Pith. sign in

REVIEW 3 major objections 8 minor 34 references

MLLP-VRAIN UPV system for the IWSLT 2025 Simultaneous Speech Translation Translation task

T0 review · 3 major / 8 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 Off-the-shelf models, lightly adapted, hit 31.96 BLEU

desk verdict A solid, transparent shared-task system paper with a cheap recipe for cascaded SimulST; the headline latency metric flatters the tails, but the paper reports them and the central claim holds. read the letter →

arxiv 2506.18828 v1 pith:LPVQB7ZE submitted 2025-06-23 cs.CL

Jorge Iranzo-Sánchez , Javier Iranzo-Sánchez , Adrià Giménez , Jorge Civera , Alfons Juan This is my paper
classification cs.CL
keywords simultaneousspeechtranslationcascadesystemstreamingmachineprefixtrainingwait-kpolicyRALCPlong-formaudioWhisper
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 argues that a simultaneous speech translation system for long-form content can be built by adapting existing pre-trained offline models rather than training a specialized end-to-end system. The authors combine Whisper Large-V3-Turbo for speech recognition with NLLB-3.3B for translation, and stream them through document-level prefix training, a wait-k start, a risk-aware emission policy, and an attention-based history tracker. On the ACL60/60 development set, the adapted cascade achieves 31.96 BLEU at 2.94 seconds mean non-computational-aware latency, and 29.8 BLEU on the official IWSLT25Instruct test set. The claim matters because it points to a cheap, modular route to real-time translation of long talks, one that does not depend on scarce in-domain speech-translation data.

What carries the argument

The central mechanism is the streaming-history adaptation of the MT model: during training, preceding sentences are prepended to the active sentence with a [SEP] sentinel, teaching a pretrained offline translation model to translate from an incomplete stream while tracking what has already been covered. At inference, the position of the last source word moved to history is computed as the argmax of the attention score from the token immediately before [SEP] across source positions. This attention-based segmentation feature, inspired by attention-map alignment, is what makes the system's history management work without a separate reverse translation model. The emission side rests on a wait-k policy that guards the start of each phrase and RALCP, which uses beam-search hypotheses to decide when to emit; the ASR side contributes a relaxed longest-common-prefix policy that stabilizes streaming recognition.

What would settle it

Reproduce the system on a public long-form speech corpus while logging per-segment computational-aware StreamLAAL; if user-visible delay averages near the computational-aware value (4.20 s on the dev set) rather than the non-computational-aware 2.94 s, the claimed real-time balance shifts to a much slower regime.

Watch

Extended reading notes

Core claim

The authors claim that carefully adapted pre-trained components can create effective simultaneous translation systems for long-form content without extensive in-domain parallel data or specialized end-to-end training. Concretely, they adapt Whisper Large-V3-Turbo on the ASR side with a longest-common-prefix streaming policy (relaxed by lowercasing, removing punctuation, and accepting a Levenshtein distance of two) and adapt NLLB-3.3B on the translation side with DoRA fine-tuning over document-level data. Prefix training prepends one to ten previous sentences with a [SEP] sentinel, teaching the model to translate from incomplete input; a learned attention feature, the maximum attention score from the token before [SEP], selects the source word that has been translated and moves it to a history buffer. Emissions are governed by a wait-k policy at phrase starts and RALCP on cleaned beam hypotheses. The resulting system beats the authors' untrained naive cascade by 5.86 BLEU while lowering mean StreamLAAL latency from 3.61 to 2.94 seconds, and its own offline-quality loss from adaptation is only 1.64 BLEU.

Load-bearing premise

The favorable quality-latency balance assumes that the non-computational-aware mean latency of 2.94 seconds represents what users actually experience, whereas the paper's own computational-aware figures are 4.20 seconds mean and 15.00 seconds maximum, driven by Whisper's temperature fallback.

