Pith. sign in

REVIEW 5 major objections 8 minor 1 cited by

From Statistical Methods to Pre-Trained Models; A Survey on Automatic Speech Recognition for Resource Scarce Urdu Language

T0 review · 5 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The survey aims to be the first comprehensive review of Urdu automatic speech recognition, mapping datasets, toolkits, and reported word error rates across monolingual and multilingual systems.

desk verdict Useful first map of Urdu ASR, but the headline numbers disagree inside the paper, so the 'comprehensive systematic' claim is not yet supportable. read the letter →

arxiv 2411.14493 v1 pith:3BMNUKL3 submitted 2024-11-20 cs.CL cs.SDeess.AS

Muhammad Sharif , Zeeshan Abbas , Jiangyan Yi , Chenglin Liu This is my paper · ORCID
classification cs.CLcs.SDeess.AS
keywords UrduASRautomaticspeechrecognitionresource-scarcelanguagedatasetsmultilingualtransferlearningpre-trainedmodelssystematicreview
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 sets out to establish that Urdu automatic speech recognition has a recordable research landscape, and that this survey is the first comprehensive systematic review of it. It claims that Urdu, spoken by over 100 million people but short on annotated speech data, has been studied mainly through small datasets, statistical HMM-GMM models, and more recently pre-trained multilingual models. The paper collects the available Urdu speech corpora, the toolkits used to build systems, and the word-error rates reported, and organizes them into monolingual and multilingual categories. A sympathetic reader would value this as a single entry point into a scattered literature: it says where the data is, which tools work, and where the field has stalled.

What carries the argument

The organizing device is a two-way taxonomy: every Urdu ASR study is classified by data type (isolated digits, read speech, spontaneous speech, broadcast, telephone) and by algorithm family (statistical HMM-GMM/SGMM, DNN variants, end-to-end, multilingual pre-trained). The accompanying tables align each study with its toolkit (Sphinx, Kaldi, SRILM, KenLM, ESPnet, Wav2Vec2/XLSR, Whisper) and its reported accuracy, and the discussion reads those tables as evidence that data scarcity, not algorithm choice, is the main constraint on Urdu ASR.

What would settle it

A bibliographic check for an earlier comprehensive Urdu ASR survey published before 2024, or the discovery of a large public Urdu corpus beyond 100 hours that is absent from the paper's dataset tables, would falsify the claim to be the first comprehensive review.

Watch

Extended reading notes

Core claim

The paper's central claim is that no earlier work systematically reviewed Urdu ASR as a whole, and that this survey fills that gap by summarizing the datasets, tools, algorithms, and results reported across two decades. It divides monolingual Urdu ASR into traditional statistical approaches (HMM-GMM, SGMM, SVM), deep neural network approaches (TDNN, LSTM, BLSTM, CNN, DNN-HMM), and end-to-end approaches, and treats multilingual systems built on Common Voice, Shrutilipi, IndicSUPERB, Vistaar, FLEURS, Whisper, and XLSR as the route for transfer learning. The best Urdu-specific result it documents comes from a self-supervised Wav2Vec2 model with KenLM at 12.6% word error rate, while Whisper and XLSR fine-tuned on Urdu reach 17.5% and comparable accuracies. On the paper's own account, end-to-end techniques remain mostly untried for Urdu because the available datasets are too small, and the most promising direction is fine-tuning large multilingual pre-trained models.

Load-bearing premise

The load-bearing premise is that the paper's literature search actually found all or nearly all relevant Urdu ASR work: Section 3 names no databases, date range, screening counts, or inclusion criteria, so the 'first comprehensive systematic review' claim rests on the keyword combination 'Urdu Speech Recognition' being sufficient.

Editorial extensions

If this is right

  • A new researcher can use the survey's dataset and toolkit tables to reproduce the field's baselines without repeating the search.
  • The comparison shows that self-supervised multilingual models currently give Urdu's lowest word error rates, suggesting transfer learning rather than new statistical modeling is the productive path.
  • Because most large Urdu corpora are private, public data like Common Voice and FLEURS will remain the practical foundation for reproducible Urdu ASR unless new open corpora appear.
  • End-to-end ASR for Urdu stays unexplored, so the survey implies that building a sufficiently large open Urdu corpus would unlock a class of methods not yet evaluated on this language.

