Pith. sign in

REVIEW 4 major objections 6 minor 35 references

PSRB: A Comprehensive Benchmark for Evaluating Persian ASR Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A new benchmark for Persian speech recognition, built from 3,372 real-world clips, ranks ten ASR systems and shows they all degrade on informal speech, regional accents, and children's voices; the paper also introduces a…

desk verdict A genuinely useful Persian ASR benchmark with a modest new metric, but the unvalidated reference transcriptions put the bias findings on shaky ground. read the letter →

arxiv 2505.21230 v1 pith:CI4UTQ5Q submitted 2025-05-27 eess.AS cs.AIcs.CLcs.SD

Nima Sedghiyeh , Sara Sadeghi , Reza Khodadadi , Farzin Kashani , Omid Aghdaei , Somayeh Rahimi , Mohammad Sadegh Safari This is my paper
classification eess.AScs.AIcs.CLcs.SD
keywords AutomaticSpeechRecognitionPersianASRbenchmarkdemographicbiaserroranalysisSW-WERlow-resourcelanguagesregionalaccents
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

Persian is spoken by over 100 million people, but until now there has been no widely accepted benchmark that tests automatic speech recognition (ASR) systems on the messy conditions of real Persian speech. This paper introduces PSRB, a benchmark of 3,372 audio clips spanning formal and informal speech, twelve regional accents, multiple age groups, and clean, noisy, and reverberant settings. The authors evaluate ten commercial and open-source ASR systems on the same clips and report a clear performance hierarchy plus systematic biases: systems do well on standard formal Persian but struggle with regional accents, children's speech, informal phrasing, and multi-speaker audio. They also propose a new evaluation metric, Substitution Weighted WER (SW-WER), which weights substitution errors by how far the recognized word actually is from the reference, arguing that it gives a fairer score for Persian than raw word or character error rates. If the benchmark and metric hold up, they give researchers a reusable instrument for improving Persian ASR and a template for other low-resource languages.

What carries the argument

The machinery has two parts. The first is the PSRB corpus: 10.4 hours of audio split into 3,372 clips, 78,526 words, and 756 speakers, designed to cover eight categories (data source, speech style, semantic content, linguistic diversity, speaker diversity, accent, number of speakers, and acoustic environment) with at least 50 utterances per category. The second is the metric SW-WER, defined as $$\mathrm{SW\text{-}WER} = \frac{S + I + D}{N_{\mathrm{sub}} + C + D}$$, where $I$ and $D$ count insertions and deletions, $C$ counts correct words, and each substitution contributes $s_i = n_i \times \mathrm{CER}(\mathrm{ref}_i, \mathrm{hyp}_i)$, so a substitution that differs by one character counts much less than one that changes the whole word. The corpus is what makes the bias claims measurable; the metric is what makes errors comparable across models whose mistakes differ in severity. The paper also uses a decoder-type ablation on one model to show that the benchmark can expose hidden failure modes that aggregate error rates do not show.

What would settle it

Take a random sample of PSRB clips, say 300 spanning accents, ages, and formality levels, have two independent teams transcribe them, and compare their transcripts to the released ground truth; if agreement falls below a conventional threshold, or if disagreement concentrates in the same categories the paper flags as failure modes (informal speech, children, Mashhadi and Yazdi accents), then the reported CER, WER, and SW-WER gaps and model rankings are not yet established.

Watch

Extended reading notes

Core claim

The paper's central claim is that PSRB is the kind of benchmark Persian ASR has been missing: a balanced, multi-dimensional test set whose measurements are diagnostic rather than merely comparative. On it, the best result is posted by a commercial Persian system (Avanegar), with two other commercial services close behind; open-source and multilingual models trail, with word error rates roughly double. Across all ten systems, error rates rise sharply for informal speech, children's voices, regional accents such as Mashhadi and Yazdi, noisy conditions, and multi-speaker clips, while adult standard Persian remains comparatively easy. A secondary discovery is that the transducer decoder of one FastConformer-based Persian model hallucinates far more than its CTC decoder, with character error rate jumping from 19.42% to 40.73% on the same audio. The proposed SW-WER metric separates meaningful word-level mistakes from cosmetic character-level differences, making model comparisons less distorted by Persian's flexible word boundaries.