Editorial extensions

If this is right

  • A simultaneous translation system for long-form audio can be assembled in days from open-weight models, with training limited to generic document-level data and no in-domain ACL speech-translation corpus.
  • Cascade architectures remain a competitive route for real-time translation: the adapted cascade surpasses both the organizers' baseline and the authors' naive baseline on quality and latency.
  • Latency risk in this design concentrates in the ASR component's temperature-fallback behavior, which causes rare spikes to 15.00 seconds in the computational-aware worst case, so improving ASR emission stability is a higher-leverage target than further MT tuning.
  • The full system's peak memory footprint of about 20 GB of VRAM fits on a single high-end GPU, making the recipe practical for deployment.

Reading between the lines

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

  • The same adaptation recipe is language-agnostic in principle, so the English-to-German result suggests a cheap path to other long-form language pairs, though the paper only validates one direction.
  • The attention-based history segmentation could double as a general alignment signal, potentially useful for forcing monotonic decoding or diagnosing when a pretrained MT model loses track of the source span in streaming conditions.
  • If user-perceived latency includes ASR computation time, the system's practical edge narrows: the paper reports a computational-aware mean of 4.20 seconds and a maximum of 15.00 seconds, so deployments would need to suppress or bound the temperature-fallback spikes to keep the advertised real-time feel.
  • The authors' rejection of MADLAD-400 due to suspiciously high offline BLEU (over 95) highlights how data contamination can distort model selection in this task, suggesting that contamination checks should be standard practice when choosing pretrained backbones.
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

3 major / 8 minor

Summary. The paper describes the MLLP-VRAIN UPV system for the IWSLT 2025 Simultaneous Speech Translation task (English–German). The system is a cascade: Whisper Large-V3-Turbo for streaming ASR and NLLB-3.3B for MT, adapted via document-level prefix training with a [SEP] sentinel token, an attention-based history mechanism, and wait-k plus RALCP emission policies. On the ACL60/60 development set, the adapted system achieves 31.96 BLEU and 0.732 COMET with a mean non-computational-aware StreamLAAL of 2.94 s (computational-aware: 4.20 s); the official test set score is a preliminary 29.8 BLEU. The central claim is that carefully adapted pre-trained components can create effective simultaneous translation systems for long-form content without extensive in-domain parallel data or end-to-end training.

Significance. The paper provides a practical demonstration that a cascade of strong pre-trained ASR and MT models, with lightweight adaptation, can yield a competitive SimulST system on a challenging long-form domain. The use of public models and an open training recipe (Europarl, News Commentary with Paradocs) makes the approach fairly reproducible. The authors are transparent in reporting both NCA and CA latency and in showing tail percentiles (Table 2), which is a creditworthy aspect of the evaluation. If the system's latency is acceptable to the task's requirements, the work supports the view that end-to-end training is not the only route to effective SimulST. The main weaknesses are the heavy reliance on a single dev set, the lack of significance tests, and the fact that the 'favorable balance' claim depends on the choice of latency metric.

major comments (3)
  1. [Section 3 (Table 2) and Abstract] The headline claim of a 'favorable balance between translation quality and latency' is based on the mean non-computational-aware (NCA) StreamLAAL of 2.94 s, but the paper's own Table 2 reports a computational-aware (CA) mean of 4.20 s, a p99 CA of 10.73 s, and a maximum of 15.00 s. For real-time interpretation, users experience latency per segment rather than as a talk-level mean, so tail events such as a 15 s stall are likely to be noticeable and could be disqualifying in a production setting. The authors honestly report these numbers, but the abstract and the 'production ready' language in Section 1 overstate the case; please qualify the latency claim by reporting CA as primary or adding an explicit discussion of tail behavior.
  2. [Section 2.2 (History Mechanism, Eq. (1))] The attention-based history mechanism is described only schematically. Equation (1) selects the source position via an argmax of attention scores A(x_i, y_[SEP]-1), but the paper does not specify how A is aggregated across layers and heads, even though the hyperparameter table lists a specific 'Attention Head Layer' (6). The 'lightweight log-linear model' mentioned in the same section is not described at all. This makes the method hard to reproduce and the contribution difficult to evaluate; please provide the missing details or a precise pointer to the earlier work that defines them.
  3. [Section 3 (Table 1)] The evaluation on the ACL60/60 development set reports single-point estimates with no error bars, confidence intervals, or significance tests. The central claim about the benefit of adaptation rests on the gap between the adapted (31.96 BLEU) and baseline (26.10 BLEU) systems; it would be valuable to know whether this difference is statistically stable, particularly since the official test set score is only preliminary. Please add at least a statement of this limitation or a resampling-based uncertainty estimate.
minor comments (8)
  1. [Title] The title contains a duplicated word: 'Translation Translation'.
  2. [Section 2.1] 'V AD filtering' should be 'VAD filtering'.
  3. [Section 2.2] 'News Comentary' should be 'News Commentary'.
  4. [Section 3] 'recomendations' should be 'recommendations'.
  5. [Section 3 (Table 1 caption)] 'computationally an computationally aware' should be 'non-computational and computational aware'.
  6. [Section 3] 'Surpassingly' should be 'Surprisingly'.
  7. [Section 2.2] 'founding BLEU scores' should be 'finding BLEU scores'.
  8. [Section 2.2] The adjustment of RALCP λ after filtering invalid beam hypotheses is described only qualitatively; please report the specific λ value used.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the system's quality and latency results are genuine held-out measurements, and self-citations are descriptive rather than load-bearing.

