with limited data. arXiv preprint arXiv:220314688 find
  59. Vaswani A, Shazeer N, Parmar N, et al (2017) Attention is all you need. Advances in neural information processing systems 30 find
  60. Yi C, Wang J, Cheng N, et al (2020) Applying wav2vec2. 0 to speech recognition in various low-resource languages. arXiv preprint arXiv:201212121 find
  61. Yousef WA, Ibrahime OM, Madbouly TM, et al (2018) Poem comprehensive dataset (pcd). URL https://hci-lab.github.io/ ArabicPoetry-1-Private/#PCD ↗
  62. Yousef WA, Ibrahime OM, Madbouly TM, et al (2019) Learning meters of arabic and english poems with recurrent neural networks: a step forward for language understanding and synthesis. arXiv preprint arXiv:190505700 find
  63. Zhao L, Han Z (2010) Speech recognition system based on integrating feature and hmm. In: 2010 international conference on measuring technology and mechatronics automation, IEEE, pp 449–452 find
  64. Ziyovuddinova M (2021) Arud system in view of metric theory. The American Journal of Applied sciences 3(05):234–239 find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/al-2025-shaibani-poem-meter-classification-recited-arabic/