Pith. sign in

REVIEW 3 major objections 8 minor 1 cited by

CMU's IWSLT 2025 Simultaneous Speech Translation System

T0 review · 3 major / 8 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 This paper reports that a simultaneous speech translation system combining a chunkwise causal Wav2Vec 2.0 encoder with a 7B-parameter LLM decoder achieves 44.3 BLEU for English-to-Chinese and 25.1 BLEU for English-to-German on the…

desk verdict A clean, honest competition system paper with a useful synthetic-data recipe; the unvalidated alignment-based training targets are the main caveat, not a fatal flaw. read the letter →

arxiv 2506.13143 v1 pith:OZQQANDM submitted 2025-06-16 cs.CL

Siqi Ouyang , Xi Xu , Lei Li This is my paper
classification cs.CL
keywords simultaneousspeechtranslationspeech-to-textWav2Vec2.0largelanguagemodeldecodersynthetictrainingdatastreaminginferencelatencyEnglish-to-Chinese
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

The paper claims that an end-to-end simultaneous speech translation system can reach strong BLEU scores—44.3 for English-to-Chinese and 25.1 for English-to-German on the ACL60/60 development set—at a few seconds of latency by combining a chunkwise causal speech encoder with a 7B-parameter instruction-tuned LLM decoder. The training relies on roughly 3,850 hours of synthetic speech-translation data created by machine-translating transcripts of public ASR corpora. The paper argues that expanding synthetic data beyond 3,000 hours still yields quality gains, and that the choice of the Qwen2.5-7B-Instruct decoder over Llama-3.1-8B-Instruct matters most for Chinese. If correct, this shows that LLM-decoded streaming translation can be trained largely on synthesized parallel data rather than scarce human-translated speech corpora.

What carries the argument

The central object is the translation trajectory: for each utterance, speech is aligned to transcript words, transcript words to translation words, and the resulting word-level right boundaries are made monotone by taking running maxima; a 960 ms fixed chunk then carries the translation span whose boundary timestamps fall inside that chunk. This trajectory supervision is what lets a chunkwise causal encoder and a multi-turn LLM decoder train on streaming speech without seeing the future. The encoder itself uses chunk-wise causal attention with a sliding window, and the decoder maintains a KV cache that concatenates the initial instruction with the most recent 1K tokens.

What would settle it

A concrete test: take a set of English utterances with human word-level time alignments, run the paper's alignment pipeline on them, measure the boundary error, and then retrain the same model with the human boundaries in place of the estimated ones; if BLEU does not change, the trajectory construction is not the load-bearing part of the reported quality.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a simultaneous speech-to-text model built from a causal Wav2Vec2.0 encoder and a frozen-large-language-model decoder, trained with plain cross-entropy loss on synthetic speech-translation trajectories, translates unsegmented English speech into Chinese and German with competitive BLEU while emitting output at 1.7–2.7 seconds of latency. The mechanism that enables this is the construction of a monotone word-level translation trajectory via forced alignment of speech to transcript and word alignment of transcript to translation, which turns each 960 ms speech chunk into a supervised target span. The paper further claims that adding synthetic data from a third corpus (VoxPopuli) improves BLEU by at least one point even after 3,000 hours of training, and that the choice of decoder LLM yields over three BLEU points on English-to-Chinese.

Load-bearing premise

The training targets assume that the two-stage alignment (speech-to-transcript then transcript-to-translation) plus a monotonicity correction produces word-level time boundaries accurate enough to tell the model which translation words correspond to each 960 ms speech chunk; if those boundaries are systematically wrong, every chunk's supervision is corrupted.