Load-bearing premise

The entire benchmark sits on the assumption that the human transcriptions used as ground truth are correct, consistent, and unbiased; the paper reports selecting and reviewing annotators but gives no inter-annotator agreement or independent audit, so if transcription quality varies with accent, formality, or child speech, every error rate, ranking, and bias conclusion would shift.

Editorial extensions

If this is right

  • Persian ASR comparisons should shift from clean read-speech test sets to PSRB-style conditions, because the same model can appear competitive on standard speech while falling behind on informal and accented speech.
  • The reported rankings imply that fine-tuning on informal, child, and accented Persian data is a direct path to improving real-world performance, since these are the categories where every system loses accuracy.
  • If adopted, SW-WER would change how benchmarks report Persian errors, reducing the influence of minor spelling or ZWNJ issues while still penalizing whole-word substitutions.
  • The decoder ablation shows that benchmark choice matters even for the same model: the transducer variant's 40.73% CER would be the headline result if the CTC variant had not been selected.
  • PSRB can serve as a template for other low-resource languages that need benchmarks spanning demographics, accents, and spontaneity rather than read speech alone.

Reading between the lines

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

  • A natural next step the paper does not take is to test whether LLM-based post-editing of ASR output narrows the SW-WER gaps between commercial and open-source models; PSRB's public subset would allow that comparison directly.
  • The benchmark's value for fairness audits depends on full release of the dataset and annotation guidelines; if only a subset is public, the rankings cannot be independently reproduced by other groups.
  • The SW-WER idea may transfer to other morphologically rich or boundary-ambiguous languages, since its weighting mechanism is not Persian-specific, but the paper leaves that generalization untested.
  • Because inter-annotator agreement is not reported, a prudent reader should treat the absolute error rates as provisional until transcription reliability is demonstrated, even if the relative pattern across models is real.
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 / 6 minor

Summary. The paper introduces PSRB, a Persian ASR benchmark consisting of 3,372 clips (about 10.4 hours) spanning multiple accents, age groups, formality levels, data sources, acoustic conditions, and multi-speaker scenarios. Ten ASR systems, including commercial and open-source models, are evaluated with CER, WER, and a proposed SW-WER metric that weights word substitutions by character-level distance. The reported results show commercial models leading the ranking, substantial degradation on children's speech and regional accents, and an ablation of the NVIDIA FastConformer decoder types. A subset of the dataset is released on Hugging Face.

Significance. If the reference transcriptions are valid and the comparisons are statistically supported, PSRB would be a valuable reusable resource for Persian ASR evaluation, with unusually broad demographic and acoustic coverage and an interpretable, parameter-free metric. Strengths of the manuscript include the detailed multidimensional design, the large number of speakers and unique words, the inclusion of both commercial and open-source systems, and the fact that SW-WER is a fixed deterministic combination of WER and CER rather than a fitted metric. The main limitations are the lack of validation of the reference transcriptions, the absence of confidence intervals and significance tests for the comparative claims, and the release of only an unspecified subset of the benchmark.

