Pith. sign in

REVIEW 4 major objections 4 minor 50 references

High-precision medical speech recognition through synthetic data and semantic correction: UNITED-MEDASR

T0 review · 4 major / 4 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 Synthetic medical audio cuts ASR word error below 1 percent

desk verdict The dataset is a real artifact, but the headline WERs are unreproducible and internally contradictory: the abstract and body report different numbers for the same benchmarks. read the letter →

arxiv 2412.00055 v1 pith:BMH7HSP5 submitted 2024-11-24 eess.AS cs.CLcs.SD

Sourav Banerjee , Ayushi Agarwal , Promila Ghosh This is my paper · ORCID
classification eess.AScs.CLcs.SD
keywords medicalspeechrecognitionsyntheticdataWhisperfine-tuningsemanticenhancementBARTdomainadaptationworderrorrateStyleTTS2
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

United-MedASR is a claim about how to build high-accuracy, domain-specific speech recognition without large amounts of real, privacy-protected clinical audio. The paper argues that scraping authoritative medical text sources, turning them into sentences with a language model, synthesizing speech with StyleTTS2, fine-tuning Whisper on that synthetic audio, and then passing transcripts through a BART-based semantic corrector yields word error rates below 1% on standard English benchmarks. If true, clinical transcription could bypass the usual data-collection bottleneck while still handling rare drug names and procedures. The architecture is presented as transferable to other specialized vocabularies.

What carries the argument

The load-bearing mechanism is the synthetic data pipeline feeding a two-stage model chain. Text scraped from medical references is expanded by a language model into contextually realistic sentences, each tagged with a unique ID; StyleTTS2 converts those sentences to clean 30-second audio; the Whisper medium model is fine-tuned on the resulting labeled corpus; and a BART-base sequence-to-sequence model, fine-tuned on clinical dictation-style text, rewrites the ASR hypothesis into the correct medical phrasing. The semantic enhancer is what the paper credits with absorbing errors on complex terms that acoustic fine-tuning alone would miss, and the WER is computed on the BART-corrected output.

What would settle it

Run the published pipeline with the released synthetic dataset and fine-tuning hyperparameters, evaluate on the official LibriSpeech test-clean with the standard benchmark normalization, and search the 395,000 generated sentences for verbatim or near-verbatim fragments of the four benchmark references; if the WER rises above 1% or any overlap appears, the sub-1% claims collapse.

Watch

Extended reading notes

Core claim

The central claim is that a fully synthetic training corpus, built from ICD-10, MIMS, and FDA text, is sufficient to adapt a general ASR model to medical language at a level that beats both general-purpose and earlier domain-specific systems. Specifically, United-MedASR generates 395,000 unique medical sentences, synthesizes about 790,000 thirty-second audio clips (roughly 5,486 hours) with StyleTTS2 voices, fine-tunes Whisper medium with Faster Whisper and voice-activity detection, and then applies a separately fine-tuned BART-base model to semantically correct the raw transcript. The paper reports 0.985% WER on LibriSpeech test-clean, 0.26% on Europarl-ASR EN Guest-test, 0.29% on TED-LIUM, and 0.336% on FLEURS (English). The intended conclusion is that the pipeline's layered design—synthetic acoustic exposure plus semantic post-editing—is what makes specialized vocabulary recognition possible at near-human accuracy.

Load-bearing premise

The results rest on the assumption that the reported WERs were computed on the full official test sets under standard normalization and that none of the benchmark reference transcripts leaked into the synthetic corpus, the Whisper fine-tuning, or the BART semantic enhancer.

Editorial extensions

If this is right

  • Medical ASR can be brought to production accuracy without collecting real patient voice data, reducing privacy and annotation costs.
  • The same scrape-synthesize-fine-tune-correct recipe should transfer to other terminology-heavy domains such as law, pharmacology, or radiology.
  • The sub-1% results imply that synthetic audio, at scale, can substitute for a large fraction of human speech data in a constrained vocabulary.
  • Faster Whisper conversion and VAD filtering make the accuracy gains available in near-real-time transcription, not just offline scoring.
  • The released synthetic dataset gives other teams a concrete starting point for reproducing or extending the pipeline.