Editorial extensions

If this is right

  • Adding synthetic data from another ASR corpus improves BLEU by at least one point even past 3,000 hours, so further scaling the synthesis pipeline should keep helping both language pairs.
  • The decoder LLM is a major quality lever: Qwen2.5-7B-Instruct beats Llama-3.1-8B-Instruct by more than three BLEU points for English-to-Chinese while staying comparable for English-to-German.
  • The latency multiplier provides a single test-time knob that trades output latency for translation quality, with the reported numbers coming from multipliers of 3 (Chinese) and 2 (German).
  • Two-stage training—first freezing the LLM and updating only encoder and adapter, then LoRA-finetuning the LLM—with standard cross-entropy loss is sufficient to learn streaming speech translation without a specialized simultaneous-training objective.
  • Because the model emits an EOS token when it needs more speech, the same architecture can process unbounded input without explicit speech segmentation.

Reading between the lines

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

  • If the synthetic-data scaling pattern holds, any large unlabeled speech corpus with ASR transcripts can be converted into streaming translation training data, shifting the practical bottleneck from parallel speech-text availability to alignment accuracy.
  • The monotonicity correction (taking running maxima of word boundaries) discards future context for non-monotone language pairs, so languages with systematic reordering, such as German verb-final structures, may benefit from a bounded lookahead rather than strict chunk monotonicity.
  • Since the training data is generated from ASR transcripts, translation quality is implicitly capped by ASR errors; comparing against oracle transcripts would isolate how much of the reported BLEU is lost to transcription noise.
  • The robustness segments created by splicing utterances with silence intervals may not reproduce real acoustic conditions like overlapping speakers or background noise, so stress-testing on noisy unsegmented audio would reveal how much of the reported robustness transfers.
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 manuscript reports CMU's IWSLT 2025 simultaneous speech-to-text translation system for English-to-Chinese and English-to-German, built on the InfiniSST framework. The architecture combines a chunk-wise causal Wav2Vec 2.0 encoder, a convolutional adapter, and a 7B-parameter instruct-tuned LLM decoder (Qwen2.5-7B-Instruct or Llama-3.1-8B-Instruct) in a multi-turn dialogue formulation. Training uses two stages (encoder+adapter update followed by LoRA on the LLM) on roughly 3,850 hours of synthetic speech-translation data derived from LibriSpeech, CommonVoice, and VoxPopuli, with word-level alignments computed by MFA and SimAlign to construct chunk-level training trajectories. The paper reports 44.3/25.1 BLEU on the ACL60/60 development set with theoretical latencies of 2.2/1.7 seconds and computation-aware latencies of 2.7/2.3 seconds for English-to-Chinese/English-to-German, and an adjustable latency multiplier that trades quality against latency.

Significance. If the results hold, this is a competitive end-to-end streaming speech translation system worth recording: it demonstrates that synthetic data scaling beyond 3,000 hours continues to improve BLEU and that a 7B instruct-tuned decoder is effective for low-latency simultaneous translation. The paper's strengths are its concreteness and reproducibility: the architecture modifications to Wav2Vec 2.0 (RoPE, chunk-wise causal attention, sliding window), the explicit two-stage training recipe, the data synthesis pipeline, and the latency-multiplier design are all described in sufficient detail for a systems paper. The ablations across data sources and decoder choices are informative and internally consistent. The main weakness is that the alignment-based training-target construction is not validated, and all empirical claims rest on single point estimates on one development set.