major comments (4)
  1. [Section 3.1.1] The manuscript states that four top-performing annotators were selected and a two-tier review was applied, but it reports no inter-annotator agreement measure, no gold-standard audit, no breakdown of review disagreements, and does not release the annotation guidelines. Because every CER, WER, and SW-WER in Tables 4-8 is computed against these transcriptions, systematic annotator conventions—for example, normalizing informal morphology, resolving ZWNJ/word-boundary choices, or transcribing accented realizations as standard orthography—would propagate directly into the model rankings and into the age, accent, and formality bias comparisons. Section 5.1 itself identifies ZWNJ, formality, and He-Kasreh as ambiguities, which are exactly the places where annotator policy can masquerade as ASR error. Please provide inter-annotator agreement metrics on a held-out subset, quantify disagreement resolution, and release the annotation guidelines.
  2. [Tables 4, 6, 7, 8; Figures 4-7] All benchmark results are presented as point estimates without confidence intervals or significance tests. The claims that Avanegar is best (CER 8.75 vs Chirp 9.05 vs Aipa 10.43), that Chirp is best in informal noisy conditions, and that Faster-Whisper surpasses Azure in informal noisy conditions all rest on small differences that could easily be within utterance-level variability. Please report per-utterance variances, paired bootstrap confidence intervals, or paired significance tests for all model comparisons and subgroup comparisons, and restrict ranking statements to differences supported by those intervals.
  3. [Section 3.2, Eqs. (1)-(5) and Table 3] The definition of SW-WER is not internally consistent as written, and the illustrative numbers do not reproduce the formula. In the first row of Table 3, WER=100.00% and CER=12.90%, but SW-WER=33.33%; under Eq. (1) a single substituted word with n=1 would give SW-WER=CER, while a multi-substitution case depends on the unstated segmentation. Moreover, the notion of a 'reference substitution segment' with n_i words is not defined for standard word-level Levenshtein alignment, where each substitution replaces one word. Please specify the alignment and segmentation algorithm, clarify how CER is computed for segments of unequal length, and provide a step-by-step worked example that reproduces the values in Table 3.
  4. [Abstract and Section 1 (data availability)] The paper states that only 'a subset' of PSRB is publicly available, without specifying which subset, how it was selected, or whether it preserves the category balance of the full benchmark. A benchmark whose full reference corpus is private cannot be independently audited or reused, which weakens the central claim that PSRB is a reusable evaluation instrument. Please release the full dataset, or a large clearly documented sample with stratified category counts, along with the annotation guidelines and metadata schema.
minor comments (6)
  1. [Section 2] The sentence 'Szymański et al. [6] Kuhn et al. [11] note...' is missing a conjunction and appears to cite the wrong reference; the second citation should be to Ardila et al. (Common Voice), not 'Kuhn et al.'.
  2. [Abstract and Table captions] There are spacing errors such as 'Benchmark(PSRB)' in the abstract and 'T able' in the captions of Tables 1 and 5; please correct these typographical issues.
  3. [Table 2 and Section 3.1.2] Table 2 reports a maximum PSRB duration of 383.7 seconds, which contradicts the stated clip constraint of 1-100 seconds in Section 3.1.2 and the x-axis of Figure 2; please reconcile these numbers.
  4. [Section 6.1 and Table 7] The text says 'the SLPL model... performs identically to the whisper model' in single-speaker mode, but Table 7 shows materially different values (WER 43.4 vs 38.4, SW-WER 29.3 vs 28.0); please correct the claim.
  5. [Section 6.3] The statement that 'outlier data points were excluded from the analysis' is made without defining the outlier criterion; please specify the rule or remove the exclusion.
  6. [Figures 6 and 7] The labels CM1 and CM2 in the figure captions are not defined; please name the models explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SW-WER is a fixed, parameter-free formula over standard ASR alignment quantities, and the benchmark results are measurements against external reference transcriptions rather than outputs of the metric.

full rationale

The paper's central derivation chain is the definition of SW-WER in Equations (1)-(5). This metric is a closed-form function of standard Levenshtein-alignment quantities: substitution errors S, insertions I, deletions D, and correct words C, with each substitution segment weighted by the CER of that segment normalized by the substituted word count. No parameter in this formula is fitted to the benchmark data, and the metric is not used to generate or normalize the reference transcriptions. The claim that SW-WER correlates with WER and CER is presented as an empirical observation in Figure 3, not as a derivation of the benchmark results. The model rankings and bias findings in Tables 4-8 are computed by applying standard metrics to external ASR system outputs against reference transcriptions; the metric and the benchmark are independent inputs to those measurements. The principal validity concern is that the reference transcriptions lack reported inter-annotator agreement or a gold-standard audit, as noted in Section 3.1.1, but this is a data-quality and external-validity limitation, not a circularity: the transcriptions are not defined in terms of the ASR outputs, and no prediction is constructed from the quantity it claims to measure. The only self-referential element is that the new metric is demonstrated on the authors' own benchmark, which affects generalization but does not make the derivation equivalent to its inputs. No load-bearing self-citation, imported uniqueness theorem, or ansatz hidden in a citation was found. Therefore the derivation is self-contained with respect to circularity.

