Pith. sign in

REVIEW 5 major objections 5 minor 32 references

Benchmarking Akan ASR Models Across Domain-Specific Datasets: A Comparative Evaluation of Performance, Scalability, and Adaptability

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Seven Akan speech-recognition models perform best only in the domain they were trained on, and Whisper and Wav2Vec2 make different, visible kinds of errors on unfamiliar speech.

desk verdict Useful first cross-dataset Akan ASR benchmark, but the domain-dependency claim is confounded by corpus differences, and the Common Voice in-domain failure plus missing confidence intervals weaken it. read the letter →

arxiv 2507.02407 v1 pith:BSQA5QRF submitted 2025-07-03 cs.CL cs.LGcs.SDeess.AS

Mark Atta Mensah , Isaac Wiafe , Akon Ekpezu , Justice Kwame Appati , Jamal-Deen Abdulai , Akosua Nyarkoa Wiafe-Akenten , Frank Ernest Yeboah , Gifty Odame This is my paper · ORCID
classification cs.CLcs.LGcs.SDeess.AS
keywords automaticspeechrecognitionAkanlow-resourcelanguagesdomaingeneralizationcross-datasetvalidationWhisperWav2Vec2worderrorrate
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 asks whether seven existing Akan speech-recognition models can recognize Akan speech outside the specific corpus each was fine-tuned on. Testing them on four Akan datasets—spontaneous image descriptions, crowdsourced informal clips, formal Bible readings, and short financial phrases—it finds that every model performs best on its own training domain and degrades sharply elsewhere. It also finds that Whisper and Wav2Vec2 fail in different ways: Whisper produces fluent, plausible-sounding Akan even when the transcript is wrong, while Wav2Vec2 produces fragmented, obviously incomplete output. The paper concludes that domain mismatch, not just data scarcity, is a central obstacle for low-resource Akan ASR, and that architecture choice should consider which error type is safer for the user.

What carries the argument

The machinery is a cross-dataset evaluation protocol: seven publicly available Akan models—Whisper-small, Whisper-large, and Wav2Vec2-XLS-R variants—are run unchanged on test splits from four Akan corpora representing distinct speech domains, with WER and CER computed after uniform text normalization. The protocol treats the four corpora as a matrix of training-domain and test-domain combinations, so every model's in-domain score can be compared with its out-of-domain scores. A qualitative error taxonomy (phoneme confusion, out-of-vocabulary terms, dialect mismatch, insertions/omissions, lexical substitution) separates Whisper's fluent-but-wrong errors from Wav2Vec2's fragmented ones, and the sequence-to-sequence vs. CTC distinction explains why the two architectures err differently.

What would settle it

Run a controlled replication in which the same architecture is fine-tuned on equal-sized, separately balanced subsets of the four Akan corpora with identical hyperparameters, then cross-evaluated on the same test splits. If in-domain error rates stop beating out-of-domain rates, or if Whisper and Wav2Vec2 no longer show fluent-versus-fragmented error styles, the paper's central claims would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that Akan ASR models are domain-bound: a model fine-tuned on one Akan corpus reaches its best word error rate on that corpus and loses accuracy, often dramatically, on the others. For example, the financial-dialogue model drops from about 10% WER in-domain to roughly 86% on spontaneous speech and 98% on crowdsourced clips, and the UGSpeechData-trained Whisper model moves from 30% in-domain to 77% on Common Voice and above 100% on financial dialogues, a pattern the paper calls decoder collapse. The second claim is that the two architectures have distinct error profiles: Whisper's built-in language model makes its mistakes fluent and semantically plausible—'Nebukadnezar' becomes 'Neburankan'—while Wav2Vec2's CTC decoding produces transparently broken fragments such as 'NebukaX.' These findings are offered as the first cross-dataset evaluation of Akan transformer ASR, with the practical consequence that deployment choices should weigh domain specialization against general robustness, and error readability against the risk of misleadingly fluent output.