Reading between the lines

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

  • A controlled ablation that removes the BART post-corrector would reveal how much of the gain is semantic cleanup versus acoustic adaptation; the paper does not isolate these contributions.
  • If the method is as domain-transferable as claimed, the same architecture should produce comparable sub-1% WER on a non-medical benchmark after swapping the source vocabulary, a test the paper does not run.
  • The dependence on StyleTTS2 means that future improvements in neural text-to-speech could directly lower the ASR error floor for this class of pipelines.
  • Because the synthetic audio is generated from text, the pipeline can in principle be extended to low-resource languages simply by changing the text source and TTS voice.
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 / 4 minor

Summary. The paper proposes United-MedASR, a medical ASR pipeline that combines synthetic text generation (GPT), synthetic speech generation (StyleTTS2), fine-tuning of Whisper/Faster-Whisper, and a BART-based semantic correction model. The central claim is that this pipeline achieves sub-1% Word Error Rates on four standard benchmarks: LibriSpeech test-clean (0.985%), Europarl-ASR EN Guest-test (0.26%), TED-LIUM (0.29%), and FLEURS English (0.336%). The evaluation is described in Section 4.3 and Algorithm 7.

Significance. If the reported WERs were correct and reproducible, this would be an important result for domain-specific ASR, demonstrating that synthetic data plus a semantic post-processor can surpass general-purpose models on standard English benchmarks by a large margin. The paper also releases the United-Syn-Med dataset on HuggingFace, which is a useful community resource. However, the central claim is currently unsupported because the paper's own tables and text report conflicting numbers for the same benchmarks, and the evaluation protocol is underspecified. The claimed mean WER is never actually reported.

major comments (4)
  1. [§4.3.1, Table 3, Abstract] Section 4.3.1 states that WER is computed per sample and 'then averaging them out,' yet Table 3 reports only minimum, maximum, and median WER, not the mean. The abstract's Europarl-ASR value (0.26%) matches the Table 3 median (0.26), and the LibriSpeech value (0.985%) is close to the median (0.98), but the body text (§4.3.3, §4.3.4) reports Europarl-ASR as 0.412 and TED-LIUM as 0.514, while the abstract's TED-LIUM value (0.29%) matches no reported statistic. The claimed mean is therefore never reported, and the central sub-1% claim rests on inconsistent statistics.
  2. [Algorithm 7, §4.3.1] The benchmark evaluation protocol is underspecified. Algorithm 7 calls LOADBENCHMARK, APPLYNOISE REDUCTION, and CALCULATE WER without defining test-set sizes, subset selection, audio preprocessing, or text normalization beyond lowercasing and space removal. No statement is made about excluding benchmark transcripts from the synthetic corpus, GPT-generated sentences, or BART training data. Without these details and leakage controls, the reported WERs are not independently verifiable.
  3. [Table 2, §4.3.3, §4.3.4] The comparison baselines in Table 2 are not shown to be evaluated under the same protocol. For example, the Europarl-ASR prior best (7.0) is presumably the original corpus result, while United-MedASR is evaluated on 'Guest-test' after noise reduction; no evidence is given that the comparison uses the same test split, audio preprocessing, or normalization. An improvement of this magnitude (7.0 to 0.412) requires a controlled comparison to be credible.
  4. [§4.3.2, §4.3.5, Abstract] The units and exact values of the reported WERs are inconsistent. The abstract gives percentages (0.985%, 0.26%, 0.29%, 0.336%), while body sections use bare numbers (0.985, 0.412, 0.514) and one percentage (0.336%). LibriSpeech is 0.985 in §4.3.2 but 0.98 in Table 2, and 0.985% in the abstract; these are not trivially reconcilable. This ambiguity makes it impossible to compare against prior work or to reproduce the measurements.
minor comments (4)
  1. [§4.3.1] The reference to noise reduction '[37][38]' is not clearly tied to the algorithm; please specify the exact noisereduce configuration and version used.
  2. [Table 3] Table 3 reports only minimum, maximum, and median WER, but the abstract appears to use some of these medians as headline numbers; please clarify which statistic is reported in each location and provide the mean as stated in Section 4.3.1.
  3. [§3.5, Algorithm 4] Algorithm 4 lists batch_size 16, but the text says batch size 1 per device; please reconcile these values.
  4. [Throughout] The paper uses 'WER' and 'CER' both as fractions and percentages, with inconsistent use of the percent sign; please standardize on one convention and ensure all values are explicitly labeled.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claimed benchmark WERs are independent empirical evaluations on external test sets after model training, not quantities derived from or equivalent to the pipeline's own inputs.