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

No numerical free parameters are fitted and no new physical or conceptual entities are introduced; SW-WER is a deliberately parameter-free weighting of standard WER. The real burden is carried by assumptions about annotation quality, metadata correctness, and sample representativeness.

assumptions (3)
  • domain assumption Manual transcriptions in PSRB are accurate and consistent enough to serve as ground truth.
    Section 3.1.1 describes two-tier review and top-annotator selection, but no inter-annotator agreement or external validation is reported; every model score is relative to these transcriptions.
  • domain assumption Category labels for accent, age, gender, formality, spontaneity, and acoustic environment are correctly and consistently assigned.
    Section 3.1.1 says clips were labeled and Table 1 shows examples, but no metadata labeling protocol or quality metric is given. Section 6's bias and robustness analyses rely entirely on these labels.
  • domain assumption A sample with at least 50 utterances per category is sufficient to support the reported model comparisons and bias conclusions.
    Section 3.1.2 asserts this balance, but no power analysis or per-cell counts are given, and 3,372 utterances cannot densely cover all combinations of 12 accents, 6 age groups, 2 genders, and 2 spontaneity levels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PSRB: A Comprehensive Benchmark for Evaluating Persian ASR Systems." pith.science (2026). https://pith.science/paper/CI4UTQ5Q

@misc{pith2026250521230,
  author       = {Pith},
  title        = {Pith review of: PSRB: A Comprehensive Benchmark for Evaluating Persian ASR Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CI4UTQ5Q}},
  note         = {Machine review of arXiv:2505.21230}
}
read the original abstract

Although Automatic Speech Recognition (ASR) systems have become an integral part of modern technology, their evaluation remains challenging, particularly for low-resource languages such as Persian. This paper introduces Persian Speech Recognition Benchmark(PSRB), a comprehensive benchmark designed to address this gap by incorporating diverse linguistic and acoustic conditions. We evaluate ten ASR systems, including state-of-the-art commercial and open-source models, to examine performance variations and inherent biases. Additionally, we conduct an in-depth analysis of Persian ASR transcriptions, identifying key error types and proposing a novel metric that weights substitution errors. This metric enhances evaluation robustness by reducing the impact of minor and partial errors, thereby improving the precision of performance assessment. Our findings indicate that while ASR models generally perform well on standard Persian, they struggle with regional accents, children's speech, and specific linguistic challenges. These results highlight the necessity of fine-tuning and incorporating diverse, representative training datasets to mitigate biases and enhance overall ASR performance. PSRB provides a valuable resource for advancing ASR research in Persian and serves as a framework for developing benchmarks in other low-resource languages. A subset of the PSRB dataset is publicly available at https://huggingface.co/datasets/PartAI/PSRB.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    In: 2023 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU), pp

    Rekesh, D., Koluguri, N.R., Kriman, S., Majumdar, S., Noroozi, V., Huang, H., Hrinchuk, O., Puvvada, K., Kumar, A., Balam, J., et al.: Fast conformer with linearly scalable attention for efficient speech recognition. In: 2023 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU), pp. 1–8 (2023). 22 IEEE

    work page 2023

  2. [2]

    In: 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp

    Chung, Y.-A., Zhang, Y., Han, W., Chiu, C.-C., Qin, J., Pang, R., Wu, Y.: W2v- bert: Combining contrastive learning and masked language modeling for self- supervised speech pre-training. In: 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 244–250 (2021). IEEE

    work page 2021

  3. [3]

    In: International Conference on Machine Learning, pp

    Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning, pp. 28492–28518 (2023). PMLR

    work page 2023

  4. [4]

    In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp

    Panayotov, V., Chen, G., Povey, D., Khudanpur, S.: Librispeech: an asr corpus based on public domain audio books. In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5206–5210 (2015). IEEE

    work page 2015

  5. [5]

    22–34 (2021)

    Aks¨ enova, A., Esch, D., Flynn, J., Golik, P.: How might we create better benchmarks for speech recognition? In: Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pp. 22–34 (2021)

    work page 2021

  6. [6]

    WER we are and WER we think we are

    Szyma´ nski, P., ˙Zelasko, P., Morzy, M., Szymczak, A., ˙Zy la-Hoppe, M., Banaszczak, J., Augustyniak, L., Mizgajski, J., Carmiel, Y.: Wer we are and wer we think we are. arXiv preprint arXiv:2010.03432 (2020)

    work page Pith review arXiv 2020

  7. [7]

    arXiv preprint arXiv:2104.11348 (2021)