Reading between the lines

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

  • If the survey's coverage holds, the most urgent next step is not another model but a standardized public evaluation set: the WER figures across tables were produced on different vocabularies and test splits, so they are indicative rather than directly comparable.
  • The same comparison suggests a concrete testable prediction: a Whisper or XLSR model fine-tuned on a combined 300+ hour Urdu corpus with a KenLM language model should beat the reported 12.6% WER monolingual baseline.
  • The paper's own remarks about dialect and orthography variation imply that a single national Urdu benchmark will likely mislead; a dialect-stratified evaluation (Pakistani, Indian, and diaspora Urdu) would be a more informative yardstick.
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

5 major / 8 minor

Summary. This paper surveys automatic speech recognition (ASR) for Urdu, covering the transition from HMM-GMM systems to deep neural networks and to multilingual pre-trained models such as Whisper and XLSR. The manuscript is organized into sections on motivation, background, methodology, Urdu datasets, tools, monolingual and multilingual ASR approaches, comparison tables, discussion, open challenges, and conclusion. Its central claim, repeated in Sections 1.2, 7.2, and 8, is that it is the first comprehensive systematic review of Urdu ASR. The qualitative storyline is plausible and consistent with the cited literature in outline, but the quantitative comparison tables contain unresolved internal contradictions, and Section 3 does not document a reproducible literature search.

Significance. If the internal inconsistencies are corrected and the search protocol is documented, the survey would be a useful starting point for researchers entering Urdu ASR: it collects scattered dataset descriptions, toolkit choices, and accuracy figures into one document, and it identifies open problems such as lack of standardized orthography, limited transfer learning, and missing end-to-end systems. The claimed novelty as the first Urdu-specific ASR survey is plausible but not yet evidenced, because the authors do not demonstrate that the literature collection was exhaustive or that no prior survey covers the same ground. I therefore view the contribution as defensible in outline but currently blocked by data-quality problems in the central tables.

major comments (5)
  1. [Sec. 4.4; Table 2; Table 4] The Whisper Urdu WER is internally inconsistent: Section 4.4 reports 17.5% WER, Table 2 row 15 reports 22.6 for 'Whisper Large-V2', and Table 4 row 4 reports 22.6 on FLEURS and 24.2 on Common Voice 9. Because Whisper is the strongest pre-trained model discussed and this number is the principal evidence for its suitability, the authors must trace each value to its primary source, state the evaluation condition (dataset split, use of language model, fine-tuning), and reconcile the text with the tables.
  2. [Sec. 4.1.3; Table 1; Table 5] The FLEURS Urdu audio size is reported in three incompatible ways: 1.4 hours in Section 4.1.3, 8.5 hours in Table 1 row 23, and 12 hours per language in Table 5. This affects how readers judge the resource status of Urdu within FLEURS, so it is not a formatting slip. The authors should consult the FLEURS source and give a single figure with explicit split information (e.g., train/validation/test) rather than three different totals.
  3. [Table 2; Sec. 4.5] Table 2 row 14 reports the XLSR-Wav2Vec2 result as '0.49' without specifying whether this is WER, CER, accuracy, or some other metric, and Section 4.5 describes the same study without providing any quantitative result. A unitless value cannot be compared with the other rows in the table; the authors must state the metric and verify the Urdu-specific value against the primary source.
  4. [Sec. 3] The methodology does not currently support the claim of a comprehensive systematic review. Section 3 states that the keyword combination 'Urdu Speech Recognition' 'efficiently helped to get most of the Urdu ASR studies across years', but it names no bibliographic databases, no date range, no screening counts, and no inclusion or exclusion criteria, and Figure 2 is a generic four-box diagram. I ask for a reproducible description of the search and screening process, and for an explicit demonstration that no earlier Urdu-specific ASR survey exists (e.g., by discussing how the authors positioned the work relative to Besacier et al. 2014, which already surveys ASR for under-resourced languages including Urdu).
  5. [Table 3; Table 4] The comparison tables mix accuracy percentages with WER values without a consistent metric column, which makes higher-is-better versus lower-is-better ambiguous. For example, Table 3 lists 'SVM, CNN 0.97' next to 'HMM 74%' in the same column, and Table 4 row 5 lists FLEURS CER values of 82.9/83.1 that Table 5 later calls '%CER reductions'. The authors should standardize the tables by adding separate columns for metric and direction, and verifying each entry against the original publication.
minor comments (8)
  1. [Sec. 2.2] The sentence beginning 'Traditionally researchers of ASR have explored several techniques, including Gaussian Mixture Models - Hidden Markov Models' is duplicated verbatim at the start of two consecutive paragraphs; one copy should be removed.
  2. [Sec. 6] The paragraph beginning 'The insights derived and summarized in this review are poised to captivate...' appears twice, almost verbatim, in the discussion; remove the duplicate.
  3. [Sec. 4.1] Subsection numbering is duplicated: '4.1.1' and '4.1.3' each appear twice. The subsections for Common Voice and FLEURS should be renumbered so the dataset discussion has a clean sequence.
  4. [Table 1] The S.No column skips entries 11-14, jumping from 10 to 15; renumber the rows sequentially.
  5. [References] Several references are duplicated with identical author-year labels, including two Ashraf (2010) entries, two Bhogale (2023) entries, and two Khan (2021) entries; disambiguate these with letter suffixes and update the in-text citations accordingly.
  6. [Eq. (1)] The typeset WER formula appears garbled in the manuscript text; please check that it reads (S+D+I)/N × 100.
  7. [Sec. 4.1.3] Whisper is described as a 'massive dataset' in the text; Whisper is a model trained on a large weakly supervised dataset, so the phrasing should be corrected.
  8. [Introduction] There are small terminology slips: 'Discrete Cosine Transform DTS' should likely be DCT, and 'Short Fourier Transforms SFT' is usually called the short-time Fourier transform (STFT).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; this literature survey performs no derivation whose output is fed back as evidence.

