Pith. sign in

REVIEW 5 major objections 5 minor 5 cited by

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

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

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

Pith's one-line read Seed LiveInterpret 2.0 is an end-to-end speech-to-speech interpreter that clones each speaker's voice, cuts cloned-speech latency from about 10 to about 3 seconds, and exceeds 70 percent correctness in human evaluation.

desk verdict A genuinely interesting RL training scheme is wrapped in an evaluation that cannot yet support the headline accuracy and latency claims; worth refereeing, but revision must demand transparency. read the letter →

arxiv 2507.17527 v3 pith:3QZOCBOJ submitted 2025-07-23 cs.CL cs.SDeess.AS

Shanbo Cheng , Yu Bao , Zhichao Huang , Yu Lu , Ningxin Peng , Lu Xu , Runsheng Yu , Rong Cao
show 20 more authors
Yujiao Du Ting Han Yuxiang Hu Zeyang Li Sitong Liu Shengtao Ma Shiguang Pan Jiongchen Xiao Nuo Xu Meng Yang Rong Ye Yiming Yu Jun Zhang Ruofei Zhang Wanyi Zhang Wenhao Zhu Liehao Zou Lu Lu Yuxuan Wang Yonghui Wu
This is my paper
classification cs.CLcs.SDeess.AS
keywords simultaneousinterpretationspeech-to-speechtranslationvoicecloningreinforcementlearningduplexstreamingarchitecturelatency-qualitytrade-offrewardhackingmultimodallargelanguagemodel
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 a single end-to-end model, Seed LiveInterpret 2.0, can do simultaneous interpretation the way a human interpreter does: listen to each speaker in a live conversation, emit a translation in the other language within a few seconds, and deliver it in the speaker's own cloned voice. If the claim holds, bilingual meetings could run without a human booth interpreter, with each participant hearing the other person's words almost immediately and in that person's voice. The authors report that human interpreters scored the system above 70 percent on long-form, unscripted scenarios, and that cloned-speech latency fell from roughly ten seconds to about three seconds, a near 70 percent reduction. The mechanism is a duplex architecture that reads streaming audio and writes speech tokens continuously, refined by a two-stage reinforcement learning scheme that balances step-by-step translation fidelity against whole-session coherence.

What carries the argument

The load-bearing machinery is the duplex speech-to-speech understanding–generating framework: the model continuously ingests streaming audio chunks and, at each step, decides whether to stay silent or emit a translation segment, a data-driven read-write policy rather than a fixed rule. The argument's engine is the two-stage reinforcement learning scheme. Multi-dimensional single-turn rewards (detection accuracy, translation initiative, translation quality, time compliance, format consistency) give stepwise feedback that keeps each partial output semantically and temporally faithful for intra-segment consistency. Multi-turn rewards (a lagging penalty and a sequence-level alignment score) enforce global coherence and latency across the whole session for inter-segment coherence. Training first optimizes only the single-turn rewards to internalize human priors, then adds the multi-turn reward, under Proximal Policy Optimization with an adaptively tuned KL penalty. A second key object is the Speech Valid Information Proportion (SVIP) metric, an extension of the VIP metric that counts a speech fragment as valid only if it also meets latency, speech-rate, pronunciation, and fluency criteria.

What would settle it

An independent lab could replay the same RealSI audio segments through Seed LiveInterpret 2.0 and the commercial baselines, score the outputs blind with professional interpreters using the published VIP and SVIP rubrics, and time cloned-speech output end to end; if the commercial systems close the reported quality gap or the measured end-to-end latency for cloned speech comes out well above three seconds, the central claim fails.

Watch

Extended reading notes

Core claim

Seed LiveInterpret 2.0 is an end-to-end speech-to-speech simultaneous translation model: a single network, built from a text language model extended with a streaming audio encoder and a speech-token decoder, hears speech in one language and writes text and audio tokens in another while the speaker is still talking. The central discovery is that this architecture, after large-scale multimodal pretraining and carefully staged reinforcement learning, can hold all three requirements of practical interpretation at once: detecting complete semantic units before committing output, delivering each translation within a few seconds of the source, and reproducing the source speaker's voice in the target language. On the RealSI long-form benchmark, human interpreters gave it Valid Information Proportion (VIP) scores of 79.5 (Chinese-to-English) and 70.1 (English-to-Chinese) for speech-to-text, and Speech Valid Information Proportion (SVIP) scores of 67.8 and 64.7 for speech-to-speech, the top scores in the comparison and the only system supporting voice cloning. The paper further shows that the two-stage reward scheme is what keeps these results stable: trained on a time-compliance reward alone, the model collapses into output that is 35 percent shorter with clearly degraded translation quality, while pairing that reward with a quality reward preserves both output length and fidelity.