full rationale

The central claim, that carefully adapted pre-trained components can form an effective simultaneous translation system, is supported by measurements on held-out data: the ACL60/60 development set and the official IWSLT25Instruct test set (Section 3). Neither set is used for training the adapted MT model, which is fine-tuned only on Europarl and News Commentary documents (Section 2.2). No parameter is fitted to the evaluation targets, and the reported BLEU and StreamLAAL numbers are genuine system outputs rather than quantities derived from the inputs by construction. The self-citations to Iranzo-Sánchez et al. (2022, 2024) describe the prefix-training and streaming-history mechanism that inspires the adaptation, but the paper supplies its own implementation details and evaluates the resulting system independently; no uniqueness theorem or unverified ansatz is imported through these citations. The choice of headline latency (NCA mean 2.94s) over computational-aware or tail metrics is a legitimate measurement-interpretation concern, not circularity, and the paper transparently reports CA mean, p99, and max values in Table 2. The note about MADLAD-400's suspiciously high offline BLEU is a data-contamination observation, not a derivation step. The stated limitation on limited hyperparameter search is an honest scope note, not evidence that results are forced. Overall, the derivation chain is self-contained: the adaptation recipe, the baselines, and the evaluation are all concretely described and empirically grounded, with no fitted input renamed as a prediction and no load-bearing argument resting solely on the authors' prior work.

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

The system is an engineering composition of existing models and methods. The tuned hyperparameters (wait-k, RALCP lambda, buffers, DoRA rank) are free parameters that affect the reported quality-latency tradeoff. The axioms are standard domain assumptions about metrics, model quality, and evaluation data, none of which are ad hoc to this paper. No new conceptual entities are introduced.

free parameters (6)
  • MT wait-k = 3
    Number of source tokens collected before translation starts at the beginning of a phrase; tuned on the ACL60/60 dev set and reported in Appendix A.
  • RALCP lambda = 0.5
    Threshold controlling when the [SEP] sentinel is emitted; adjusted after filtering empty beam hypotheses so the ratio of hypotheses to lambda remains similar (Section 2.2).
  • Attention head layer = 6
    Decoder layer from which attention scores are taken for the [SEP] segmentation feature; chosen during development and listed in Appendix A.
  • DoRA rank = r=16
    Rank of the low-rank adaptation applied to NLLB during fine-tuning; a hyperparameter of the adaptation procedure (Appendix A).
  • ASR initial wait / LCP chunk = 1s / 1s
    Streaming parameters for Whisper, set after initial tests on development data (Section 2.1).
  • Max audio buffer = 30s
    Upper limit on the audio window used by the ASR streaming policy (Section 2.1).
assumptions (4)
  • domain assumption Cascade systems generally achieve superior performance over current end-to-end alternatives in speech translation.
    Stated in Section 1 and referenced to prior IWSLT campaigns; this motivates the architectural choice.
  • domain assumption BLEU, COMET, and StreamLAAL are valid metrics for evaluating simultaneous speech translation quality and latency.
    Used throughout Section 3; the paper follows the shared task recommendations, including setting COMET to 0 for empty translations.
  • domain assumption Whisper Large-V3-Turbo and NLLB-3.3B are strong, general-purpose pretrained models for ASR and MT respectively.
    Invocations in Sections 2.1 and 2.2, based on public benchmarks rather than fine-tuning for the target domain.
  • domain assumption The ACL60/60 development set is representative enough for tuning system hyperparameters and comparing against baselines.
    Used for all reported development results in Section 3; the official test set is reported separately.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MLLP-VRAIN UPV system for the IWSLT 2025 Simultaneous Speech Translation Translation task." pith.science (2026). https://pith.science/paper/LPVQB7ZE