full rationale

United-MedASR's derivation chain is an empirical engineering pipeline, not a formal derivation. Synthetic medical text is scraped from ICD-10/MIMS/FDA sources (Section 3.2), converted to speech with StyleTTS2 (Section 3.3), used to fine-tune Whisper (Section 3.5), and followed by a BART-based semantic enhancer (Section 3.7). The headline results are then computed by running the trained system on external benchmarks—LibriSpeech test-clean, Europarl-ASR, TED-LIUM, and FLEURS (Section 4.3, Algorithm 7)—and comparing transcripts against the benchmarks' own reference texts. There is no fitted parameter, self-definition, or imported uniqueness theorem that makes any reported WER equal to a training target by construction. The only self-referential element is reference [49], the authors' own publicly released synthetic dataset, but it is cited only as a training-data resource and is not load-bearing evidence for the benchmark claims. The internal inconsistencies among the abstract, Table 2, and Table 3 WER values (e.g., Europarl-ASR 0.26% vs 0.412%, TED-LIUM 0.29% vs 0.514%) and the unspecified normalization/leakage controls in Algorithm 7 are serious reproducibility and validity concerns, but they are not circularity: no equation in the paper reduces to its own inputs, and the benchmark numbers are not forced by the training procedure in the sense required for a circularity finding. The central claim therefore has independent empirical content, even though its reliability is compromised by the reporting inconsistencies.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

All training data is synthetic: scraped medical text, GPT-generated sentences, and StyleTTS2 audio. The WER claims depend on the transfer of that synthetic speech to real benchmark audio, on a clean evaluation protocol, and on the BART post-editor improving rather than distorting transcripts. No independent evidence is provided for any of these assumptions.

free parameters (6)
  • StyleTTS2 alpha = 0.3
    Blend weight for reference versus predicted style in synthetic audio; chosen by hand in Algorithm 2 and used to generate all training audio.
  • StyleTTS2 beta = 0.7
    Second style-vector blend weight in Algorithm 2; no sensitivity analysis is reported.
  • StyleTTS2 diffusion_steps = 6
    Quality and refinement trade-off for style embeddings; no sweep is reported.
  • Whisper fine-tuning hyperparameters = learning rate 1e-5, batch size 16 (Algorithm 4) or 1 per device (Section 3.5), max steps 5000, warmup 500
    These values determine the fine-tuned model; no search or sensitivity analysis is given.
  • BART fine-tuning hyperparameters = learning rate 4e-4, batch size 8
    Used for the semantic enhancer; no tuning details or training data sizes are given.
  • Synthetic audio duration = 30 seconds per clip
    Audio is standardized to Whisper's 30-second window; this choice affects utterance context and evaluation alignment.
assumptions (4)
  • domain assumption Synthetic speech from StyleTTS2, generated from GPT-paraphrased scraped medical text, is a faithful enough proxy for real clinical and benchmark speech.
    The entire Whisper fine-tuning set (Section 3.3) is synthetic; no human validation of the audio or transcripts is reported.
  • domain assumption WERs were calculated on the official benchmark test sets with standard lowercasing and normalization, and the reported statistic is the sample mean as claimed in Section 4.3.1.
    Algorithm 7 is the only stated protocol; Section 4.3.6 reports min/max/median instead, and the headline numbers match the median column.
  • domain assumption Neither LibriSpeech test-clean, Europarl-ASR guest test, TED-LIUM, nor FLEURS transcripts leaked into the synthetic training corpus or the BART semantic enhancer training.
    No leakage control is described; because the synthetic sentences are generated by GPT and the benchmark text is public, contamination is possible.
  • domain assumption The BART semantic enhancer's post-edits reduce or preserve WER rather than inflating it.
    The paper reports only the final WER after BART correction; it does not compare with and without the enhancer, so the assumed net benefit is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-precision medical speech recognition through synthetic data and semantic correction: UNITED-MEDASR." pith.science (2026). https://pith.science/paper/BMH7HSP5