Load-bearing premise

The central claim rests on the assumption that the RealSI benchmark and the human scoring procedure built by the authors themselves measure real simultaneous-interpretation quality fairly, and that the four commercial systems were tested under comparable conditions.

Editorial extensions

If this is right

  • If the measured scores hold, a machine interpreter can operate in the range of human interpreters on the benchmark's complex long-form scenarios: 79.5 and 70.1 VIP for the two directions, with SVIP scores of 67.8 and 64.7 for speech-to-speech output.
  • Voice cloning is integrated in the same end-to-end pipeline, so each participant can hear the translation in the original speaker's voice, a capability none of the four commercial systems compared in the paper offers.
  • Cutting cloned-speech latency from roughly ten seconds to about three seconds moves the system into the range where live conversation feels natural, which is what makes product-level deployment plausible.
  • The two-stage reward scheme demonstrably prevents a specific failure mode: a time-compliance reward alone drives the model to shorten output by 35 percent and lose translation quality, while the combined reward setup keeps length and fidelity intact.
  • Because the model supports both speech-to-text and speech-to-speech output from one network, a deployed system can serve read-along subtitles and voiced interpretation from the same computation.

Reading between the lines

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

  • A natural extension the paper does not report is applying the same duplex architecture and two-stage reward scheme to language pairs beyond Chinese and English, and to meetings with three or more speakers, where multi-speaker discrimination would face a harder test than the two-speaker scenario illustrated.
  • The reward-hacking result points to a transferable design rule for other latency-constrained generation systems: whenever a verifiable reward correlates with output brevity, it needs an explicit quality reward paired against it, or the policy will silently trade content for speed.
  • Because voice cloning emerges from the same decoder that writes speech tokens, a plausible next step is cloning paralinguistic delivery, such as emphasis, pauses, and emotion, rather than just timbre; the current SVIP rubric does not yet measure such aspects.
  • Independent verification could take the form of a blind re-scoring of the same outputs by professional interpreters outside the developing team, since the RealSI benchmark and the human-evaluation protocol come from the authors' own pipeline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper presents Seed LiveInterpret 2.0, an end-to-end simultaneous speech-to-speech translation system with voice cloning. The authors describe a pipeline based on continual pretraining, supervised fine-tuning, and a two-stage reinforcement learning procedure that combines single-turn stepwise rewards with multi-turn sequence-level rewards. The main empirical claims are that the system achieves high human-evaluated translation quality (VIP 79.5/70.1 and SVIP 67.8/64.7 on the authors' RealSI longform benchmark), substantially outperforms commercial and open-source baselines, reduces cloned-speech latency to about three seconds, and is the only evaluated system supporting voice cloning. The paper also reports sentence-level automatic-metric results, ablation studies on reward design and reward hacking, and latency metrics AL/LAAL/FLAL. The central claims rest on a human evaluation protocol that is not described in sufficient detail and on a benchmark and metrics introduced in the authors' own prior work.

Significance. If the reported results are valid, the contribution is significant: an end-to-end speech-to-speech simultaneous interpreter with voice cloning, low latency, and human-judged quality exceeding commercial systems would be a meaningful advance for practical simultaneous interpretation. The proposed two-stage RL framework with complementary single-turn and multi-turn rewards, and the explicit analysis of reward hacking in Section 4.2, are useful ideas. The paper also provides falsifiable quantitative claims on a public benchmark, although the benchmark itself is the authors' own. However, the significance is conditional on the trustworthiness of the evaluation: the headline VIP/SVIP numbers come from an undocumented human annotation process, and the comparison to commercial baselines is not independently reproducible from the information given. The paper would be considerably strengthened by releasing the evaluation protocol, raw per-fragment scores, and baseline interaction details.