@misc{pith2026250618828,
  author       = {Pith},
  title        = {Pith review of: MLLP-VRAIN UPV system for the IWSLT 2025 Simultaneous Speech Translation Translation task},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPVQB7ZE}},
  note         = {Machine review of arXiv:2506.18828}
}
abstract

This work describes the participation of the MLLP-VRAIN research group in the shared task of the IWSLT 2025 Simultaneous Speech Translation track. Our submission addresses the unique challenges of real-time translation of long-form speech by developing a modular cascade system that adapts strong pre-trained models to streaming scenarios. We combine Whisper Large-V3-Turbo for ASR with the multilingual NLLB-3.3B model for MT, implementing lightweight adaptation techniques rather than training new end-to-end models from scratch. Our approach employs document-level adaptation with prefix training to enhance the MT model's ability to handle incomplete inputs, while incorporating adaptive emission policies including a wait-$k$ strategy and RALCP for managing the translation stream. Specialized buffer management techniques and segmentation strategies ensure coherent translations across long audio sequences. Experimental results on the ACL60/60 dataset demonstrate that our system achieves a favorable balance between translation quality and latency, with a BLEU score of 31.96 and non-computational-aware StreamLAAL latency of 2.94 seconds. Our final model achieves a preliminary score on the official test set (IWSLT25Instruct) of 29.8 BLEU. Our work demonstrates that carefully adapted pre-trained components can create effective simultaneous translation systems for long-form content without requiring extensive in-domain parallel data or specialized end-to-end training.

Figures

Figures reproduced from arXiv: 2506.18828 by the authors.

Figure 1
Figure 1. System diagram of our cascaded system for [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 22 canonical work pages

  1. [1]

    McCrae, Salima Mdhaffar, Yasmin Moslem, Kenton Murray, Satoshi Nakamura, Matteo Negri, Jan Niehues, Atul Kr

    Idris Abdulmumin, Victor Agostinelli, Tanel Alumäe, Antonios Anastasopoulos, Ashwin, Luisa Bentivogli, Ondřej Bojar, Claudia Borg, Fethi Bougares, Roldano Cattoni, Mauro Cettolo, Lizhong Chen, William Chen, Raj Dabre, Yannick Estève, Marcello Federico, Marco Gaido, Dávid Javorský, Marek Kasztelnik, Tsz Kin Lam, Danni Liu, Evgeny Matusov, Chandresh Kumar M...

    work page 2025

  2. [2]

    Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 2022. https://openreview.net/forum?id=shpkpVXzo3h 8-bit optimizers via block-wise quantization . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    work page 2022

  3. [3]

    Kosuke Doi, Yuka Ko, Mana Makinae, Katsuhito Sudoh, and Satoshi Nakamura. 2024. https://doi.org/10.18653/v1/2024.iwslt-1.30 Word order in E nglish- J apanese simultaneous interpretation: Analyses and evaluation using chunk-wise monotonic translation . In Proceedings of the 21st International Conference on Spoken Language Translation (IWSLT 2024), pages 25...

    work page doi:10.18653/v1/2024.iwslt-1.30 2024

  4. [4]

    Javier Iranzo-S \'a nchez, Jorge Civera, and Alfons Juan. 2022. https://doi.org/10.18653/v1/2022.acl-long.480 From simultaneous to streaming machine translation by leveraging streaming history . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6972--6985, Dublin, Ireland. Association...

