Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A phonetic script that encodes every Tajweed rule can be learned by a multi-level CTC model, reaching 0.16% phoneme error on held-out reciters.

desk verdict The dataset and QPS script are real contributions, but the 0.16% PER is self-consistency, not validated error detection; still worth refereeing if the authors fix the evaluation. read the letter →

arxiv 2509.00094 v1 pith:5YJTDADG submitted 2025-08-27 eess.AS cs.AIcs.CLcs.LGcs.SD

Abdullah Abdelfattah , Mahmoud I. Khalil , Hazem Abbas This is my paper
classification eess.AScs.AIcs.CLcs.LGcs.SD
keywords TajweedQuranicrecitationanalysisphoneticscriptdesignpronunciationerrordetectionmulti-levelCTCwav2vec2Arabicspeechprocessingself-supervisedmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that Quranic recitation, with its elaborate Tajweed rules and articulation attributes, can be captured in a compact two-level phonetic script and that this script is learnable by a speech recognition model. The authors train a multi-level CTC model on 850+ hours of expert recitations and report a 0.16% average phoneme error rate on unseen reciters, which they take as proof that the script is internally consistent and learnable. They also show the model can flag Madd, Ghunna, Qalqala, and Tafkheem errors in real samples even though it was trained only on error-free golden recitations. A 98%-automated pipeline for producing such datasets is presented alongside the release of code, data, and models.

What carries the argument

The Quranic Phonetic Script (QPS) is the load-bearing object: a phoneme level (43 symbols representing Arabic letters, vowels, and Tajweed markers) and a Sifat level (10 binary articulation attributes such as hams/jahr, shidda/rakhawa, tafkheem/taqeeq, and ghonna). It is produced from the Uthmani script through 26 sequential regex operations. The second mechanism is the Multi-level CTC architecture: one speech encoder with 11 linear output heads, each trained with its own CTC loss, averaged with a weight of 0.4 on the phoneme head; this lets the model emit phonemes and all ten articulation attributes in parallel from unsegmented audio.

What would settle it

Have several certified Tajweed teachers independently transcribe a set of expert recitations into QPS, then run the paper's pipeline on the same audio; if the pipeline's phoneme error rate relative to the human transcriptions is substantially above 0.16%, the claim of learnability of the script as true phonetic ground truth fails. A complementary test: feed the model deliberately mispronounced recitations with known Madd, Ghunna, Qalqala, and Tafkheem errors and measure detection accuracy against expert judges.

Watch

Extended reading notes

Core claim

The central claim is that the Quranic Phonetic Script (QPS) is learnable: a two-level script that encodes 43 phonemes plus 10 articulation attributes (Sifat) per phoneme, generated deterministically from the Uthmani text by 26 rule-based transformations. A multi-level CTC model with 11 parallel output heads, fine-tuned from a pretrained Wav2Vec2-BERT encoder, reaches 0.16% average phoneme error rate on two held-out complete recitations. The same model, trained only on correct recitations, detects Madd, Ghunna, Qalqala, and Tafkheem errors in real learner recordings, which the authors interpret as evidence that the phonetic representation itself carries enough information for error detection

Load-bearing premise

The 26 rule-based transformations that turn the Uthmani text into QPS labels are treated as correct phonetic ground truth for expert audio; no human-expert validation of these labels is reported, so the 0.16% error rate measures consistency between audio and the authors' own labels, not correctness of the script.