major comments (5)
  1. [Section 3.3, Table 1] The central VIP and SVIP scores are human judgments, but the paper never reports the evaluation protocol: number of scored fragments, number of annotators per fragment, annotator recruitment or language competence, instructions, inter-annotator agreement, or confidence intervals. With binary fragment-level validity judgments, the reported margins (e.g., VIP 79.5 vs 50.1 for zh-en S2T) cannot be distinguished from annotation noise. Please provide these details and report per-fragment scores or confidence intervals for all headline numbers.
  2. [Appendix A] The SVIP metric is introduced in this paper without validation. The rule that a fragment scores zero if any indicator scores 1, one if all indicators score at least 3, and an intermediate case depends on whether 'the overall message remains comprehensible at the sentence level' is underspecified: the paper does not say who makes that final judgment, how the indicators are aggregated, or how annotator disagreements are resolved. Report the complete aggregation procedure and inter-annotator agreement for each indicator and for the final SVIP decision.
  3. [Section 3.3, Table 1] For zh-en speech-to-speech translation, the text states that the method outperforms all other speech-to-speech baselines on FLAL, but Table 1 lists SeamlessStreaming with FLAL 2.38 and Ours with FLAL 2.71. This latency claim is not supported by the table for this direction. Please correct the claim or explain why SeamlessStreaming is excluded from that comparison.
  4. [Sections 3.1 and 3.3] The longform evaluation uses the RealSI benchmark [6], which was introduced in the authors' own prior work, and the VIP and FLAL metrics are also inherited from that work. The paper does not describe how the commercial baselines were accessed, configured, or interacted with (e.g., segmentation, rewriting strategy, API parameters, or latency measurement conditions). Without this information, the claimed large margins over commercial systems cannot be independently verified. Please describe the baseline evaluation setup in detail and, if possible, release the evaluation data and raw annotations.
  5. [Introduction and Figure 1] The claim of 'translation performance approaching human-level accuracy' is not supported by the data reported in the paper, because no human-interpreter VIP or SVIP scores are provided in Table 1 or Figure 1. Response efficiency is defined relative to human interpreter latency, but interpreter-level translation quality is never quantified. Add a human baseline for VIP/SVIP or remove the 'approaching human-level' claim.
minor comments (5)
  1. [Figure 1] The axis label 'Response Efficency' should be 'Response Efficiency', and the header line 'Offical Page' should be 'Official Page'.
  2. [Section 3.3, Table 1] Several entries in Table 1 use inconsistent leading zeros (e.g., '03.27', '03.0') and inconsistent decimal formatting; unify the formatting throughout the table.
  3. [Section 2.2.2] The reward weights w_l, w_s, w_q, w_c, w_f, w_L, and w_Q, as well as the constants c, l, and K_beta, are never specified; report their values or provide a sensitivity analysis so the RL recipe is reproducible.
  4. [Section 3.1] The sentence-level evaluation datasets are not named; even when combining public and proprietary data, the paper should state corpus sizes, domains, and split details so that Table 2 can be interpreted and reproduced.
  5. [Table 6] The VIP scores in Table 6 are reported as rounded percentages (70%, 76%) without denominators or raw counts; state the number of fragments and the valid/invalid counts underlying these percentages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model's accuracy and latency claims are empirical measurements, not derivations that reduce to their inputs; self-authored benchmark and metric citations supply reusable instruments rather than circular premises.

full rationale

The paper's central claims are empirical: Seed LiveInterpret 2.0 is trained with a two-stage RL objective (Section 2.2) and then measured on the RealSI long-form benchmark and sentence-level datasets (Section 3.1). No equation in the paper defines the reported VIP/SVIP scores or latency values in terms of the reward functions, and no parameter is fitted to the test-set numbers. The RL rewards (Eq. 2, rL, rQ) use reference translations and reference timings as training signals; the test metrics VIP, SVIP, AL, FLAL are evaluation instruments, so optimizing similar objectives and then reporting those metrics is standard practice, not circularity. The RealSI benchmark, VIP, and FLAL are cited to the authors' prior work [6], and SVIP is defined in Appendix A; these are self-citations with overlapping authors. However, they supply a reusable benchmark and metric definitions rather than importing the conclusion that the present system performs well. The sentence-level evaluation additionally uses public datasets with external metrics BLEURT and COMET, giving independent grounding. Concerns about the unreported human-evaluation protocol (annotator counts, inter-annotator agreement, confidence intervals) are validity and transparency issues, not circular reductions; they affect whether the empirical claim is convincing, not whether the claim is equivalent to its inputs by construction. Under the hard rule requiring a quoted equation-level reduction, no circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims depend on several unverified premises: fairness of the self-authored RealSI benchmark and SVIP metric, validity of automated reward proxies for partial translation quality, adequacy of inherited Seed TTS and LLM components, and stability of the RL recipe. No model weights, data, or hyperparameters are released, so the ledger cannot be checked externally.