major comments (3)
  1. [Section 3.2] The speech-to-translation-word mapping that defines every chunk-level training target (MFA for speech-to-transcript, SimAlign with LaBSE for transcript-to-translation, and the monotonicity correction m_i = max(m_i, m_{i-1})) is never validated. No alignment error rate, no manual sample inspection, and no ablation against an alternative trajectory (e.g., uniform duration-proportional boundaries) is reported. Since both training stages use these targets, a systematic alignment error would corrupt the supervision and invalidate the attribution of the reported BLEU/latency gains to the model and data. Please add an alignment-quality evaluation (for instance, AER on a subset or a few visual examples) and an ablation demonstrating that the alignment-based trajectory is more informative than a simple uniform-chunk baseline.
  2. [Table 2 / Section 4.2] The central empirical claims—that adding VoxPopuli synthetic data 'consistently improves BLEU scores by at least 1 point' and that Qwen2.5 'significantly outperforms' Llama-3.1—are based on single runs on one development set, with no variance estimates, significance tests, or held-out test results. BLEU differences of 1.5–2.6 points on a small dev set can lie within run-to-run noise. Please soften these claims to describe point estimates on this development set, and if possible provide multiple runs or a significance check; at a minimum, explicitly state that these are single development-set observations.
  3. [Section 3.2, monotonicity correction] The operation m_i = max(m_i, m_{i-1}) forces monotone boundaries by postponing any target word whose raw alignment boundary precedes the previous word's boundary to the later chunk. For English-to-Chinese and English-to-German, where reordering is frequent, this systematically delays target tokens relative to their source evidence even under perfect alignments, potentially biasing the learned latency-quality policy. The paper neither quantifies the magnitude of this correction (e.g., the fraction of words whose m_i is changed) nor discusses its effect on the reported latency. Please measure the amount of correction applied and comment on its consequences for the learned simultaneous policy.
minor comments (8)
  1. [Section 2] StreamLAAL is cited as Papi et al. (2024a), but the reference list identifies 2024a as the StreamAtt paper; the StreamLAAL definition appears in Papi et al. (2024b). Please correct the citation and also define StreamLAAL_CA explicitly, as it is used in Table 2 without explanation.
  2. [Table 2] The units for columns B and C are not stated in the table; the abstract reports seconds, whereas the table values are in milliseconds. Please specify the units in the table header or caption.
  3. [Abstract / Section 3.2] The dataset name is spelled 'CommonV oice' in the abstract and table, whereas the correct spelling is 'CommonVoice' (as used elsewhere in the text). Please fix the typo.
  4. [Section 3.2] The distribution for sampling the latency multiplier m (m ≤ 12) is not specified; state whether it is uniform over {1, ..., 12} or follows some other distribution, as this affects the reproducibility of the data augmentation.
  5. [Section 3.2] When a robust segment is shifted to align its start with an utterance boundary, clarify whether the segment length is kept to 30 chunks or allowed to vary; the current description is ambiguous.
  6. [Section 4.1] The phrase 'maximum effective batch size of 57.6K tokens' is ambiguous; specify whether this is per GPU, per step, or the global batch, and how the maximum is enforced.
  7. [Section 4.1 / Conclusion] Minor language issues: 'ngram_no_repeat 5' should be written as 'no_repeat_ngram_size=5' for clarity, and the conclusion contains 'a adapter', which should be 'an adapter'.
  8. [Figure 1] Figure 1 is referenced in Section 3.1 but the callout is not explicit in the provided text; please add a visible reference to the figure at the point where the three components are described.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: headline BLEU and latency figures are external measurements on ACL60/60; the InfiniSST self-citation is framework-level provenance, not a derivation forcing the reported numbers.

full rationale

The paper is an experimental system report: the headline claims (44.3/25.1 BLEU; StreamLAAL 2.189/1.689 s; StreamLAAL_CA 2.739/2.306 s on the ACL60/60 development set) are measurements against human references by the task's official metric tooling, not quantities derived by construction from the paper's own equations. The only self-citation with any structural role is the reliance on InfiniSST (Ouyang, Xu, and Li, the same three authors) for the multi-turn LLM decoder, KV-cache strategy, and chunk-wise causal encoder (Sections 1 and 3.1). That citation is architectural provenance, not a load-bearing proof: the actual contributions—3,854 hours of synthesized training data, the Llama-vs-Qwen decoder comparison, and the final operating points—are independently measured and would not change if the InfiniSST reference were removed. The latency multiplier is an explicitly hand-set inference knob (Section 4.1: 'We set test-time latency multiplier to 3 for English-to-Chinese and 2 for English-to-German'), so the reported latencies are measurements at a chosen operating point, not fitted values renamed as predictions. Section 3.2's trajectory construction (MFA speech-transcript alignment, SimAlign transcript-translation alignment, monotone correction m_i = max(m_i, m_{i-1})) defines training targets only; BLEU and StreamLAAL are computed on held-out speech with independent references. The skeptic's concern that MFA/SimAlign boundaries might be systematically wrong and corrupt every chunk-level target is a legitimate data-validity risk that could bias the reported numbers, but it is not circularity: corrupted targets would degrade or distort the measured scores, not make those scores equal their inputs by construction. Verdict: no significant circularity; score 2 solely for the minor, non-load-bearing self-citation.