Editorial extensions

If this is right

  • Automatic Tajweed tutoring could give learners per-rule feedback (length of Madd, presence of Ghunna, Qalqala echo, Tafkheem emphasis) from a single utterance without needing a pronunciation-error dataset.
  • The 98%-automated pipeline, applied to other riwayat (recitation traditions) or to the same Hafs variants, can produce large annotated Quranic corpora where only a small manual annotation set is needed.
  • Because the model was trained only on golden recitations, any detected deviation is framed as a potential error; this suggests an unsupervised error-detection path for other strictly well-formed spoken liturgical texts.
  • If the QPS representation is adopted as a standard, different ASR backends can be compared on the same phonetic transcription task, making Quranic recitation a repeatable benchmark for Arabic speech models.
  • The low phoneme error rate on held-out reciters indicates the script is deterministic enough to serve as a target for forced alignment, opening the way to word- or even phoneme-level alignment of recitations to the written text.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 0.16% phoneme error rate should be read as a measure of internal consistency between the audio and the authors' own rule-generated labels; whether those labels match expert-human phonetics remains untested, since the 5,400 manually annotated samples are mentioned but no accuracy figure is reported.
  • Treating each Tajweed rule as a separate prediction head offers a natural diagnostic: if a particular Sifat attribute (e.g., Istitala or Tikraar) is rarely present in the data, its head may be hard to train; the paper itself notes this limitation for attributes applying to single letters.
  • The same multi-head CTC scheme could be applied to Modern Standard Arabic or other liturgical languages with codified pronunciation rules, where the existence of a strict canonical form makes error detection a structured prediction task rather than a free-form assessment.
  • A direct testable extension would be to train on error-containing recitations with human-verified QPS labels and compare detection accuracy against the current error-free-trained model, which would quantify how much of the demonstrated detection ability comes from the script's structure versus from exposure to actual mistakes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents a Quran-specific phonetic script (QPS) that encodes both phonemes and ten articulation attributes (sifa), a largely automated pipeline that converts digitized Uthmani text into QPS labels and segments expert recitations at pause points, and a multi-level CTC model fine-tuned from Wav2Vec2-BERT with 11 parallel output heads. The authors report a 0.16% average phoneme error rate on two held-out reciters and claim this proves the learnability of QPS and enables detection of Tajweed errors such as Madd, Ghunna, Qalqala, and Tafkheem. Code, data, and models are released open-source.

Significance. If the QPS labels were independently validated, the released 850+ hour dataset with Tajweed-rule-level annotations would be a valuable resource for Quranic speech processing and for low-resource Arabic ASR more generally. The two-level encoding of phonemes plus articulatory attributes is a conceptually interesting alternative to IPA-based schemes, and the multi-level CTC architecture is a simple but sensible way to model parallel phonological targets. The authors are also to be credited for committing to open-source release and for providing a highly automated, reproducible data pipeline. However, the central quantitative evidence currently rests on labels generated by the authors' own rule-based phonetizer, without human validation; the error-detection claim is supported only by anecdotal samples, as the authors themselves concede. These gaps are load-bearing for the paper's main claims.

major comments (4)
  1. [Section VI, Table V; Sections III-D and Appendix VIII-A] The 0.16% average PER is computed against QPS labels produced by the authors' 26 regex operations from the Uthmani text, not against independently verified pronunciations. The 5,400 manually annotated samples mentioned in Section IV are never used to report label-level accuracy or inter-annotator agreement. Consequently, the low PER demonstrates that the model reproduces the authors' transcription conventions, not that QPS correctly captures expert-Quranic pronunciation. This directly weakens the claim that the results 'prove the Quranic phonetic script is learnable,' if learnability is intended as a statement about a correct script. Please provide expert validation of a sample of QPS labels, including agreement statistics, or substantially soften the claim to internal consistency.
  2. [Section VI and Section VII] The error-detection claim is anecdotal. The text states that 'we tested some actual samples with errors' in Madd, Ghunna, Qalqala, and Tafkheem and that the model 'was able to detect them,' but no sample size, metrics (precision/recall, F1, detection accuracy), or comparison with a baseline is given. Section VII then states the primary limitation: the dataset contains 'golden recitations with no errors,' limiting evaluation. Since the model was trained only on correct recitations, the mechanism by which it can detect errors is not established. A systematic evaluation on error-containing data, even if small, is essential before the title's 'Error Detection' claim can be supported.
  3. [Appendix VIII-A, operations 11 and 12; Section IV-D] Several phonetization operations are pause-dependent: 'SkoonMostateel' and 'MaddAlewad' add or remove letters depending on whether a pause (waqf) occurs. The segmentation is therefore not merely a pre-processing step; segmentation errors can change the ground-truth labels themselves. The segmenter's reported frame-level F1 of 0.99476 (Table IV) is not a measure of pause-boundary accuracy, and no sensitivity analysis is provided for how boundary shifts affect the QPS labels. This is a load-bearing issue because these labels are used as training targets, and label noise from segmentation may be baked into the model.
  4. [Section V and Table V] The multi-level loss formulation is under-specified: the text says losses are averaged and the phoneme level is assigned a weight of 0.4, but the normalization of the remaining ten levels is not stated. If the weights sum to 1, each sifa level receives 0.06, whereas if 0.4 is used as an additional multiplier, the effective balance is different. The reported 'average_per' also appears to be a simple mean of the per-level PERs, not the weighted optimization objective. Please clarify the loss definition and report how the average PER is computed, since this is the headline number.
minor comments (5)
  1. [Section VII] The sentence 'Consequently, we expect our model will be unable without Istitala without Tikrar' is incomplete and garbled. It should be rephrased to state which rules are expected to be undetectable and why.
  2. [Throughout] Many Arabic strings and phonetic symbols are corrupted or missing (e.g., Tables I, VI, VII; Appendix VIII-A), making it difficult to verify the phoneme and sifa inventories. A PDF preview pass is needed.
  3. [Figure 3] Figure 3, captioned 'VAD architecture vs. standard streaming models,' appears unrelated to the segmenter discussion in Section IV-D. Please either replace the figure with the segmenter architecture or explain its relevance.
  4. [Algorithm 1] The complexity expression 'O(N· W· L2)' should read 'O(N·W·L^2)' and the variable definitions (e.g., 'overlap', 'penalty') should be stated before the loop.
  5. [References] References [24] and [25] are given only in Arabic with no English translation of the titles or venue names. Also, spelling of 'Moshaf'/'Mushaf' and 'Tasmeea'/'Tasmee' is inconsistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the 0.16% PER is a genuine held-out-reciter generalization result; the rule-generated labels are a validity concern, not a circularity.