free parameters (4)
  • Reward weights w_l, w_s, w_q, w_c, w_f, w_L, w_Q = not reported
    Balancing coefficients in Eq. (2) and the multi-turn reward rS; chosen by hand or tuned, not reported, and central to the RL objective.
  • Time compliance constant c = not reported
    Appears in rc_t = clip(1 - (1/c) * max(0, Time_y / Time_y* - 1), -1, 1); controls how strongly duration mismatch is penalized.
  • Lagging threshold l = not reported
    Reference threshold in rL := -max(l, (1/K) * sum d_k); defines acceptable waiting time and is not specified.
  • KL target and adaptive KL gain K_beta = not reported
    Used in the adaptive KL penalty update in Section 2.2.3; important for training stability but not reported.
assumptions (5)
  • domain assumption The RealSI benchmark, introduced in the authors' prior paper [6], is a valid and unbiased measure of real-world simultaneous interpretation quality.
    Used for all headline long-form evaluations in Section 3.3; no external benchmark confirms the result.
  • domain assumption The single-turn reward components, especially Trans(y_t, y*_t), correlate strongly with human evaluation of partial simultaneous translations.
    Asserted in Section 2.2.2 without supporting evidence; the RL objective depends on this correlation.
  • domain assumption The pre-trained Seed LLM family and pre-trained audio encoder provide a suitable base for multimodal streaming translation.
    The paper inherits these components from prior work [2,4,36] and does not describe their internal design.
  • domain assumption Human evaluation under the SVIP rules yields stable and comparable scores across systems.
    Appendix A defines indicators, but annotator selection, instructions, and inter-annotator agreement are not reported.
  • standard math PPO with GAE and the adaptive KL penalty stabilizes training of this mixed audio-text policy.
    Adopted from Schulman et al. and Ziegler et al.; not proven here but standard in the RL literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice." pith.science (2026). https://pith.science/paper/3QZOCBOJ

@misc{pith2026250717527,
  author       = {Pith},
  title        = {Pith review of: Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3QZOCBOJ}},
  note         = {Machine review of arXiv:2507.17527}
}
read the original abstract

Simultaneous Interpretation (SI) represents one of the most daunting frontiers in the translation industry, with product-level automatic systems long plagued by intractable challenges: subpar transcription and translation quality, lack of real-time speech generation, multi-speaker confusion, and translated speech inflation, especially in long-form discourses. In this study, we introduce Seed-LiveInterpret 2.0, an end-to-end SI model that delivers high-fidelity, ultra-low-latency speech-to-speech generation with voice cloning capabilities. As a fully operational product-level solution, Seed-LiveInterpret 2.0 tackles these challenges head-on through our novel duplex speech-to-speech understanding-generating framework. Experimental results demonstrate that through large-scale pretraining and reinforcement learning, the model achieves a significantly better balance between translation accuracy and latency, validated by human interpreters to exceed 70% correctness in complex scenarios. Notably, Seed-LiveInterpret 2.0 outperforms commercial SI solutions by significant margins in translation quality, while slashing the average latency of cloned speech from nearly 10 seconds to a near-real-time 3 seconds, which is around a near 70% reduction that drastically enhances practical usability.

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. ProactiveLLM: Learning Active Interaction for Streaming Large Language Models

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    ProactiveLLM enables active interaction in streaming LLMs by learning semantic sufficiency cues from partial inputs through mask-based modeling and synchronized privileged self-distillation without external supervision.