    Del Rio, M., Delworth, N., Westerman, R., Huang, M., Bhandari, N., Palakapilly, J., McNamara, Q., Dong, J., Zelasko, P., Jett´ e, M.: Earnings-21: A practical benchmark for asr in the wild. arXiv preprint arXiv:2104.11348 (2021)

    arXiv 2021

  8. [8]

    In: Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization, pp

    Cao, J., Ganesh, A., Cai, J., Southwell, R., Perkoff, E.M., Regan, M., Kann, K., Martin, J.H., Palmer, M., D’Mello, S.: A comparative analysis of automatic speech recognition errors in small group classroom discourse. In: Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization, pp. 250–262 (2023)

    work page 2023

Show all 35 references
  1. [9]

    arXiv preprint arXiv:2210.13352 (2022)

    Gandhi, S., Von Platen, P., Rush, A.M.: Esb: A benchmark for multi-domain end-to-end speech recognition. arXiv preprint arXiv:2210.13352 (2022)

    2022 arXiv

  2. [10]

    In: 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp

    Zeinali, H., Burget, L., ˇCernock` y, J.H.: A multi purpose and large scale speech corpus in persian and english for speaker and speech recognition: the deep- mine database. In: 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 397–402 (2019). IEEE

    2019

  3. [11]

    arXiv preprint arXiv:1912.06670 (2019)

    Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., Morais, R., Saunders, L., Tyers, F.M., Weber, G.: Common voice: A massively- multilingual speech corpus. arXiv preprint arXiv:1912.06670 (2019)

    2019 arXiv

  4. [12]

    In: 2022 IEEE Spoken Language Technology Workshop (SLT), pp

    Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., 23 Rivera, C., Bapna, A.: Fleurs: Few-shot learning evaluation of universal represen- tations of speech. In: 2022 IEEE Spoken Language Technology Workshop (SLT), pp. 798–805 (2023). IEEE

    2023

  5. [13]

    In: 2023 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), pp

    Fuckner, M., Horsman, S., Wiggers, P., Janssen, I.: Uncovering bias in asr sys- tems: Evaluating wav2vec2 and whisper for dutch speakers. In: 2023 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), pp. 146–151 (2023). IEEE

    2023

  6. [14]

    Computer Speech & Language 84, 101567 (2024)

    Feng, S., Halpern, B.M., Kudina, O., Scharenborg, O.: Towards inclusive auto- matic speech recognition. Computer Speech & Language 84, 101567 (2024)

    2024

  7. [15]

    Interspeech 2024, 3045– 3049 (2024)

    Schubert, M., Duran, D., Siegert, I.: Challenges of german speech recognition: A study on multi-ethnolectal speech among adolescents. Interspeech 2024, 3045– 3049 (2024)

    2024

  8. [16]

    arXiv preprint arXiv:2402.07513 (2024)

    Kulkarni, A., Tokareva, A., Qureshi, R., Couceiro, M.: The balancing act: Unmasking and alleviating asr biases in portuguese. arXiv preprint arXiv:2402.07513 (2024)

    2024 arXiv

  9. [17]

    arXiv preprint arXiv:2204.05617 (2022)

    Wirth, J., Peinl, R.: Asr in german: A detailed error analysis. arXiv preprint arXiv:2204.05617 (2022)

    2022 arXiv

  10. [18]

    arXiv preprint arXiv:2409.09554 (2024)

    Ma, R., Qian, M., Gales, M., Knill, K.: Asr error correction using large language models. arXiv preprint arXiv:2409.09554 (2024)

    2024 arXiv

  11. [19]

    arXiv preprint arXiv:2111.09296 (2021)

    Babu, A., Wang, C., Tjandra, A., Lakhotia, K., Xu, Q., Goyal, N., Singh, K., Von Platen, P., Saraf, Y., Pino, J., et al.: Xls-r: Self-supervised cross-lingual speech representation learning at scale. arXiv preprint arXiv:2111.09296 (2021)

    2021 arXiv

  12. [20]

    arXiv preprint arXiv:2312.05187 (2023)