full rationale

The paper's central quantitative claim is that a multi-level CTC model achieves 0.16% average phoneme error rate on held-out reciters, which the authors interpret as evidence that their Quran Phonetic Script (QPS) is learnable. This is a standard supervised-learning evaluation: the model is trained on audio with QPS labels and tested on reciters (Mushaf 26.1 and 19.0) that were not used in training. The test labels are produced by the same 26-rule phonetizer as the training labels, so the PER measures how well the model can reproduce the authors' own transcription conventions from acoustic input. This is a real validity threat for any claim that QPS captures correct pronunciation or that the model detects actual errors, but it is not circular in the specific sense required here. The 'learnability' claim is exactly the claim that a model can learn the mapping from speech to QPS labels, and that mapping is empirically tested across unseen reciters; the low PER is not forced by construction, since the model could have failed to generalize. No parameter is fitted to the test labels, no prediction is measured against the same data points used for fitting, and no load-bearing self-citation or imported uniqueness theorem is present. The anecdotal error-detection results are weak evidence, and the paper itself acknowledges the lack of error-containing data as a primary limitation, but this is an evidence-quality issue rather than a circularity. Therefore no circular step rises to the level of the defined patterns, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated. The QPS is a symbolic notation, not an empirical entity. The main load-bearing inputs are the assumed correctness of the digitized text, the classical Tajweed rules, and the expert reciters, none of which are independently validated against human phonetic transcription in this paper.

free parameters (4)
  • CTC phoneme weight = 0.4
    Hand-chosen loss weighting for the phoneme level (Section V) to balance the larger vocabulary.
  • Training hyperparameters = learning rate 5e-5, batch size 64, 1 epoch
    Chosen without reported tuning or sensitivity analysis (Section V).
  • Segmenter tuning parameters = threshold, min silence, min speech, padding per Moshaf
    Tuned per Moshaf against EveryAyah ground truth (Section IV-D1); these affect pause boundaries and hence Tajweed label correctness.
  • Tasmeea algorithm parameters = overlap_words=6, window_words=30, acceptance_ratio=0.5
    Hand-set matching parameters in Algorithm 1 that determine which segments are accepted or rejected.
assumptions (5)
  • domain assumption The Tanzil digitized Uthmani Quran text is accurate and canonical
    Used as the foundation for all phonetic labels (Section IV-A).
  • domain assumption The classical Tajweed rules encoded in QPS are correct and complete for Hafs, except Ishmam
    The script is built on classical scholarship and the authors' expert guidance (Section III, Table VII).
  • domain assumption The 22 selected reciters produce error-free golden recitations
    Dataset collection relies on this to create ground truth (Section IV-C).
  • domain assumption The Tasmeea algorithm correctly validates transcription and segmentation
    Used as the verification step for 98% of the dataset (Section IV-F, Algorithm 1).
  • ad hoc to paper A model trained only on correct recitations can detect mispronunciations
    Tested only anecdotally in Section VI; no systematic evaluation supports this premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning." pith.science (2026). https://pith.science/paper/5YJTDADG