  2. X-Translator: A Real-Time Multilingual Speaker-Aware Speech-to-Speech Translation System

    eess.AS 2026-07 conditional novelty 5.0 of 10

    An open, modular cascaded system (streaming ASR + MT + prompt-conditioned TTS) preserves speaker identity in long-form multi-speaker translation, at higher latency and slightly lower translation quality than proprietary APIs.

  3. Regularized Entropy Information Adaptation with Temporal-Awareness Networks for Simultaneous Speech Translation

    cs.LG 2026-04 unverdicted novelty 5.0 of 10

    REINA-SAN and REINA-TAN add temporal context to information-based read/write policies, improving the quality-latency tradeoff in simultaneous speech translation by up to 7.1% on Normalized Streaming Efficiency.

  4. Afrispeech Semantics: Evaluating Audio Semantic Reasoning in Spoken Language Models Across Domains and Accents

    cs.CL 2026-05 unverdicted novelty 4.0 of 10

    Audio language models are benchmarked on five semantic and paralinguistic reasoning tasks to reveal limitations in handling spoken audio evidence, accent variation, and domain shifts.

  5. MLLP-VRAIN UPV system for the IWSLT 2026 Simultaneous Speech Translation task

    cs.CL 2026-06 unverdicted novelty 3.0 of 10

    A cascaded SimulST system using Parakeet and Qwen 3.5 with adaptive black-box policies and RAG context achieves +5.82 XCOMET-XL improvement on En→De for IWSLT 2026.

Reference graph

Works this paper leans on

52 extracted references · 21 canonical work pages · cited by 5 Pith papers

  1. [6]

    Towards achieving human parity on end-to-end simultaneous speech translation via llm agent, 2024

    Shanbo Cheng, Zhichao Huang, Tom Ko, Hang Li, Ningxin Peng, Lu Xu, and Qini Zhang. Towards achieving human parity on end-to-end simultaneous speech translation via llm agent, 2024. URLhttps://arxiv.org/abs/ 2407.21646

    arXiv 2024

  2. [1]

    Tower: An open multilingual large language model for translation-related tasks

    Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint arXiv:2402.17733 , 2024

    arXiv 2024

  3. [2]

    Seed-tts: A family of high-quality versatile speech generation models.arXiv preprint arXiv:2406.02430, 2024

    Philip Anastassiou, Jiawei Chen, Jitong Chen, Yuanzhe Chen, Zhuo Chen, Ziyi Chen, Jian Cong, Lelai Deng, Chuang Ding, Lu Gao, et al. Seed-tts: A family of high-quality versatile speech generation models.arXiv preprint arXiv:2406.02430, 2024

    arXiv 2024

  4. [3]

    Monotonic Infinite Lookback Attention for Simultaneous Machine Translation

    Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. Monotonic infinite lookback attention for simultaneous machine translation.arXiv preprint arXiv:1906.05218, 2019

    work page Pith review arXiv 1906

  5. [4]

    Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition

    Ye Bai, Jingping Chen, Jitong Chen, Wei Chen, Zhuo Chen, Chuang Ding, Linhao Dong, Qianqian Dong, Yujiao Du, Kepan Gao, et al. Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition. arXiv preprint arXiv:2407.04675 , 2024

    arXiv 2024

  6. [5]

    Seamless: Multilingual expressive and streaming speech translation

    Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, et al. Seamless: Multilingual expressive and streaming speech translation. arXiv preprint arXiv:2312.05187 , 2023

    arXiv 2023

  7. [7]

    Can neural machine translation do simultaneous translation?arXiv preprint arXiv:1606.02012, 2016

    Kyunghyun Cho and Masha Esipova. Can neural machine translation do simultaneous translation?arXiv preprint arXiv:1606.02012, 2016

    arXiv 2016

  8. [8]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models.arXiv preprint arXiv:2311.07919, 2023

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models.arXiv preprint arXiv:2311.07919, 2023

    arXiv 2023

Show all 52 references
  1. [9]

    Efficient wait-k models for simultaneous machine translation

    Maha Elbayad, Laurent Besacier, and Jakob Verbeek. Efficient wait-k models for simultaneous machine translation. arXiv preprint arXiv:2005.08595 , 2020

    2005 arXiv

  2. [10]

    Mt-r1-zero: Advancing llm-based machine translation via r1-zero-like reinforcement learning