@misc{pith2026241200055,
  author       = {Pith},
  title        = {Pith review of: High-precision medical speech recognition through synthetic data and semantic correction: UNITED-MEDASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMH7HSP5}},
  note         = {Machine review of arXiv:2412.00055}
}
read the original abstract

Automatic Speech Recognition (ASR) systems in the clinical domain face significant challenges, notably the need to recognise specialised medical vocabulary accurately and meet stringent precision requirements. We introduce United-MedASR, a novel architecture that addresses these challenges by integrating synthetic data generation, precision ASR fine-tuning, and advanced semantic enhancement techniques. United-MedASR constructs a specialised medical vocabulary by synthesising data from authoritative sources such as ICD-10 (International Classification of Diseases, 10th Revision), MIMS (Monthly Index of Medical Specialties), and FDA databases. This enriched vocabulary helps finetune the Whisper ASR model to better cater to clinical needs. To enhance processing speed, we incorporate Faster Whisper, ensuring streamlined and high-speed ASR performance. Additionally, we employ a customised BART-based semantic enhancer to handle intricate medical terminology, thereby increasing accuracy efficiently. Our layered approach establishes new benchmarks in ASR performance, achieving a Word Error Rate (WER) of 0.985% on LibriSpeech test-clean, 0.26% on Europarl-ASR EN Guest-test, and demonstrating robust performance on Tedlium (0.29% WER) and FLEURS (0.336% WER). Furthermore, we present an adaptable architecture that can be replicated across different domains, making it a versatile solution for domain-specific ASR systems.

Figures

Figures reproduced from arXiv: 2412.00055 by the authors.

Figure 1
Figure 1. End-to-End Workflow of United-MedASR ASR System Development [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Synthetic Data Pipeline and United-MedASR Training Process [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Synthetic Data Pipeline and United-MedASR Training Process [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance Metrics of Fine Tuning of the Whisper and Bart-Base on Clinical Data. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: United-MedASR Benchmarks Evaluation Flow. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The Word Error Rate (WER) performance of different models on four benchmark datasets: (a) Librispeech [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 30 canonical work pages

  1. [1]

    Y . Lu, M. Shen, H. Wang, X. Wang, C. van Rechem, and W. Wei. Machine Learning for Synthetic Data Generation: A Review. arXiv preprint arXiv:2302.04062, 2023. https://ar5iv.labs.arxiv.org/html/2302.04062

    arXiv 2023

  2. [2]

    Fonseca and F

    J. Fonseca and F. Bacao. Tabular and Latent Space Synthetic Data Generation: A Literature Re- view. Journal of Big Data , 2023. https://journalofbigdata.springeropen.com/articles/10.1186/ s40537-023-00792-7

    work page 2023

  3. [3]

    Deep Generative Models, Synthetic Tabular Data, and Differential Privacy: An Overview and Synthesis

    C. Hassan, R. Salomone, and K. Mengersen. Deep Generative Models, Synthetic Tabular Data, and Differential Privacy: An Overview and Synthesis. arXiv preprint arXiv:2307.15424, 2023. https://ar5iv.labs.arxiv. org/html/2307.15424

    work page Pith review arXiv 2023

  4. [4]

    Schneider

    F. Schneider. ArchiSound: Audio Generation with Diffusion. arXiv preprint arXiv:2301.13267v1, 2023. https: //arxiv.org/pdf/2301.13267v1

    arXiv 2023

  5. [5]