@misc{pith2026250900094,
  author       = {Pith},
  title        = {Pith review of: Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5YJTDADG}},
  note         = {Machine review of arXiv:2509.00094}
}
read the original abstract

Assessing spoken language is challenging, and quantifying pronunciation metrics for machine learning models is even harder. However, for the Holy Quran, this task is simplified by the rigorous recitation rules (tajweed) established by Muslim scholars, enabling highly effective assessment. Despite this advantage, the scarcity of high-quality annotated data remains a significant barrier. In this work, we bridge these gaps by introducing: (1) A 98% automated pipeline to produce high-quality Quranic datasets -- encompassing: Collection of recitations from expert reciters, Segmentation at pause points (waqf) using our fine-tuned wav2vec2-BERT model, Transcription of segments, Transcript verification via our novel Tasmeea algorithm; (2) 850+ hours of audio (~300K annotated utterances); (3) A novel ASR-based approach for pronunciation error detection, utilizing our custom Quran Phonetic Script (QPS) to encode Tajweed rules (unlike the IPA standard for Modern Standard Arabic). QPS uses a two-level script: (Phoneme level): Encodes Arabic letters with short/long vowels. (Sifa level): Encodes articulation characteristics of every phoneme. We further include comprehensive modeling with our novel multi-level CTC Model which achieved 0.16% average Phoneme Error Rate (PER) on the testset. We release all code, data, and models as open-source: https://obadx.github.io/prepare-quran-dataset/

Figures

Figures reproduced from arXiv: 2509.00094 by the authors.

Figure 1
Figure 1. Database Collection Statistics [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Reciters Statistics We developed a web GUI using Streamlit5 that: • Downloads and extracts metadata for each track • Organizes data by Moshaf (each chapter as ”001.mp3”) • Annotates Moshaf attributes IV-D Segment Recitations Since Tajweed rules are affected by pauses (وڢژ(, accurate segmentation is crucial. We initially tested open-source Voice Activity Detection (VAD) models including SileroVAD [27] and PyAnnotate … view at source ↗
Figure 3
Figure 3. VAD architecture vs. standard streaming models This work has been submitted to the IEEE International Conference on Intelligent Computing and Systems (ICICoS) 2025 for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Multi-level CTC loss Architecture composed of 11 Heads for every level and CTC loss for every level with weighted average loss [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Recitations lengths in seconds for the whole dataset Tikrar. This limitation similarly applies to Tajweed rules that occur less frequently in the Holy Quran, such as Imala, Rawm, and Tasheel. VIII Conclusion We present a novel approach for assessing pronunciation error…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 48 canonical work pages

  1. [1]

    Automatic pronunciation assessment–a review,

    Y. E. Kheir, A. Ali, and S. A. Chowdhury, “Automatic pronunciation assessment–a review,”arXiv preprint arXiv:2310.13974, 2023

    arXiv 2023

  2. [2]

    Enhancing usability of capl system for quran recitation learning

    M. Sherif, A. Samir, A. Khalil, and R. Mohsen, “Enhancing usability of capl system for quran recitation learning.” INTERSPEECH, 2007

    work page 2007

  3. [3]

    Qdat: adatasetforreciting the quran,

    H.M.Osman, B.S.Mustafa, andY.Faisal, “Qdat: adatasetforreciting the quran,” International Journal on Islamic Applications in Computer Science And Technology, vol. 9, no. 1, pp. 1–9, 2021

    work page 2021

  4. [4]