Load-bearing premise

The load-bearing assumption is that the error-rate gaps are caused by domain mismatch rather than by the models' differences in training-data size, corpus quality, architecture, and undisclosed hyperparameters; Section 5.4 concedes that these factors were not isolated.

Editorial extensions

If this is right

  • No existing Akan ASR model should be assumed to work outside the speech domain it was trained on; teams should evaluate on target-domain audio before deployment.
  • Domain-specialized models should be paired with broadly trained fallback models or adaptive routing that selects the best model for the detected domain.
  • Deployment decisions should weigh error interpretability: Wav2Vec2's obvious failures are easier for humans to catch and correct, while Whisper's fluent errors are riskier in medical, legal, or financial transcription.
  • Cross-dataset validation, not single-corpus benchmarking, should become the standard way to report low-resource ASR performance.
  • Akan-English code-switching remains a major unsolved failure mode for both architectures, so language identification plus multilingual transcription is a priority for future models.

Reading between the lines

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

  • If the fluent-error result generalizes, a practical next step is to measure human error-detection rates on Whisper vs. Wav2Vec2 transcriptions; the safer architecture may be the one whose mistakes humans catch more often, even if raw WER is higher.
  • The domain-dependency conclusion could be tested more cleanly by training the same architecture on equal-sized, balanced subsets of the four corpora; the paper itself notes its datasets differ in size and quality, so a controlled replication would show whether domain or data quantity drives the gap.
  • A routing or ensemble system that classifies incoming audio by domain and then selects the specialist model, with a generalist fallback, is an implicit solution the paper recommends but does not implement.
  • The same four-corpus matrix could be applied to neighboring low-resource West African languages to see whether the domain-bound pattern and architecture error split are general or specific to Akan.
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 / 5 minor

Summary. The paper reports a cross-dataset benchmarking study of seven publicly available Akan ASR models based on Whisper and Wav2Vec2 architectures, evaluated on four Akan speech corpora: UGSpeechData, Common Voice, the Lagyamfi Akan Bible corpus, and a financial inclusion speech dataset. Performance is measured with WER and CER, and the authors interpret the results as showing strong domain dependency, with models performing best on their training domains and degrading on mismatched domains. The paper also claims a qualitative difference in error types: Whisper tends to produce fluent but incorrect transcriptions, while Wav2Vec2 produces fragmented and more obviously erroneous outputs. The authors use these findings to argue for domain adaptation and architecture-aware selection in low-resource ASR deployment.

Significance. If the central claims were fully supported, the paper would make a useful contribution to low-resource ASR evaluation by moving beyond single-dataset reporting, using real public corpora, and proposing a practical trade-off between readability and transparency of ASR errors. The scope is systematic: four domains, seven models, uniform preprocessing, and standard metrics are used, which is a strength. However, the main domain-dependency claim is currently confounded by dataset identity and size, and the paper's own Common Voice results contain a clear counterexample. The qualitative error-style claim is supported only anecdotally. The work is therefore a promising benchmark study whose interpretation requires substantial revision.