    work page doi:10.18653/v1/2022.acl-long.480 2022

  5. [5]

    Javier Iranzo-S \'a nchez, Jorge Iranzo-S \'a nchez, Adri \`a Gim \'e nez, Jorge Civera, and Alfons Juan. 2024. https://doi.org/10.1162/tacl_a_00691 Segmentation-free streaming machine translation . Transactions of the Association for Computational Linguistics, 12:1104--1121

    doi:10.1162/tacl_a_00691 2024

  6. [6]

    Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusupati, Romi Stella, Ankur Bapna, and Orhan Firat. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/d49042a5d49818711c401d34172f9900-Abstract-Datasets\_and\_Benchmarks.html MADLAD-400: A multilingual and document-level large audited dataset . In Advances in Neural Inf...

    work page 2023

  7. [7]

    Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, Hua Wu, and Haifeng Wang. 2019. https://doi.org/10.18653/v1/P19-1289 STACL : Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework . In Proceedings of the 57th Annual Meetin...

    doi:10.18653/v1/p19-1289 2019

  8. [8]

    Dominik Mach \'a c ek, Raj Dabre, and Ond r ej Bojar. 2023. https://doi.org/10.18653/v1/2023.ijcnlp-demo.3 Turning whisper into real-time transcription system . In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: System...

    work page doi:10.18653/v1/2023.ijcnlp-demo.3 2023

Show all 34 references
  1. [9]

    Yulong Mao, Kaiyu Huang, Changhao Guan, Ganglin Bao, Fengran Mo, and Jinan Xu. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.626 Dora: Enhancing parameter-efficient fine-tuning with dynamic rank distribution . In Proceedings of the 62nd Annual Meeting of the Association for ...

    2024 doi

  2. [10]

    Evgeny Matusov, Gregor Leusch, Oliver Bender, and Hermann Ney. 2005. https://aclanthology.org/2005.iwslt-1.19 Evaluating machine translation output with automatic sentence segmentation . In Proceedings of the Second International Workshop on Spoken Language Translation, Pittsb...

    2005

  3. [11]

    Jan Niehues, Ngoc - Quan Pham, Thanh - Le Ha, Matthias Sperber, and Alex Waibel. 2018. https://doi.org/10.21437/INTERSPEECH.2018-1055 Low-latency neural speech translation . In 19th Annual Conference of the International Speech Communication Association, Interspeech 2018, Hyde...

    2018 doi

  4. [12]

    Siqi Ouyang, Xi Xu, and Lei Li. 2025. https://arxiv.org/abs/2503.02969 Infinisst: Simultaneous translation of unbounded speech with large language model . Preprint, arXiv:2503.02969

    2025 arXiv

  5. [13]

    Sara Papi, Marco Gaido, Matteo Negri, and Luisa Bentivogli. 2024. https://doi.org/10.18653/v1/2024.acl-long.202 S tream A tt: Direct streaming speech-to-text translation with attention-based audio history selection . In Proceedings of the 62nd Annual Meeting of the Association...

    2024 doi

  6. [14]

    Sara Papi, Marco Gaido, Matteo Negri, and Marco Turchi. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.11 Does Simultaneous Speech Translation need Simultaneous Models ? In Findings of the Association for Computational Linguistics : EMNLP 2022 , pages 141--153

    2022 doi

  7. [15]

    Sara Papi, Marco Turchi, and Matteo Negri. 2023. https://doi.org/10.21437/Interspeech.2023-170 AlignAtt : Using Attention-based Audio-Translation Alignments as a Guide for Simultaneous Speech Translation . In INTERSPEECH 2023 , pages 3974--3978

    2023 doi

  8. [16]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 B leu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311--3...