    Automatic detection of some tajweed rules,

    D. Omran, S. Fawzi, and A. Kandil, “Automatic detection of some tajweed rules,” in2023 20th Learning and Technology Conference (LT) , 2023, pp. 157–160

    work page 2023

  5. [5]

    Evaluation of the pronunciation of tajweed rules based on dnn as a step towards interactive recitation learning,

    D. Shaiakhmetov, G. Gimaletdinova, K. Momunov, and S. Cankurt, “Evaluation of the pronunciation of tajweed rules based on dnn as a step towards interactive recitation learning,” arXiv preprint arXiv:2503.23470, 2025

    work page arXiv 2025

  6. [6]

    The tarteel dataset: crowd-sourced and labeled quranic recitation,

    H. I. Khan, A. Abid, M. M. Moussa, and A. Abou-Allaban, “The tarteel dataset: crowd-sourced and labeled quranic recitation,” 2021

    work page 2021

  7. [7]

    Towards a unified benchmark for arabic pronunciation assessment: Quranic recitation as case study,

    Y. E. Kheir, O. Ibrahim, A. Meghanani, N. Almarwani, H. O. Toyin, S.Alharbi, M.Alfadly, L.Alkanhal, I.Selim, S.Elbatal et al., “Towards a unified benchmark for arabic pronunciation assessment: Quranic recitation as case study,”arXiv preprint arXiv:2506.07722, 2025

    arXiv 2025

  8. [8]

    A computer aided pronunciation learning system for teaching the holy quran recitation rules,

    S. M. Abdou and M. Rashwan, “A computer aided pronunciation learning system for teaching the holy quran recitation rules,” in2014 IEEE/ACS 11th International Conference on Computer Systems and Applications (AICCSA). IEEE, 2014, pp. 543–550

    work page 2014

Show all 55 references
  1. [9]

    Computeraidedqur’anpronunciationusingdnn,

    M. Al-Marri, H. Raafat, M. Abdallah, S. Abdou, and M. Rashwan, “Computeraidedqur’anpronunciationusingdnn,” Journal of Intelligent & Fuzzy Systems , vol. 34, no. 5, pp. 3257–3271, 2018

    2018

  2. [10]

    Recognition of holy quran recitation rules using phoneme duration,

    A. Mohammed, M. S. B. Sunar, and M. S. H. Salam, “Recognition of holy quran recitation rules using phoneme duration,” inInternational Conference of Reliable Information and Communication Technology . Springer, 2017, pp. 343–352

    2017

  3. [11]

    Improving automatic forced alignment for phoneme segmentation in quranic recitation,

    A. M. A. Alqadasi, A. M. Zeki, M. S. Sunar, M. S. B. H. Salam, R. Ab- dulghafor, and N. A. Khaled, “Improving automatic forced alignment for phoneme segmentation in quranic recitation,”IEEE Access, vol. 12, pp. 229–244, 2023

    2023

  4. [12]

    Empirical study on mispronunciation detection for tajweed rules during quran recitation,

    Y. S. Alsahafi and M. Asad, “Empirical study on mispronunciation detection for tajweed rules during quran recitation,” in2024 6th Inter- national Conference on Computing and Informatics (ICCI) , 2024, pp. 39–45

    2024

  5. [13]

    Developing speech recog- nition system for quranic verse recitation learning software,

    B. Putra, B. T. Atmaja, and D. Prananto, “Developing speech recog- nition system for quranic verse recitation learning software,” IJID (International Journal on Informatics for Development) , vol. 1, no. 2, pp. 1–8, 2012

    2012

  6. [14]

    Reducing the dimensionality of data with neural networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,”science, vol. 313, no. 5786, pp. 504–507, 2006

    2006

  7. [15]

    Neural networks and physical systems with emer- gent collective computational abilities,

    J. J. Hopfield, “Neural networks and physical systems with emer- gent collective computational abilities,” Proceedings of the National Academy of Sciences , vol. 79, no. 8, pp. 2554–2558, 1982

    1982

  8. [16]

    Attention is all you need,

    A.Vaswani,N.Shazeer,N.Parmar,J.Uszkoreit,L.Jones,A.N.Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in Neural Information Processing Systems , vol. 30, pp. 5998–6008, 2017

    2017

  9. [17]

    Bert: Pre- training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre- training of deep bidirectional transformers for language understanding,” Proceedings of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL- HLT), pp. 4171...