major comments (5)
  1. [Section 3.2, Sections 4.1-4.4] Section 3.2 states that WER and CER were 'accompanied by the standard deviation and 95% confidence intervals,' but Sections 4.1-4.4 and Figures 1-4 report only point estimates. Because the UGSpeechData test set contains only 120 utterances and the Common Voice evaluation involves only nine speakers, variability estimates are essential for interpreting cross-model differences. Please add the promised SD/CIs or remove the claim.
  2. [Section 4.2] The Common Voice result directly contradicts the abstract's claim that models perform optimally only within their training domains: Model 4, which was trained on Common Voice, obtains a WER of 82%, worse than Models 1, 6, 7, and 3 on the same test set. This is compounded by Section 2.2, which reports Model 4's in-domain WER as 30%. The paper acknowledges the underperformance but does not reconcile it with the central claim; please explain the discrepancy and qualify the domain-dependency conclusion accordingly.
  3. [Sections 3.1 and 5.4] The comparison changes domain simultaneously with dataset identity, total duration (about 2 to 107 hours), number of speakers, utterance length, recording conditions, and the original test splits used for evaluation. Observed degradation on out-of-domain data therefore cannot be uniquely attributed to domain mismatch. The paper itself concedes in Section 5.4 that 'variability in the dataset size and quality... limited the validity of architectural comparisons.' A fully crossed evaluation with matched test conditions, or at least an explicit analysis separating domain from corpus confounds, is needed before the central claim is established.
  4. [Sections 3.3 and 5.3] The claim that Whisper produces fluent-but-misleading errors while Wav2Vec2 produces transparent-but-fragmented errors rests on two anecdotal examples ('Nebukadnezar' being transcribed as 'Neburankan' versus 'NebukaX'). No systematic error annotation, sample size, error-type counts, or inter-annotator agreement is reported. Please support this load-bearing qualitative claim with a quantitative error taxonomy over a representative sample of outputs.
  5. [Sections 4.2-4.4] The evaluation matrix is incomplete in ways that weaken architecture-level conclusions: Model 2 is not evaluated on Common Voice or the Bible corpus, Model 1 is excluded from the financial dataset, and Models 2 and 3 produce WER exceeding 100% on the financial dataset. The paper should either complete the matrix or explicitly mark missing cells and discuss how the missingness affects the claimed architecture differences and domain-dependency patterns.
minor comments (5)
  1. [Section 4.2] The text reports 'WER: 767%' for Model 3 on Common Voice; this appears to be a typo for 76%, consistent with Figure 2. Please correct it.
  2. [Section 3.1] The text cites 'Common Voice Corpus 18.0' but reference [16] is labeled 'Common Voice 17.0'; please align the version and citation.
  3. [Section 3.1] The description of the financial inclusion dataset says a 10-hour Akan subset was extracted, yet the same paragraph states that 'original training-test splits provided by the dataset authors were maintained.' Please clarify how the subset relates to the original splits.
  4. [Figure 4] Figure 4 appears to cap error rates at 1.00 even though the text states that Models 2 and 3 exceeded 100% WER; please report the actual values or indicate truncation.
  5. [References] Reference [13] duplicates reference [5] (both are the same Yoruba speech corpus paper); please remove the duplicate and renumber.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency found: the benchmark's measurements are external and the central claims are falsifiable, not definitional.

full rationale

The paper reports an empirical cross-dataset benchmark; there is no fitted parameter that is later renamed a prediction and no derivation chain in which an output equals an input by construction. The central claim that Akan ASR models show domain dependency rests on direct WER/CER measurements on the original held-out test splits of four corpora. These measurements are external to any definition: in-domain performance is not asserted to be best by construction, and indeed the paper reports a counterexample (Model 4, trained on Common Voice, has WER 82% on Common Voice, worse than several out-of-domain models), which shows the claim is falsifiable rather than circular. Self-citations are present (UGSpeechData [6] and the authors' private models), but they are the objects of evaluation, not load-bearing evidence for a conclusion: the same conclusions are also supported by the public Common Voice, Bible, and Financial Inclusion datasets and public Hugging Face models. The Section 5.4 limitation statement that dataset size and quality variability "may have introduced biases and limited the validity of architectural comparisons" is a validity caveat about confounded corpora, not a circular step; it weakens causal attribution but does not make any result equivalent to its inputs. The qualitative error-style claim rests on small anecdotal samples, but that is an evidence-strength concern, not circularity. Overall, no step reduces to its own input, so the circularity score is 0.