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

The headline numbers are empirical outputs of a heavily engineered pipeline; they rest on the transferability of pretrained components, the fidelity of LLM-synthesized translations, and the accuracy of automatic alignments used to build chunk-level supervision. These are domain assumptions rather than derived results, and none are independently validated in the paper.

free parameters (6)
  • test-time latency multiplier = 3 for EN->ZH, 2 for EN->DE
    Hand-selected during inference to meet the dev-set latency targets; the reported latency numbers depend on this choice.
  • speech chunk duration = 960 ms (48 wav2vec frames)
    Fixed chunk size used both to build training trajectories and to segment speech at inference; chosen without a reported sweep.
  • encoder sliding window size ws = 10 chunks
    Limits context for the causal chunk-wise encoder; no sensitivity analysis is reported.
  • LoRA rank, alpha, dropout = rank 32, alpha 16, dropout 0.1
    LoRA hyperparameters applied to all linear layers in the second training stage; not swept.
  • beam search decoding parameters = beam 4, repetition penalty 1.2, ngram_no_repeat 5
    Decoding configuration used for all reported results; no ablation is provided.
  • training latency multiplier sampling range = random m up to 12
    Used to merge consecutive chunks during training as data augmentation; the range is chosen by hand.
assumptions (4)
  • domain assumption Pretrained Wav2Vec2.0 and Qwen2.5-7B-Instruct weights transfer to the simultaneous speech translation task.
    The full system is initialized from off-the-shelf checkpoints; the paper does not test alternative encoders or pretraining schemes.
  • domain assumption MFA and SimAlign/LaBSE alignments produce correct word-level time boundaries for trajectory construction.
    Section 3.2 uses these alignments to map each translation word to speech timestamps, but their accuracy is never validated or ablated.
  • domain assumption LLM-translated transcripts from 4-bit Qwen2.5-32B are faithful training targets.
    Section 3.2 uses these synthetic translations as ground truth; translation quality and filtering are not reported.
  • domain assumption The ACL60/60 development set is representative of the IWSLT 2025 evaluation conditions.
    All reported quality and latency conclusions are drawn from this single development set, with no test-set results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CMU's IWSLT 2025 Simultaneous Speech Translation System." pith.science (2026). https://pith.science/paper/OZQQANDM

@misc{pith2026250613143,
  author       = {Pith},
  title        = {Pith review of: CMU's IWSLT 2025 Simultaneous Speech Translation System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZQQANDM}},
  note         = {Machine review of arXiv:2506.13143}
}
read the original abstract

This paper presents CMU's submission to the IWSLT 2025 Simultaneous Speech Translation (SST) task for translating unsegmented English speech into Chinese and German text in a streaming manner. Our end-to-end speech-to-text system integrates a chunkwise causal Wav2Vec 2.0 speech encoder, an adapter, and the Qwen2.5-7B-Instruct as the decoder. We use a two-stage simultaneous training procedure on robust speech segments curated from LibriSpeech, CommonVoice, and VoxPopuli datasets, utilizing standard cross-entropy loss. Our model supports adjustable latency through a configurable latency multiplier. Experimental results demonstrate that our system achieves 44.3 BLEU for English-to-Chinese and 25.1 BLEU for English-to-German translations on the ACL60/60 development set, with computation-aware latencies of 2.7 seconds and 2.3 seconds, and theoretical latencies of 2.2 and 1.7 seconds, respectively.