full rationale

This is a literature survey, not a derivation chain. It contains no equations that produce a numeric result subsequently used as evidence, no fitted parameters, and no prediction that is equivalent to an input by construction. The central claim that this is 'the first comprehensive systematic review' is an assertion about literature coverage, not a conclusion derived from the papers it summarizes; even if that claim is debatable, it is not circular in the sense of reducing to its own inputs. The Section 3 methodology states that the keyword combination 'Urdu Speech Recognition' 'efficiently helped to get most of the Urdu ASR studies across years,' but an under-documented search protocol is a completeness and reproducibility limitation, not a circular step. Likewise, the reported inconsistencies in Whisper WER, FLEURS hours, and the unlabeled XLSR '0.49' entry are correctness and data-quality problems that undermine the comprehensiveness claim, but they do not constitute self-definition, fitted-input-as-prediction, or self-citation load-bearing reasoning. No author self-citation appears in the reference list, no uniqueness theorem is imported from the authors' own prior work, and no known result is renamed and presented as a derivation. The honest verdict is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This survey performs no experiments, so it has no fitted free parameters and no invented entities. Its load-bearing assumptions are about literature coverage and the fidelity of compiled figures: (1) the single keyword search in Section 3 retrieves essentially all relevant Urdu ASR studies, so 'comprehensive' holds; (2) the WER/CER values copied from cited papers are accurate and comparable across different datasets, splits, and evaluation protocols, so the comparative tables are meaningful; and (3) the assertion that no prior Urdu-ASR-specific survey exists. Assumptions (1) and (2) are the ones most likely to be silently breached: a missed corpus or a misread metric corrupts the tables without leaving a trace in the text. Assumption (3) is asserted self-referentially ('to the best of our knowledge') and is a bibliographic claim the paper does not verify against a systematic registry of surveys.

assumptions (3)
  • domain assumption The keyword search described in Section 3 retrieves essentially all relevant Urdu ASR studies, making the review 'comprehensive'.
    Section 3 says the combination 'Urdu Speech Recognition' 'efficiently helped to get most of the Urdu ASR studies across years', but no databases, date ranges, screening counts, or inclusion criteria are reported, so coverage is assumed, not demonstrated.
  • domain assumption The WER, CER, and accuracy figures compiled in Tables 2-5 are accurate and comparable across different datasets, speaker splits, and evaluation protocols.
    The survey's comparative tables (Section 5) list single error-rate values per system without normalizing for dataset or setup; the internal conflicts (Whisper 17.5 vs 22.6/24.2; FLEURS 1.4 vs 12 hours) show this assumption is fragile.
  • ad hoc to paper No prior comprehensive Urdu-ASR-specific survey exists, supporting the 'first' claim.
    Section 1.2 asserts this 'to the best of our knowledge'; the paper cites broader surveys (Besacier 2014, Daud 2017) but does not systematically rule out overlapping surveys of Urdu or South Asian ASR.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Statistical Methods to Pre-Trained Models; A Survey on Automatic Speech Recognition for Resource Scarce Urdu Language." pith.science (2026). https://pith.science/paper/3BMNUKL3

@misc{pith2026241114493,
  author       = {Pith},
  title        = {Pith review of: From Statistical Methods to Pre-Trained Models; A Survey on Automatic Speech Recognition for Resource Scarce Urdu Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BMNUKL3}},
  note         = {Machine review of arXiv:2411.14493}
}
read the original abstract