    2019

  10. [18]

    wav2vec: Unsupervised pre-training for speech recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,” arXiv preprint arXiv:1904.05862, 2019

    1904 arXiv

  11. [19]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems , vol. 33, pp. 12449– 12460, 2020

    2020

  12. [20]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wuet al., “Conformer: Convolution-augmented transformer for speech recognition,”arXiv preprint arXiv:2005.08100 , 2020

    2005 arXiv

  13. [21]

    W2v-bert: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,

    Y.-A. Chung, Y. Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y. Wu, “W2v-bert: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . IEEE, 2021, pp...

    2021

  14. [22]

    Seamless: Multilingual expressive and streaming speech translation,

    L. Barrault, Y.-A. Chung, M. C. Meglioli, D. Dale, N. Dong, M. Dup- penthaler, P.-A. Duquenne, B. Ellis, H. Elsahar, J. Haaheim et al. , “Seamless: Multilingual expressive and streaming speech translation,” arXiv preprint arXiv:2312.05187, 2023

    2023 arXiv

  15. [23]

    Automatic detection of some tajweed rules,

    D. Omran, S. Fawzi, and A. Kandil, “Automatic detection of some tajweed rules,” in 2023 20th Learning and Technology Conference (L&T). IEEE, 2023, pp. 157–160. [24]۱ٴ“, ෛ ੼ ﺍ”༥ ᄭᄥ෠ ੼, vol. 31, pp. 165–190, 02 2023. [25]ﺃ݁, ﺍܳٺ۠ިࢴࣖﺍ .ﺩﺍﺭﺍܳ؞ިٔ؇ 2021. [26]ঌॻ༟,༃຃๤ཛྷ .݁ܝٺٴ۰ﻭ݁ޚٴأ۰݁ݱޚ...

    2023

  16. [24]

    Madd (݁ڎAdds madd symbols for all madd types, inserting madd_alif (ﺍ,)madd_waw (ۥand madd_yaa (ۦ.)

  17. [25]

    Qalqla (ڢܹگ۰Adds echoing effect toﻕ,ﻁ,ﺏ,ﺝ,ﺩ letters with sukoon

  18. [26]

    Algorithm 1 Tasmeea Algorithm Require: text_segments = [ s1, s2,

    RemoveRasHaaAndShadda(ﺇﺯﺍ:) Deletes sukoon diacritic marks. Algorithm 1 Tasmeea Algorithm Require: text_segments = [ s1, s2, . . . , sn], sura_idx, overlap_words = 6 , window_words = 30 , acceptance_ratio = 0.5, flags for special phrases Ensure: List of tuples(match, ratio) pe...

    2025

  19. [27]

    Silerovad: pre-trainedenterprise-gradevoiceactivitydetector (vad), number detector and language classifier,

    S.Team,“Silerovad: pre-trainedenterprise-gradevoiceactivitydetector (vad), number detector and language classifier,” https://github.com/ snakers4/silero-vad, 2024

    2024

  20. [28]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” inProc. INTERSPEECH 2023, 2023

    2023

  21. [29]

    Audiomentations: Apythonlibraryforaudio data augmentation,

    I.JordalandContributors,“Audiomentations: Apythonlibraryforaudio data augmentation,”https://github.com/iver56/audiomentations, 2025

    2025

  22. [30]

    Whisper base arabic quran (automatic speech recognition for quranic recitation),

    T. AI, “Whisper base arabic quran (automatic speech recognition for quranic recitation),” https://huggingface.co/tarteel-ai/ whisper-base-ar-quran, 2023, model by Tarteel AI. Company website: https://www.tarteel.ai/

    2023

  23. [31]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational conference on machine learning . PMLR, 2023, pp. 28492–28518

    2023

  24. [32]

    Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks,

    A.Graves, S.Fernández, F.Gomez, andJ.Schmidhuber, “Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks,” in Proceedings of the 23rd International Conference on Machine Learning (ICML 2006) . ACM, 2006, pp. 369– 376. Appendi...