    Fazel, W

    A. Fazel, W. Yang, Y . Liu, R. Barra-Chicote, Y . Meng, R. Maas, and J. Droppo. SynthASR: Unlocking Synthetic Data for Speech Recognition. arXiv preprint arXiv:2106.07803, 2021. https://arxiv.org/abs/2106.07803

    arXiv 2021

  6. [6]

    Amazon Transcribe

    Amazon Web Services. Amazon Transcribe. https://aws.amazon.com/transcribe

    work page

  7. [7]

    Microsoft Azure Speech-to-Text

    Microsoft Corporation. Microsoft Azure Speech-to-Text. https://azure.microsoft.com/en-us/services/ cognitive-services/speech-to-text

    work page

  8. [8]

    OpenAI. OpenAI. https://openai.com/

    work page

Show all 50 references
  1. [9]

    Anibal, H

    J. Anibal, H. Huth, M. Li, L. Hazen, Y . M. Lam, H. Nguyen, P. Hong, M. Kleinman, S. Ost, C. Jackson, L. Sprabery, C. Elangovan, B. Krishnaiah, L. Akst, I. Lina, I. Elyazar, L. Ekwati, S. Jansen, R. Nduwayezu, C. Garcia, J. Plum, J. Brenner, M. Song, E. Ricotta, D. Clifton, C....

    2024 arXiv

  2. [10]

    Ibrahim, Y

    M. Ibrahim, Y . A. Khalil, S. Amirrajab, C. Sun, M. Breeuwer, J. Pluim, B. Elen, G. Ertaylan, and M. Dumon- tier. Generative AI for Synthetic Data Across Multiple Medical Modalities: A Systematic Review of Recent Developments and Challenges. arXiv preprint arXiv:2407.00116, 20...

    2024 arXiv

  3. [11]

    J. Lee, Y . Jeon, W. Lee, Y . Kim, and G. G. Lee. Exploring the Viability of Synthetic Audio Data for Audio-Based Dialogue State Tracking. arXiv preprint arXiv:2312.01842, 2023. https://arxiv.org/abs/2312.01842

    2023 arXiv

  4. [12]

    Goncalves, P

    A. Goncalves, P. Ray, B. Soper, J. Stevens, L. Coyle, and A. P. Sales. Generation and Evaluation of Synthetic Patient Data. PubMed, 2020. https://pubmed.ncbi.nlm.nih.gov/32381039

    2020

  5. [13]

    Yu and L

    D. Yu and L. Deng. Deep Neural Network-Hidden Markov Model Hybrid Systems. In Springer Handbook of Speech Processing, 2017. https://link.springer.com/chapter/10.1007/978-3-319-57624-9_16

    2017 doi

  6. [14]

    Arriaga, A

    C. Arriaga, A. Pozo, J. Conde, and A. Alonso. Evaluation of Real-Time Transcriptions Using End-to-End ASR Models. arXiv preprint arXiv:2409.05674, 2024. https://arxiv.org/abs/2409.05674

    2024

  7. [15]

    Radford, J

    A. Radford, J. W. Kim, T. Xu, G. Brockman, and C. McLeavey. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv preprint arXiv:2212.04356, 2022. https://arxiv.org/pdf/2212.04356

    2022 arXiv

  8. [16]

    D. Hwang. FAdam: Adam is a Natural Gradient Optimizer Using Diagonal Empirical Fisher Information. arXiv preprint arXiv:2405.12807v9, 2024. https://arxiv.org/pdf/2405.12807v9

    2024 arXiv

  9. [17]

    Rekesh, N

    D. Rekesh, N. R. Koluguri, S. Kriman, S. Majumdar, V . Noroozi, H. Huang, O. Hrinchuk, K. Puvvada, A. Kumar, J. Balam, and B. Ginsburg. Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition. arXiv preprint arXiv:2305.05084v6, 2023. https://arxiv.org/...