Figures

Figures reproduced from arXiv: 2506.13143 by the authors.

Figure 1
Figure 1. Model Architecture 2 Task Description The IWSLT 2025 Simultaneous Speech-to-Text Translation track2 focuses on translating unseg￾mented speech into target-language text using pre￾trained large language models (LLMs) and speech encoders. The evaluation data consists of unseg￾mented ACL talks. For English-to-German, sys￾tems are additionally tested on accented speech, while a dedicated development set is provided for … view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice

    cs.CL 2025-07 conditional novelty 6.0 of 10

    An end-to-end simultaneous speech-to-speech translation model with voice cloning, trained with a two-stage reinforcement learning reward scheme, reports high accuracy and low latency on the authors' RealSI benchmark.

Reference graph

Works this paper leans on

31 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    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, and 30 others. 2025. Findings of the iwslt 2025 evaluatio...

    work page 2025

  2. [2]

    Ibrahim Said Ahmad, Antonios Anastasopoulos, Ond r ej Bojar, Claudia Borg, Marine Carpuat, Roldano Cattoni, Mauro Cettolo, William Chen, Qianqian Dong, Marcello Federico, Barry Haddow, D \'a vid Javorsk \'y , Mateusz Krubi \'n ski, Tsz Kin Lam, Xutai Ma, Prashant Mathur, Evgeny Matusov, Chandresh Maurya, John McCrae, and 25 others. 2024. https://doi.org/1...

    doi:10.18653/v1/2024.iwslt-1.1 2024

  3. [3]

    Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. https://aclanthology.org/2020.lrec-1.520/ Common voice: A massively-multilingual speech corpus . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4218--4222, Marse...

    2020

  4. [4]

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020 a . https://arxiv.org/abs/2006.11477 wav2vec 2.0: A framework for self-supervised learning of speech representations

    arXiv 2020

  5. [5]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020 b . https://proceedings.neurips.cc/paper_files/paper/2020/file/ 92d1e1eb1cd6f9fba3227870bb6d7f07-Paper.pdf wav2vec 2.0: A framework for self-supervised learning of speech representations . In Advances in Neural Information Processing Systems, volume 33, pages 12449--12460. Curran Asso...

    work page 2020

  6. [6]

    Puvvada, Jason Li, Subhankar Ghosh, Jagadeesh Balam, and Boris Ginsburg

    Zhehuai Chen, He Huang, Andrei Andrusenko, Oleksii Hrinchuk, Krishna C. Puvvada, Jason Li, Subhankar Ghosh, Jagadeesh Balam, and Boris Ginsburg. 2024. https://doi.org/10.1109/ICASSP48485.2024.10447553 Salm: Speech-augmented language model with in-context learning for speech recognition and translation . In ICASSP 2024 - 2024 IEEE International Conference ...

    arXiv 2024

  7. [7]

    Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. https://doi.org/10.18653/v1/2022.acl-long.62 Language-agnostic BERT sentence embedding . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878--891, Dublin, Ireland. Association for Computational Linguistics

    doi:10.18653/v1/2022.acl-long.62 2022

  8. [8]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. https://arxiv.org/abs/2407.21783 The llama 3...

    arXiv 2024

Show all 31 references
  1. [9]

    Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and Andr \'e F

    Nuno M. Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and Andr \'e F. T. Martins. 2024. https://doi.org/10.1162/tacl_a_00683 xcomet: Transparent machine translation evaluation through fine-grained error detection . Transactions of the Association for Co...

    2024 doi

  2. [10]

    Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. 2024. https://doi.org/10.18653/v1/2024.naacl-long.222 LM -infinite: Zero-shot extreme length generalization for large language models . In Proceedings of the 2024 Conference of the North American C...