    Zhaopeng Feng, Shaosheng Cao, Jiahan Ren, Jiayuan Su, Ruizhe Chen, Yan Zhang, Zhe Xu, Yao Hu, Jian Wu, and Zuozhu Liu. Mt-r1-zero: Advancing llm-based machine translation via r1-zero-like reinforcement learning. arXiv preprint arXiv:2504.10160 , 2025

    2025 arXiv

  3. [11]

    Naist simultaneous speech-to-speech translation system for iwslt

    Ryo Fukuda, Yuta Nishikawa, Yasumasa Kano, Yuka Ko, Tomoya Yanagita, Kosuke Doi, Mana Makinae, Sakriani Sakti, Katsuhito Sudoh, and Satoshi Nakamura. Naist simultaneous speech-to-speech translation system for iwslt

  4. [12]

    Speech translation with speech foundation models and large language models: What is there and what is missing?arXiv preprint arXiv:2402.12025 , 2024

    Marco Gaido, Sara Papi, Matteo Negri, and Luisa Bentivogli. Speech translation with speech foundation models and large language models: What is there and what is missing?arXiv preprint arXiv:2402.12025 , 2024

    2024 arXiv

  5. [13]

    Learning to translate in real-time with neural machine translation

    Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor OK Li. Learning to translate in real-time with neural machine translation. arXiv preprint arXiv:1610.00388 , 2016

    2016 arXiv

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [15]

    R1-t1: Fully incentivizing translation capability in llms via reasoning learning.arXiv preprint arXiv:2502.19735, 2025

    Minggui He, Yilun Liu, Shimin Tao, Yuanchang Luo, Hongyong Zeng, Chang Su, Li Zhang, Hongxia Ma, Daimeng Wei, Weibin Meng, et al. R1-t1: Fully incentivizing translation capability in llms via reasoning learning.arXiv preprint arXiv:2502.19735, 2025

    2025 arXiv

  8. [16]

    Improving machine translation with human feedback: An exploration of quality estimation as a reward model.arXiv preprint arXiv:2401.12873, 2024

    Zhiwei He, Xing Wang, Wenxiang Jiao, Zhuosheng Zhang, Rui Wang, Shuming Shi, and Zhaopeng Tu. Improving machine translation with human feedback: An exploration of quality estimation as a reward model.arXiv preprint arXiv:2401.12873, 2024

    2024 arXiv

  9. [17]

    Speech translation with large language models: An industrial practice.arXiv preprint arXiv:2312.13585 , 2023

    Zhichao Huang, Rong Ye, Tom Ko, Qianqian Dong, Shanbo Cheng, Mingxuan Wang, and Hang Li. Speech translation with large language models: An industrial practice.arXiv preprint arXiv:2312.13585 , 2023. 13

    2023 arXiv

  10. [18]

    Direct speech-to-speech translation with a sequence-to-sequence model.arXiv preprint arXiv:1904.06037 , 2019

    Ye Jia, Ron J Weiss, Fadi Biadsy, Wolfgang Macherey, Melvin Johnson, Zhifeng Chen, and Yonghui Wu. Direct speech-to-speech translation with a sequence-to-sequence model.arXiv preprint arXiv:1904.06037 , 2019

    1904 arXiv

  11. [19]

    Conference interpreting explained

    Roderick Jones. Conference interpreting explained. Routledge, 2014

    2014

  12. [20]

    T\" ulu 3: Pushing frontiers in open language model post-training

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\" ulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124 , 2024

    2024 arXiv

  13. [21]

    Jiahuan Li, Hao Zhou, Shujian Huang, Shanbo Cheng, and Jiajun Chen. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions.Transactions of the Association for Computational Linguistics, 12:576–592, 2024

    2024

  14. [22]

    Cross attention augmented transducer networks for simultaneous translation

    Dan Liu, Mengge Du, Xiaoxi Li, Ya Li, and Enhong Chen. Cross attention augmented transducer networks for simultaneous translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 39–55, 2021

    2021

  15. [23]

    Llamax: Scaling linguistic horizons of llm by enhancing translation capabilities beyond 100 languages.arXiv preprint arXiv:2407.05975 , 2024