    2023 arXiv

  10. [18]

    T. V . Ho, S. Horiguchi, S. Watanabe, P. Garcia, and T. Sumiyoshi. Synthetic Data Augmentation for ASR with Domain Filtering. IEEE, 2023. https://ieeexplore.ieee.org/document/10317120

    2023

  11. [19]

    Y . Fang, X. Li. Mamba for Streaming ASR Combined with Unimodal Aggregation. arXiv preprint arXiv:2410.00070v1, 2024. https://arxiv.org/html/2410.00070v1 13 A PREPRINT - DECEMBER 3, 2024

    2024 arXiv

  12. [20]

    Lindsay, J

    H. Lindsay, J. Troger, M. Mina, N. Linz, P. Muller, J. Alexandersson, and I. Ramakers. Generating Synthetic Clinical Speech Data Through Simulated ASR Deletion Error. ACL Anthology, 2022. https://aclanthology. org/2022.rapid-1.2.pdf

    2022

  13. [21]

    Rossenbach, A

    N. Rossenbach, A. Zeyer, R. Schlüter, and H. Ney. Generating Synthetic Audio Data for Attention-Based Speech Recognition Systems. IEEE, 2020. https://ieeexplore.ieee.org/document/9053008

    2020

  14. [22]

    Y . A. Li, C. Han, V . S. Raghavan, G. Mischler, and N. Mesgarani. StyleTTS 2: Towards Human-Level Text-to- Speech through Style Diffusion and Adversarial Training with Large Speech Language Models. arXiv preprint arXiv:2306.07691, 2023. https://arxiv.org/abs/2306.07691

    2023 arXiv

  15. [23]

    Gambs, M.-O

    S. Gambs, M.-O. Killijian, and M. Núñez del Prado Cortez. De-anonymization Attack on Geolocated Data. Journal of Computer and System Sciences, 2014. https://www.sciencedirect.com/science/article/pii/ S0022000014000683

    2014

  16. [24]

    Patki, R

    N. Patki, R. Wedge, and K. Veeramachaneni. The Synthetic Data Vault. IEEE, 2016. https://ieeexplore. ieee.org/document/7796926

    2016

  17. [25]

    https://www.mims.com/india

    MIMS. https://www.mims.com/india

  18. [26]

    https://www.accessdata.fda.gov/scripts/cder/daf/index.cfm

    FDA. https://www.accessdata.fda.gov/scripts/cder/daf/index.cfm

  19. [27]

    https://www.icd10data.com/ICD10CM/Codes

    ICD 10. https://www.icd10data.com/ICD10CM/Codes

  20. [28]

    https://cloud.google.com/

    Google Cloud. https://cloud.google.com/

  21. [29]

    Scikit-learn: Machine Learning in Python

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V ., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V ., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. Scikit-learn: Machine Learning in Python. Journal of...

    2011

  22. [30]

    Transformers: State-of-the-Art Natural Language Processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R’emi Louf, Morgan Funtowicz, and Jamie Brew. Transformers: State-of-the-Art Natural Language Processing. Proceedings of the 2020 Conference on Empirical Method...

    2020

  23. [31]

    Faster Whisper

    SYSTRAN. Faster Whisper. https://github.com/SYSTRAN/faster-whisper

  24. [32]

    Lewis, Y

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. CoRR, vol. abs/1910.13461, 2019. http://arxiv.org/abs/1910.13461

    1910 arXiv

  25. [33]

    Panayotov, G

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur. LIBRISPEECH: An ASR Corpus Based on Public Domain Audio Books. Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2015. https://www.danielpovey.com/files/2015_icassp_librispeech.pdf

    2015

  26. [34]

    PyTorch. PyTorch. https://pytorch.org/get-started/locally/

  27. [35]

    G. V . Garces Díaz-Munío, J. A. Silvestre-Cerda, J. Jorge, A. Gimenez, J. Iranzo-Sánchez, P. Baquero-Arnal, N. Roselló, A. Perez-González-de-Martos, J. Civera, A. Sanchis, and A. Juan. Europarl-ASR: A Large Corpus of Parliamentary Debates for Streaming ASR Benchmarking and Spe...