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

This empirical benchmark introduces no fitted parameters, no new entities, and no mathematical derivation. Its conclusions rest on dataset assumptions and evaluation choices, chiefly the representativeness of the four domains and the validity of comparing models trained under very different conditions.

assumptions (4)
  • domain assumption WER and CER computed on lowercased, punctuation-stripped transcripts are reliable proxies for Akan ASR quality.
    Section 3.2 states that no language-specific normalization for tone marks or diacritics was applied; if the reference orthography differs from model output in these aspects, error rates could be inflated.
  • domain assumption The four selected corpora represent four distinct speech domains.
    Section 3.1 classifies UGSpeechData as generic spontaneous speech, Common Voice as informal crowdsourced speech, Bible as formal read speech, and Financial as semi-spontaneous dialogue; this categorization is asserted rather than demonstrated.
  • domain assumption The Hugging Face checkpoints and the original train/test splits used in this evaluation are consistent with the training data each model claims.
    Section 3.1 and 3.2 rely on original splits and publicly available checkpoints, but for Models 2, 3, 6, and 7 the training details or exact datasets are undisclosed or private, so overlap between training and test cannot be independently verified.
  • domain assumption WER and CER differences across models are meaningfully comparable despite different architectures, parameter counts, and training corpus sizes.
    Section 2 and Section 5.4 acknowledge large variability in architecture and data; the comparison implicitly assumes that normalization to the same evaluation protocol is sufficient for cross-model comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Akan ASR Models Across Domain-Specific Datasets: A Comparative Evaluation of Performance, Scalability, and Adaptability." pith.science (2026). https://pith.science/paper/BSQA5QRF

@misc{pith2026250702407,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Akan ASR Models Across Domain-Specific Datasets: A Comparative Evaluation of Performance, Scalability, and Adaptability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BSQA5QRF}},
  note         = {Machine review of arXiv:2507.02407}
}
read the original abstract

Most existing automatic speech recognition (ASR) research evaluate models using in-domain datasets. However, they seldom evaluate how they generalize across diverse speech contexts. This study addresses this gap by benchmarking seven Akan ASR models built on transformer architectures, such as Whisper and Wav2Vec2, using four Akan speech corpora to determine their performance. These datasets encompass various domains, including culturally relevant image descriptions, informal conversations, biblical scripture readings, and spontaneous financial dialogues. A comparison of the word error rate and character error rate highlighted domain dependency, with models performing optimally only within their training domains while showing marked accuracy degradation in mismatched scenarios. This study also identified distinct error behaviors between the Whisper and Wav2Vec2 architectures. Whereas fine-tuned Whisper Akan models led to more fluent but potentially misleading transcription errors, Wav2Vec2 produced more obvious yet less interpretable outputs when encountering unfamiliar inputs. This trade-off between readability and transparency in ASR errors should be considered when selecting architectures for low-resource language (LRL) applications. These findings highlight the need for targeted domain adaptation techniques, adaptive routing strategies, and multilingual training frameworks for Akan and other LRLs.

Figures

Figures reproduced from arXiv: 2507.02407 by the authors.

Figure 1
Figure 1. Comparative analysis of model performance on UGSpeechData corpus These findings reinforce the importance of domain alignment in low-resource ASR systems. In particular, the drop in performance of models trained outside UGSpeechData confirms that narrow-domain ASR models struggle to generalize. Furthermore, the qualitative error analysis showed that frequent errors included substitution of acoustically similar Akan w… view at source ↗
Figure 2
Figure 2. Comparative analysis of model performance on the Common Voice Akan corpus Model 7 performed comparably to Models 3 and 4 (WER: 74%, CER: 26%), suggesting the moderate transferability of its acoustic representations despite the domain mismatch. Again, Model 5 which had the worst performance with the UGSpeechData corpus, recorded the weakest results (WER: 98%, CER: 65%). This indicates a consistent failure to generali… view at source ↗
Figure 3
Figure 3. Comparative analysis of model performance on the Akan Bible corpus 4.4. Financial Inclusion Speech Dataset The evaluation of ASR models on the Financial Inclusion Speech corpus exhibited some of the most divergent performances across all datasets, except for the in-domain trained model. Model 5, which was originally trained with this dataset, achieved the lowest WER of approximately 10% and CER of approximately 6%. … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparative analysis of model performance on the Financial Inclusion Speech Corpus 0.37 0.47 0.70 0.70 0.78 0.95 0.11 0.12 0.22 0.23 0.65 0.64 Model 1 Model 4 Model 6 Model 7 Model 3 Model 5 WER CER 0.10 0.86 0.86 0.95 1.00 1.00 0.06 0.39 0.41 0.47 1.00 1.00 Model 5 Mo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 31 canonical work pages