    2024 doi

  3. [11]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  4. [12]

    Zhichao Huang, Rong Ye, Tom Ko, Qianqian Dong, Shanbo Cheng, Mingxuan Wang, and Hang Li. 2023. https://arxiv.org/abs/2312.13585 Speech translation with large language models: An industrial practice

    2023 arXiv

  5. [13]

    Javier Iranzo-Sánchez, Joan Albert Silvestre-Cerdà, Javier Jorge, Nahuel Roselló, Adrià Giménez, Albert Sanchis, Jorge Civera, and Alfons Juan. 2020. https://doi.org/10.1109/ICASSP40776.2020.9054626 Europarl-st: A multilingual corpus for speech translation of parliamentary deb...

    2020

  6. [14]

    Masoud Jalili Sabet, Philipp Dufter, Fran c ois Yvon, and Hinrich Sch \"u tze. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.147 S im A lign: High quality word alignments without parallel training data using static and contextualized embeddings . In Findings of the Ass...

    2020 doi

  7. [15]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. http://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015 arXiv

  8. [16]

    Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. 2017. https://doi.org/10.21437/Interspeech.2017-1386 Montreal forced aligner: Trainable text-speech alignment using kaldi . In Interspeech 2017, pages 498--502

    2017 doi

  9. [17]

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

    2025 arXiv

  10. [18]

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. https://doi.org/10.1109/ICASSP.2015.7178964 Librispeech: An asr corpus based on public domain audio books . In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page...

    2015

  11. [19]

    Sara Papi, Marco Gaido, Matteo Negri, and Luisa Bentivogli. 2024 a . 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 Associat...

    2024 doi

  12. [20]

    Sara Papi, Peter Polák, Dominik Macháček, and Ondřej Bojar. 2024 b . https://doi.org/10.1162/tacl_a_00740 How “real” is your real-time simultaneous speech-to-text translation system? Transactions of the Association for Computational Linguistics, 13:281--313

    2024 doi

  13. [21]

    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

  14. [22]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. https://arxiv.org/abs/2412.15115 Qwe...

    2025 arXiv

  15. [23]

    Elizabeth Salesky, Kareem Darwish, Mohamed Al-Badrashiny, Mona Diab, and Jan Niehues. 2023. 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 Interna...

    2023 doi

  16. [24]

    Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. https://doi.org/10.18653/v1/2020.acl-main.704 BLEURT : Learning robust metrics for text generation . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881--7892, Online. Ass...

    2020 doi

  17. [25]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. https://doi.org/10.1016/j.neucom.2023.127063 Roformer: Enhanced transformer with rotary position embedding . Neurocomput., 568(C)

    2024

  18. [26]

    Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021 a . https://doi.org/10.18653/v1/2021.acl-long.80 V ox P opuli: A large-scale multilingual speech corpus for representation learning, semi-...

    2021 doi

  19. [27]

    Changhan Wang, Anne Wu, Jiatao Gu, and Juan Pino. 2021 b . https://doi.org/10.21437/Interspeech.2021-2027 Covost 2 and massively multilingual speech translation . In Interspeech 2021, pages 2247--2251

    2021 doi

  20. [28]

    Xi Xu, Siqi Ouyang, Brian Yan, Patrick Fernandes, William Chen, Lei Li, Graham Neubig, and Shinji Watanabe. 2024. https://doi.org/10.18653/v1/2024.iwslt-1.20 CMU `s IWSLT 2024 simultaneous speech translation system . In Proceedings of the 21st International Conference on Spoke...

    2024 doi

  21. [29]

    Hao Zhang, Nianwen Si, Yaqi Chen, Wenlin Zhang, Xukui Yang, Dan Qu, and Xiaolin Jiao. 2023. https://arxiv.org/abs/2310.02050 Tuning large language model for end-to-end speech translation

    2023 arXiv

  22. [30]

    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...

  23. [31]

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