  28. [36]

    Conneau, M

    A. Conneau, M. Ma, S. Khanuja, Y . Zhang, V . Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna. FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. arXiv preprint arXiv:2205.12446, 2022. https://arxiv.org/abs/2205.12446

    2022 arXiv

  29. [37]

    Sainburg, M

    T. Sainburg, M. Thielk, and T. Q. Gentner. Finding, Visualizing, and Quantifying Latent Structure Across Diverse Animal V ocal Repertoires.PLoS Computational Biology , vol. 16, no. 10, pp. e1008228, 2020. https: //doi.org/10.1371/journal.pcbi.1008228

    2020 doi

  30. [38]

    Sainburg

    T. Sainburg. timsainb/noisereduce: v1.0. Zenodo, version db94fe2, Jun. 2019. doi: 10.5281/zenodo.3243139. https://doi.org/10.5281/zenodo.3243139

    2019 doi

  31. [39]

    Dragon Medical One - The #1 Clinical Speech Recognition Solution

    Nuance Communications, Inc. Dragon Medical One - The #1 Clinical Speech Recognition Solution. https: //www.nuance.com/healthcare/clinical-documentation/dragon-medical-one.html

  32. [40]

    M*Modal Fluency for Transcription - Next-generation Clinical Documentation

    3M. M*Modal Fluency for Transcription - Next-generation Clinical Documentation. https://www.3m.com/ 3M/en_US/health-information-systems-us/solutions/fluency-for-transcription/

  33. [41]

    Google Speech-to-Text

    Google Cloud. Google Speech-to-Text. https://cloud.google.com/speech-to-text 14 A PREPRINT - DECEMBER 3, 2024

    2024

  34. [42]

    Chung, Y

    Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu. W2V-BERT: Combining Con- trastive Learning and Masked Language Modeling for Self-Supervised Speech Pre-Training. arXiv preprint arXiv:2108.06209v2, 2021. https://arxiv.org/pdf/2108.06209v2

    2021 arXiv

  35. [43]

    Zhang, J

    Y . Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V . Le, and Y . Wu. Pushing the Limits of Semi-Supervised Learning for Automatic Speech Recognition. arXiv preprint arXiv:2010.10504v2, 2020. https: //arxiv.org/pdf/2010.10504v2

    2010 arXiv

  36. [45]

    C. Chen, Y . Hu, C.-H. H. Yang, S. M. Siniscalchi, P.-Y . Chen, and E. S. Chng. HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models. arXiv preprint arXiv:2309.15701v2 , 2023. https://arxiv.org/pdf/2309.15701v2

    2023 arXiv

  37. [46]

    W. Chan, D. S. Park, C. A. Lee, Y . Zhang, Q. V . Le, and M. Norouzi. SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network. arXiv preprint arXiv:2104.02133v3, 2021. https: //arxiv.org/pdf/2104.02133v3

    2021 arXiv

  38. [47]

    Barrault, Y .-A

    L. Barrault, Y .-A. Chung, M. C. Meglioli, D. Dale, N. Dong, P.-A. Duquenne, H. Elsahar, H. Gong, K. Hef- fernan, J. Hoffman, et al. SeamlessM4T: Massively Multilingual & Multimodal Machine Translation, Seamless Communication. arXiv preprint arXiv:2308.11596v3, 2023. https://a...

    2023 arXiv

  39. [48]

    Rousseau, P

    A. Rousseau, P. Deléglise, and Y . Estève. TED-LIUM: An Automatic Speech Recognition Dedicated Corpus. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12) , May 2012. https://aclanthology.org/L12-1405/

    2012

  40. [49]

    United-MedSyn: Medical Speech Dataset for ASR

    United We Care. United-MedSyn: Medical Speech Dataset for ASR. Dataset, 2024. https://huggingface. co/datasets/united-we-care/United-Syn-Med License: CC BY-SA 4.0. 15

    2024

  41. [2020]

    https://arxiv.org/pdf/2010.11430v1

    2010 arXiv

  42. [2021]

    https://www.isca-archive.org/interspeech_2021/garcesdiazmunio21_interspeech.pdf

Pith tools

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