Automatic Speech Recognition (ASR) technology has witnessed significant advancements in recent years, revolutionizing human-computer interactions. While major languages have benefited from these developments, lesser-resourced languages like Urdu face unique challenges. This paper provides an extensive exploration of the dynamic landscape of ASR research, focusing particularly on the resource-constrained Urdu language, which is widely spoken across South Asian nations. It outlines current research trends, technological advancements, and potential directions for future studies in Urdu ASR, aiming to pave the way for forthcoming researchers interested in this domain. By leveraging contemporary technologies, analyzing existing datasets, and evaluating effective algorithms and tools, the paper seeks to shed light on the unique challenges and opportunities associated with Urdu language processing and its integration into the broader field of speech research.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Assessing the Feasibility of Lightweight Whisper Models for Low-Resource Urdu Transcription

    cs.CL 2025-08 conditional novelty 3.0 of 10

    Whisper-Small achieves a 33.68% word error rate on a 36-sample Urdu dataset, outperforming Tiny (67.08%) and Base (53.67%) in zero-shot transcription.

Reference graph

Works this paper leans on

9 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abadi, M. (2016 ). TensorFlow: learning functions at scale. Proceedings of the 21st ACM SIGPLAN international conference on functional programming (pp. 1-1). Abbas, F. P. (2018). The competing status of Urdu and English after declaration of Urdu as official language in Pakistan. Journal of Research (Urdu), pp.142-158. Adeeba, F. a. (2018). Acoustic featur...

    work page 2018

  2. [2]

    Deep Neural Network Acoustic Models for ASR (Doctoral dissertation, University of Toronto

    (2014). Deep Neural Network Acoustic Models for ASR (Doctoral dissertation, University of Toronto. Mohiuddin, H. A. (2023 ). UrduSpeakXLSR: Multilingual Model for Urdu Speech Recognition. 18th International Conference on Emerging Technologies (ICET) (pp. 217-221). IEEE. Morgan, D. S. (1991). Neural networks and speech processing . . Springer US., 329-348....

    work page 2014

  3. [11]

    Farooq, M. A. (2019). Improving Large Vocabulary Urdu Speech Recognition System Using Deep Neural Networks. INTERSPEECH, 2978-2982. Farooq, M. A. (2020). Enhancing Large Vocabulary Continuous Speech Recognition System for Urdu- English Conversational Code-Switched Speech. 2020 23rd Conference of the Oriental COCOSDA International Committee for the Co-ordi...

    work page 2019

  4. [28]

    Zaheer, N. A. (2023). Speech emotion recognition for the Urdu language. . Lang Resources & Evaluation, 57, 915–944. Zehra, W. J. (2021). Cross corpus multi-lingual speech emotion recognition using ensemble learning. . Complex & Intelligent Systems, 1-10. Zhang, S. R. (2022). Opt: Open pre-trained transformer language models. . arXiv preprint arXiv:2205.01...

    arXiv 2023

  5. [161]

    Javed, T. B. (2022). IndicSUPERB: A Speech Processing Universal Performance Benchmark for Indian languages. . arXiv preprint arXiv:2208.11761. Joshi, V. Z. ( 2020.). Transfer learning approaches for streaming end-to-end speech recognition system. arXiv preprint arXiv:2008.05086. Kanabur, V. H. (2019). An extensive review of feature extraction techniques, ...

    arXiv 2022

  6. [262]

    Reitmaier, T. W. (2022, April). Opportunities and challenges of automatic speech recognition systems for low-resource language speakers. . Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems , 1-17. Russell, R. (1999). Urdu in India since independence. . Economic and Political Weekly, 44-48. Sailor, H. P. (2018). Advances in Low R...

    arXiv 1999

  7. [1018]

    Watanabe, S. T. (2018). Espnet: End-to-end speech processing toolkit. arXiv preprint arXiv:1804.00015(arXiv). Xu, B. L. (2020). Discriminative multi-modality speech recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 14433-14442. Xu, J. T. (2020, August). Lrspeech: Extremely low-resource speech synthesis and re...

    arXiv 2018

  8. [1956]

    Ali, H. A. (2016). Urdu speech corpus and preliminary results on speech recognition. Aberdeen, UK: In Engineering Applications of Neural Networks: 17th International Conference, EANN, Proceedings 17 (pp. 317-325). Springer International Publishing. Ali, H. J. (2015). Automatic speech recognition of Urdu digits with optimal classification approach. . Inter...

    work page 2016

Show all 9 references
  1. [8842]

    Chen, Y. Q. (2022 ). Adversarial Meta Learning Improves Low-Resource Speech Recognition. 6th Asian Conference on Artificial Intelligence Technology (ACAIT) (pp. 1-7). IEEE. Chen, Y. Z. (2023). Task-based Meta Focal Loss for Multilingual Low-Resource Speech Recognition. ACM Tra...

    2023 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.