    2006

  25. [33]

    DisassembleHrofMoqatta (ّڰܝ٭۹Sepa- rates Quranic initials (e.g.,ﺍinto individual letters

  26. [34]

    SpecialCases (༡): Handles special words like ྟ཯ that have different pronunciation forms defined in MoshafAttributes

  27. [35]

    BeginWithHamzatWasl (ﺍܳٴڎﺀProcesses words starting with connecting hamza (ﭐand converts it to hamza (ﺀwith appropriate harakah for nouns and verbs

  28. [36]

    BeginWithSaken (ﺍܳٴڎﺀManages words begin- ning with a consonant (sakin) likeْܳ٭َگْޚَْؕ as Arabic doesn’t start utterances with consonants. Table VI:Phoneme Set (43 Symbols) Phoneme Name Symbol hamzaﺀ baaﺏ taaﺕ thaaﺙ jeemﺝ haa_mohmalaﺡ khaaﺥ daalﺩ thaalﺫ raaﺭ zayﺯ seenﺱ sheenﺵ ...

  29. [37]

    ConvertAlifMaksora (ොູ ): Convertsﻯ in Uthmani script to either yaa (ﻱor alif (ﺍbased on context

  30. [38]

    NormalizeHmazat (ّިۋ٭ڎﺍ Standardizes hamza forms (ﺃﺇﺅﺉ toﺀ

  31. [39]

    IthbatY aaY ohie(ﺇHandles words likeຬ where two yaa letters occur - resolves conflicts when pausing on words with consecutive consonants (ﺍܳٺگ؇ﺀ ﺍܳފ؇܋ٷby adding another yaa at end

  32. [40]

    RemoveKasheeda (ﺇﺯﺍDeletes elongation marks (ـــ) from text

  33. [41]

    Copyright may be transferred without notice, after which this version may no longer be accessible

    RemoveHmzatWaslMiddle(ﺇﺯﺍRe- This work has been submitted to the IEEE International Conference on Intelligent Computing and Systems (ICICoS) 2025 for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Table ...

    2025

  34. [42]

    RemoveSkoonMostadeer (༡ ݁ފٺڎߌߵEliminates letters with circular sukoon diacritics like alif inᅹ

  35. [43]

    SkoonMostateel (ݿܝިﻥ݁ފٺޚ٭ܭ Removes alif with elongatedsukoonmid-wordandaddsitattheendduring pauses (ﻭڢژ.)

  36. [44]

    MaddAlewad (݁ڎﺍܳأިﺽ Removes alif after tanween fatha mid-word and adds alif while removing tanween at pause positions (ﻭڢژ.)

  37. [45]

    WawAlsalah (ﻭﺍﻭﺍܳݱ Replaces letter waw (ﻭwith small alif above combined with alif

  38. [46]

    EnlargeSmallLetters (ّܝٴResizes miniature Arabic letters to standard proportions

  39. [47]

    CleanEnd (਍ಾ): Removes redundant diacritics and spaces at word endings

  40. [48]

    NormalizeTaa (ّިۋ٭ڎﺍܳٺ؇ﺀ Convertsﺓtaa marbuta) to ﺕorﻩbased on context, and converts finalﺓto haa (ﻩ.)

  41. [49]

    AddAlifIsmAllah (ﺇݪ؇ڣ۰ﺃܳژﺍ Inserts compen- satory alif in derivatives of ”ﺍ.”

  42. [50]

    PrepareGhonnaIdghamIqlab (ዛኤ ﻭﺍPreprocesses text for nasalization, assimilation, and conversion rules

  43. [51]

    IltiqaaAlsaknan (ﺍܳٺگ؇ﺀﺍܳފ؇܋ٷ Resolves consecutive consonants by inserting vowels

  44. [52]

    DeleteShaddaAtBeginning (༡ ): Re- moves shadda (ّfrom word-initial letters

  45. [53]

    Ghonna (ؗٷ۰Applies nasalization during pronuncia- tion of sakin noon and tanween

  46. [54]

    Tasheel (ዝ๎฽): Adds a letter representing alif with tasheel easing

  47. [55]

    Imala (ﺇ݁؇Converts fatha with imala to fatha_momala phoneme and alif with imala to alif_momala phoneme

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.