  1. [1]

    Automatic speech recognition for under-resourced languages: A survey

    Besacier L, Barnard E, Karpov A, Schultz T. Automatic speech recognition for under-resourced languages: A survey. Speech Commun 2014;56:85–100

    work page 2014

  2. [2]

    Automatic Speech Recognition: A Comprehensive Survey

    Rista A, Kadriu A. Automatic Speech Recognition: A Comprehensive Survey. SEEU Review [Internet] 2020;15:86 –112. Available from: https://www.sciendo.com/article/10.2478/seeur-2020-0019 Benchmarking Akan ASR Models 15

    work page doi:10.2478/seeur-2020-0019 2020

  3. [3]

    CUTE: A Multilingual Dataset for Enhancing Cross - Lingual Knowledge Transfer in Low -Resource Languages [Internet]

    Zhuang W, Sun Y. CUTE: A Multilingual Dataset for Enhancing Cross - Lingual Knowledge Transfer in Low -Resource Languages [Internet]. In: Proceedings ofthe 31st International Conference on Computational Linguistics. Abu Dhabi, UAE: Association for Computational Linguistics; 2025 [cited 2025 Jun 23]. page 10037–46.Available from: https://aclanthology.org/2...

    work page 2025

  4. [4]

    Development of Hausa dataset a baseline for speech recognition

    Ibrahim UA, Boukar MM, Suleiman MA. Development of Hausa dataset a baseline for speech recognition. Data Brief [Internet] 2022;40:107820. Available from: https://doi.org/10.1016/j.dib.2022.107820

    work page arXiv 2022

  5. [5]

    Developing an open-source corpus of yoruba speech

    Gutkin A, Demirsahin I, Kjartansson O, Rivera C, Túbòsún K. Developing an open-source corpus of yoruba speech. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH 2020;2020-Octob:404–8

    work page 2020

  6. [6]

    UGSpeechData

    Wiafe I, Abdulai JD, Ekpezu AO, Helegah RD, Atsakpo ED, Nutrokpor C, et al. UGSpeechData. 2023

    work page 2023

  7. [7]

    RSC: A romanian read speech corpus for automatic speech recognition

    Georgescu AL, Cucu H, Buzo A, Burileanu C. RSC: A romanian read speech corpus for automatic speech recognition. LREC 2020 - 12th International Conference on Language Resources and Evaluation, Conference Proceedings 2020;6606–12

    work page 2020

  8. [8]

    A Cookbook for Community-driven Data Collection of Impaired Speech in Low- Resource Languages

    Salihs SA, Wiafe I, Abdulai JD, Doe Atsakpo E, Ayoka G, Cave R, et al. A Cookbook for Community-driven Data Collection of Impaired Speech in Low- Resource Languages. In: August, editor. Accepted - InterSpeech Conference

    work page

Show all 32 references
  1. [9]

    wav2vec 2.0: A Framework for Self- Supervised Learning of Speech Representations [Internet]

    Baevski A, Zhou H, Mohamed A, Auli M. wav2vec 2.0: A Framework for Self- Supervised Learning of Speech Representations [Internet]. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. Vancouver, BC, Canada: Curr an Associates Inc.; 202...