    2002

  9. [17]

    Peter Pol \'a k and Ond r ej Bojar. 2023. https://doi.org/10.48550/arXiv.2309.11384 Long- Form End-to-End Speech Translation via Latent Alignment Segmentation . Preprint, arXiv:2309.11384

    2023 doi arXiv

  10. [18]

    Matt Post. 2018. https://doi.org/10.18653/v1/W18-6319 A call for clarity in reporting BLEU scores . In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186--191, Brussels, Belgium. Association for Computational Linguistics

    2018 doi

  11. [19]

    Matt Post and Marcin Junczys-Dowmunt. 2024. https://arxiv.org/abs/2304.12959 Escaping the sentence-level paradigm in machine translation . Preprint, arXiv:2304.12959

    2024 arXiv

  12. [20]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. https://proceedings.mlr.press/v202/radford23a.html Robust speech recognition via large-scale weak supervision . In International Conference on Machine Learning, ICML 2023, 23-29 J...

    2023

  13. [21]

    Ricardo Rei, Jos \'e G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and Andr \'e F. T. Martins. 2022. https://aclanthology.org/2022.wmt-1.52 COMET -22: Unbabel- IST 2022 submission for the metrics shared task . In Pro...

    2022

  14. [22]

    Elizabeth Salesky, Kareem Darwish, Mohamed Al-Badrashiny, Mona Diab, and Jan Niehues. 2023 a . https://doi.org/10.18653/v1/2023.iwslt-1.2 Evaluating multilingual speech translation under realistic conditions with resegmentation and terminology . In Proceedings of the 20th Inte...

    2023 doi

  15. [23]

    Elizabeth Salesky, Marcello Federico, and Marine Carpuat, editors. 2023 b . https://aclanthology.org/2023.iwslt-1.0 Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023) . Association for Computational Linguistics, Toronto, Canada (in-per...

    2023

  16. [24]

    Elizabeth Salesky, Marcello Federico, and Marine Carpuat, editors. 2024. https://aclanthology.org/2024.iwslt-1.0 Proceedings of the 21st International Conference on Spoken Language Translation (IWSLT 2024) . Association for Computational Linguistics, Bangkok, Thailand (in-pers...

    2024

  17. [25]

    Felix Schneider and Alexander Waibel. 2020. https://doi.org/10.18653/v1/2020.iwslt-1.28 Towards stream translation: Adaptive computation time for simultaneous machine translation . In Proceedings of the 17th International Conference on Spoken Language Translation, pages 228--2...

    2020 doi

  18. [26]

    Sukanta Sen, Ondřej Bojar, and Barry Haddow. 2022. https://arxiv.org/abs/2210.09754 Simultaneous translation for unsegmented input: A sliding window approach . Preprint, arXiv:2210.09754

    2022 arXiv

  19. [27]

    Vaibhav Srivastav, Somshubra Majumdar, Nithin Koluguri, Adel Moumen, Sanchit Gandhi, et al. 2023. Open automatic speech recognition leaderboard. https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

    2023

  20. [28]

    NLLB Team. 2024. https://doi.org/10.1038/S41586-024-07335-X Scaling neural machine translation to 200 languages . Nat., 630(8018):841--846

    2024 doi

  21. [29]

    Viet Anh Khoa Tran, David Thulke, Yingbo Gao, Christian Herold, and Hermann Ney. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.297 Does joint training really help cascaded speech translation? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Pr...

    2022 doi

  22. [30]

    Minghan Wang, Thuy-Trang Vu, Jinming Zhao, Fatemeh Shiri, Ehsan Shareghi, and Gholamreza Haffari. 2024. https://aclanthology.org/2024.alta-1.7/ Simultaneous machine translation with large language models . In Proceedings of the 22nd Annual Workshop of the Australasian Language...

    2024

  23. [31]

    Rachel Wicks, Matt Post, and Philipp Koehn. 2024. https://doi.org/10.18653/v1/2024.findings-acl.589 Recovering document annotations for sentence-level bitext . In Findings of the Association for Computational Linguistics: ACL 2024, pages 9876--9890, Bangkok, Thailand. Associat...

    2024 doi

  24. [32]

    Vil \'e m Zouhar, Pinzhen Chen, Tsz Kin Lam, Nikita Moghe, and Barry Haddow. 2024. https://doi.org/10.18653/v1/2024.wmt-1.121 Pitfalls and outlooks in using COMET . In Proceedings of the Ninth Conference on Machine Translation, pages 1272--1288, Miami, Florida, USA. Associatio...

    2024 doi

  25. [33]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  26. [34]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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