    Yinquan Lu, Wenhao Zhu, Lei Li, Yu Qiao, and Fei Yuan. Llamax: Scaling linguistic horizons of llm by enhancing translation capabilities beyond 100 languages.arXiv preprint arXiv:2407.05975 , 2024

    2024 arXiv

  16. [24]

    Reft: Reasoning with reinforced fine-tuning

    Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967 , 3, 2024

    2024 arXiv

  17. [25]

    Stacl: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework

    Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, et al. Stacl: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework. arXiv preprint arXiv:181...

    arXiv

  18. [26]

    Mt metrics correlate with human ratings of simultaneous speech translation

    Dominik Macháček, Ondřej Bojar, and Raj Dabre. Mt metrics correlate with human ratings of simultaneous speech translation. arXiv preprint arXiv:2211.08633 , 2022

    2022 arXiv

  19. [27]

    Reinforcementfine-tuning

    OpenAI. Reinforcementfine-tuning. https://platform.openai.com/docs/guides/reinforcement-fine-tuning, 2024

    2024

  20. [28]

    Cmu’s iwslt 2025 simultaneous speech translation system.arXiv preprint arXiv:2506.13143, 2025

    Siqi Ouyang, Xi Xu, and Lei Li. Cmu’s iwslt 2025 simultaneous speech translation system.arXiv preprint arXiv:2506.13143, 2025

    2025 arXiv

  21. [29]

    G-dig: Towards gradient- based diverse and high-quality instruction data selection for machine translation.arXiv preprint arXiv:2405.12915, 2024

    Xingyuan Pan, Luyang Huang, Liyan Kang, Zhicheng Liu, Yu Lu, and Shanbo Cheng. G-dig: Towards gradient- based diverse and high-quality instruction data selection for machine translation.arXiv preprint arXiv:2405.12915, 2024

    2024 arXiv

  22. [30]

    Over-generation cannot be rewarded: Length-adaptive average lagging for simultaneous speech translation.arXiv preprint arXiv:2206.05807 , 2022

    Sara Papi, Marco Gaido, Matteo Negri, and Marco Turchi. Over-generation cannot be rewarded: Length-adaptive average lagging for simultaneous speech translation.arXiv preprint arXiv:2206.05807 , 2022

    2022 arXiv

  23. [31]

    How “real” is your real-time simultaneous speech-to-text translation system? Transactions of the Association for Computational Linguistics , 13:281–313, 2025

    Sara Papi, Peter Polak, Dominik Macháček, and Ondřej Bojar. How “real” is your real-time simultaneous speech-to-text translation system? Transactions of the Association for Computational Linguistics , 13:281–313, 2025

    2025

  24. [32]

    Comet: A neural framework for mt evaluation

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. Comet: A neural framework for mt evaluation. arXiv preprint arXiv:2009.09025 , 2020

    2009 arXiv

  25. [33]

    Simulspeech: End-to-end simultaneous speech to text translation

    Yi Ren, Jinglin Liu, Xu Tan, Chen Zhang, Tao Qin, Zhou Zhao, and Tie-Yan Liu. Simulspeech: End-to-end simultaneous speech to text translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3787–3796, 2020

    2020

  26. [34]

    High-dimensional continuous control using generalized advantage estimation.arXiv preprint arXiv:1506.02438 , 2015

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation.arXiv preprint arXiv:1506.02438 , 2015

    2015 arXiv

  27. [35]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. URLhttps://api.semanticscholar.org/CorpusID:28695052

    2017 arXiv

  28. [36]

    ByteDance Seed, Jiaze Chen, Tiantian Fan, Xin Liu, Lingjun Liu, Zhiqi Lin, Mingxuan Wang, Chengyi Wang, Xiangpeng Wei, Wenyuan Xu, et al. Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning. arXiv preprint arXiv:2504.13914 , 2025

    2025

  29. [37]

    Bleurt: Learning robust metrics for text generation.arXiv preprint arXiv:2004.04696, 2020

    Thibault Sellam, Dipanjan Das, and Ankur P Parikh. Bleurt: Learning robust metrics for text generation.arXiv preprint arXiv:2004.04696, 2020. 14

    2004 arXiv

  30. [38]

    Multi-turn reinforcement learning with preference human feedback.Advances in Neural Information Processing Systems , 37:118953–118993, 2024

    Lior Shani, Aviv Rosenberg, Asaf Cassel, Oran Lang, Daniele Calandriello, Avital Zipori, Hila Noga, Orgad Keller, Bilal Piot, Idan Szpektor, et al. Multi-turn reinforcement learning with preference human feedback.Advances in Neural Information Processing Systems , 37:118953–11...