    2020

  2. [10]

    Robust Speech Recognition via Large -Scale Weak Supervision [Internet]

    Radford A, Kim JW, Xu T, Brockman G, Mcleavey C, Sutskever I. Robust Speech Recognition via Large -Scale Weak Supervision [Internet]. In: ICML’23: Proceedings of the 40th International Conference on Machine Learning. Honolulu, Hawaii, USA: Journal of M achine Learning Research

  3. [11]

    Low resource Twi -English parallel corpus for machine translation in multiple domains (Twi-2-ENG)

    Agyei E, Zhang X, Bannerman S, Quaye AB, Yussi SB, Agbesi VK. Low resource Twi -English parallel corpus for machine translation in multiple domains (Twi-2-ENG). Discover Computing [Internet] 2024;27:17. Available from: https://link.springer.com/10.1007/s10791-024-09451-8

    2024 doi

  4. [12]

    XLS -R: Self- supervised Cross -lingual Speech Representation Learning at Scale

    Babu A, Wang C, Tjandra A, Lakhotia K, Xu Q, Goyal N, et al. XLS -R: Self- supervised Cross -lingual Speech Representation Learning at Scale. In: Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH. Int ernational Speech Comm...

    2022

  5. [13]

    Developing an open-source corpus of yoruba speech

    Gutkin A, Demirsahin I, Kjartansson O, Rivera C, Túbòsún K. Developing an open-source corpus of yoruba speech. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH 2020;2020-Octob:404–8. 16 Mensah, et al

    2020

  6. [14]

    S, Sampah D, Adjepon-Yamoah D, et al

    Asamoah Owusu D, Korsah A, Quartey B, Nwolley Jnr. S, Sampah D, Adjepon-Yamoah D, et al. GitHub - Ashesi-Org/Financial-Inclusion-Speech- Dataset: A speech dataset to support financial inclusion [Internet]. GitHub - Ashesi-Org2022;Available from: https: //github.com/Ashesi-Org/...

  7. [15]

    The Bible as a Parallel Corpus: Annotating the ‘Book of 2000 Tongues.’ Comput Hum [Internet] 1999;33:129 –53

    Resnik P, Olsen MB, Diab M. The Bible as a Parallel Corpus: Annotating the ‘Book of 2000 Tongues.’ Comput Hum [Internet] 1999;33:129 –53. Available from: https://doi.org/10.1023/A:1001798929185

    2000 doi

  8. [16]

    Common Voice 17.0: Twi [Internet]

    Mozilla Foundation. Common Voice 17.0: Twi [Internet]. 2023;Available from: https://huggingface.co/datasets/mozilla - foundation/common_voice_12_0/viewer/tw/train

    2023

  9. [17]

    Language of Instruction Country Profile: Ghana

    USAID. Language of Instruction Country Profile: Ghana. 2021;1 –13. Available from: https://pdf.usaid.gov/pdf_docs/PA00XH27.pdf

    2021

  10. [18]

    Globalization , Colonization , and Linguicide : How Ghana is Losing its Local Languages through Radio and Television Broadcast

    Antwi-Boasiako KB, Agyekum K. Globalization , Colonization , and Linguicide : How Ghana is Losing its Local Languages through Radio and Television Broadcast. Int J Humanit Soc Sci 2022;12:142–51

    2022

  11. [19]

    Language shift: A case study of Ghana

    Agyekum K. Language shift: A case study of Ghana. Sociolinguistic Studies 2009;3:381–403

    2009

  12. [20]

    Research of Automatic Speech Recognition of Asante -Twi Dialect for Translation

    Boakye-Yiadom AA, Qin M, Jing R. Research of Automatic Speech Recognition of Asante -Twi Dialect for Translation. In: ACM International Conference Proceeding Series. Association for Computing Machinery; 2021. page 1086–94

    2021

  13. [21]

    The Applicability of Wav2Vec2 and Whisper for Low -Resource Maltese ASR [Internet]

    Williams A, Demarco A, Borg C. The Applicability of Wav2Vec2 and Whisper for Low -Resource Maltese ASR [Internet]. In: 2nd Annual Meeting of the ELRA/ISCA SIG on Under-resourced Languages (SIGUL 2023). ISCA: ISCA