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

    2023 arXiv

  13. [21]

    Interspeech 2020 (2020)

    Gulati, A., Qin, J., Chiu, C.-C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., et al.: Conformer: Convolution-augmented transformer for speech recognition. Interspeech 2020 (2020)

    2020

  14. [22]

    In: Proceedings of the 23rd International Conference on Machine Learning, pp

    Graves, A., Fern´ andez, S., Gomez, F., Schmidhuber, J.: Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural net- works. In: Proceedings of the 23rd International Conference on Machine Learning, pp. 369–376 (2006)

    2006

  15. [23]

    arXiv preprint arXiv:1211.3711 (2012) 24

    Graves, A.: Sequence transduction with recurrent neural networks. arXiv preprint arXiv:1211.3711 (2012) 24

    2012 arXiv

  16. [24]

    Cephei, A.: Vosk Speech Recognition Toolkit. Online. https://alphacephei.com/ vosk/

  17. [25]

    Azure, M.: Azure Speech to Text. Online. https://azure.microsoft.com/en-us/ services/cognitive-services/speech-to-text/#features

  18. [26]

    Speech recognition model

    Google: Chirp 2. Speech recognition model. https://cloud.google.com/ speech-to-text

  19. [27]

    In: 2009 International Conference on Asian Language Processing, pp

    Ghayoomi, M., Momtazi, S.: Challenges in developing persian corpora from online resources. In: 2009 International Conference on Asian Language Processing, pp. 108–113 (2009). IEEE

    2009

  20. [28]

    Language resources and evaluation 45, 143–164 (2011)

    Bijankhan, M., Sheykhzadegan, J., Bahrani, M., Ghayoomi, M.: Lessons from building a persian written corpus: Peykare. Language resources and evaluation 45, 143–164 (2011)

    2011

  21. [29]

    arXiv preprint arXiv:2401.01572 (2024)

    Frieske, R., Shi, B.E.: Hallucinations in neural automatic speech recognition: Identifying errors and hallucinatory models. arXiv preprint arXiv:2401.01572 (2024)

    2024 arXiv

  22. [30]

    arXiv preprint arXiv:2502.12414 (2025)

    Atwany, H., Waheed, A., Singh, R., Choudhury, M., Raj, B.: Lost in transcrip- tion, found in distribution shift: Demystifying hallucination in speech foundation models. arXiv preprint arXiv:2502.12414 (2025)

    2025 arXiv

  23. [31]

    arXiv preprint arXiv:2501.11378 (2025)

    Bara´ nski, M., Jasi´ nski, J., Bartolewska, J., Kacprzak, S., Witkowski, M., Kowal- czyk, K.: Investigation of whisper asr hallucinations induced by non-speech audio. arXiv preprint arXiv:2501.11378 (2025)

    2025 arXiv

  24. [32]

    In: ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp

    Huang, Z., Raj, D., Garc ´ ıa, P., Khudanpur, S.: Adapting self-supervised models to multi-talker speech recognition using speaker embeddings. In: ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5 (2023). IEEE

    2023

  25. [33]

    arXiv preprint arXiv:2407.09817 (2024)

    Meng, L., Kang, J., Wang, Y., Jin, Z., Wu, X., Liu, X., Meng, H.: Empowering whisper as a joint multi-talker and target-talker speech recognition system. arXiv preprint arXiv:2407.09817 (2024)

    2024 arXiv

  26. [34]

    arXiv preprint arXiv:2305.18747 (2023)

    Li, C., Qian, Y., Chen, Z., Kanda, N., Wang, D., Yoshioka, T., Qian, Y., Zeng, M.: Adapting multi-lingual asr models for handling multiple talkers. arXiv preprint arXiv:2305.18747 (2023)

    2023 arXiv

  27. [35]

    Applied Sciences 12(9), 4419 (2022) 25

    Bhardwaj, V., Ben Othman, M.T., Kukreja, V., Belkhier, Y., Bajaj, M., Goud, B.S., Rehman, A.U., Shafiq, M., Hamam, H.: Automatic speech recognition (asr) systems for children: A systematic literature review. Applied Sciences 12(9), 4419 (2022) 25

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.