    2024

  31. [39]

    Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289 , 2023

    Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289 , 2023

    2023 arXiv

  32. [40]

    NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prang...

    2022 arXiv

  33. [41]

    Attention is all you need.Advances in neural information processing systems , 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems , 30, 2017

    2017

  34. [42]

    Barriers to effective evaluation of simultaneous interpretation

    Shira Wein, I Te, Colin Cherry, Juraj Juraska, Dirk Padfield, and Wolfgang Macherey. Barriers to effective evaluation of simultaneous interpretation. InFindings of the Association for Computational Linguistics: EACL 2024, pages 209–219, 2024

    2024

  35. [43]

    Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144 , 2016

    Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144 , 2016

    2016 arXiv

  36. [44]

    Advancing translation preference modeling with rlhf: A step towards cost-effective solution.arXiv preprint arXiv:2402.11525, 2024

    Nuo Xu, Jun Zhao, Can Zu, Sixian Li, Lu Chen, Zhihao Zhang, Rui Zheng, Shihan Dou, Wenjuan Qin, Tao Gui, et al. Advancing translation preference modeling with rlhf: A step towards cost-effective solution.arXiv preprint arXiv:2402.11525, 2024

    2024 arXiv

  37. [45]

    Seqpo-simt: Sequential policy optimization for simultaneous machine translation.arXiv preprint arXiv:2505.20622 , 2025

    Ting Xu, Zhichao Huang, Jiankai Sun, Shanbo Cheng, and Wai Lam. Seqpo-simt: Sequential policy optimization for simultaneous machine translation.arXiv preprint arXiv:2505.20622 , 2025

    2025 arXiv

  38. [46]

    Simulpl: Aligning human preferences in simultaneous machine translation.arXiv preprint arXiv:2502.00634 , 2025

    Donglei Yu, Yang Zhao, Jie Zhu, Yangyifan Xu, Yu Zhou, and Chengqing Zong. Simulpl: Aligning human preferences in simultaneous machine translation.arXiv preprint arXiv:2502.00634 , 2025

    2025 arXiv

  39. [47]

    Realtrans: End-to-end simultaneous speech translation with convolutional weighted-shrinking transformer.arXiv preprint arXiv:2106.04833 , 2021

    Xingshan Zeng, Liangyou Li, and Qun Liu. Realtrans: End-to-end simultaneous speech translation with convolutional weighted-shrinking transformer.arXiv preprint arXiv:2106.04833 , 2021

    2021 arXiv

  40. [48]

    Hidden markov transformer for simultaneous machine translation.arXiv preprint arXiv:2303.00257, 2023

    Shaolei Zhang and Yang Feng. Hidden markov transformer for simultaneous machine translation.arXiv preprint arXiv:2303.00257, 2023

    2023 arXiv

  41. [49]

    Streamspeech: Simultaneous speech-to-speech translation with multi-task learning.arXiv preprint arXiv:2406.03049 , 2024

    Shaolei Zhang, Qingkai Fang, Shoutao Guo, Zhengrui Ma, Min Zhang, and Yang Feng. Streamspeech: Simultaneous speech-to-speech translation with multi-task learning.arXiv preprint arXiv:2406.03049 , 2024

    2024 arXiv

  42. [50]

    Simpler and faster learning of adaptive policies for simultaneous translation

    Baigong Zheng, Renjie Zheng, Mingbo Ma, and Liang Huang. Simpler and faster learning of adaptive policies for simultaneous translation. arXiv preprint arXiv:1909.01559 , 2019

    1909 arXiv

  43. [51]

    Fine-tuning language models from human preferences.arXiv preprint arXiv:1909.08593 , 2019

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences.arXiv preprint arXiv:1909.08593 , 2019. 15 Appendix A Guidelines of SVIP The SVIP assessment uses a t...

    1909 arXiv

  44. [2023]

    In Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023) , pages 330–340, 2023

    2023

Pith tools

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