    2023

  14. [22]

    AraConv: Developing an Arabic Task -Oriented Dialogue System Using Multi -Lingual Transformer Model mT5

    Fuad A, Al -Yahya M. AraConv: Developing an Arabic Task -Oriented Dialogue System Using Multi -Lingual Transformer Model mT5. Applied Sciences (Switzerland) 2022;12

    2022

  15. [23]

    page 39 –43.Available from: https ://www.isca- archive.org/sigul_2023/williams23_sigul.html

  16. [24]

    CDEvalSumm: An Empirical Study of Cross -Dataset Evaluation for Neural Summarization Systems

    Chen Y, Liu P, Zhong M, Dou ZY, Wang D, Qiu X, et al. CDEvalSumm: An Empirical Study of Cross -Dataset Evaluation for Neural Summarization Systems. Findings of the Association for Computational Linguistics Findings of ACL: EMNLP 2020 [Internet] 2020 [cited 2025 Jun 22];3679–91...

    2020

  17. [25]

    End -To-End Multilingual Automatic Speech Recognition for Less -Resourced Languages: The Case of Four Ethiopian Languages

    Abate ST, Tachbelie MY, Schultz T. End -To-End Multilingual Automatic Speech Recognition for Less -Resourced Languages: The Case of Four Ethiopian Languages. In: ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings. Institute of Electr...

    2021

  18. [26]

    Cross -Domain Generalization of Neural Constituency Parsers

    Fried D, Kitaev N, Klein D. Cross -Domain Generalization of Neural Constituency Parsers. ACL 2019 - 57th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference [Internet] 2019 Benchmarking Akan ASR Models 17 [cited 2025 Jun 22];323 –30. ...

    2019

  19. [27]

    Cross -Dataset Evaluation of Multimodal Neural Networks for Glaucoma Diagnosis

    Das S, Jain A, Durai A, Gabbita S, Vasantharao A, Kotha V. Cross -Dataset Evaluation of Multimodal Neural Networks for Glaucoma Diagnosis. Proceedings - 2022 IEEE 9th International Conference on Data Science and Advanced Analytics, DSAA 2022 2022

    2022

  20. [28]

    Recent advancements in automatic disordered speech recognition: A survey paper

    Gohider N, Basir OA. Recent advancements in automatic disordered speech recognition: A survey paper. Natural Language Processing Journal [Internet] 2024 [cited 2025 Jun 23];9:100110. Available from: https://linkinghub.elsevier.com/retrieve/pii/S294971912400058X

    2024

  21. [29]

    An overview of high -resource automatic speech recognition methods and their empirical evaluation in low - resource environments

    Fatehi K, Torres Torres M, Kucukyilmaz A. An overview of high -resource automatic speech recognition methods and their empirical evaluation in low - resource environments. Speech Commun [Internet] 2025 [cited 2025 Apr 27];167:103151. Available from: https://www.sciencedirect.c...

    2025

  22. [30]

    Unveiling Biases while Embracing Sustainability: Assessing the Dual Challenges of Automatic Speech Recognition Systems

    Kulkarni A, Kulkarni A, Couceiro M, Trancoso I. Unveiling Biases while Embracing Sustainability: Assessing the Dual Challenges of Automatic Speech Recognition Systems. Proceedings of the Annual Conference of the International Speech Communication Asso ciation, INTERSPEECH 2024;4628–32

    2024

  23. [31]

    BibleTTS: a large, high -fidelity, multilingual, and uniquely African speech corpus

    Meyer J, Adelani DI, Casanova E, Öktem A, Weber DWJ, Kabongo S, et al. BibleTTS: a large, high -fidelity, multilingual, and uniquely African speech corpus. arXiv preprint [Internet] 2022;Available from: http://arxiv.org/abs/2207.03546

    2022 arXiv

  24. [2023]

    page 28492–518.Available from: https://github.com/openai/

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.