pith. sign in

arxiv: 2606.03948 · v1 · pith:TNF2VMZ6new · submitted 2026-06-02 · 💻 cs.CL

A Pocket Offline Model for Simultaneous Speech Translation as CUNI Submission to IWSLT 2026

Aziz Sharipov Ortega , Dominik Mach\'a\v{c}ek This is my paper

Pith reviewed 2026-06-28 09:47 UTC · model grok-4.3

classification 💻 cs.CL
keywords simultaneous speech translationAlignAttCanary modelIWSLT 2026multilinguallow latencyspeech to textCzech English translation
0
0 comments X

The pith

A 1B-parameter offline model outperforms baselines in simultaneous speech translation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a system for simultaneous speech translation by applying the AlignAtt policy to the existing offline Canary model without major retraining. This system is submitted to the IWSLT 2026 shared task covering Czech to English and English to German and Italian. The authors report that the approach achieves high translation quality, outperforming similarly sized baselines in both low- and high-latency regimes in simulations, while using only 1B parameters and supporting 25 languages. A sympathetic reader would care if this shows that offline models can be repurposed for real-time translation tasks efficiently.

Core claim

The paper claims that the AlignAtt policy can be used with the Canary offline speech-to-text model to implement simultaneous translation, resulting in a system that has high translation quality, low computational requirements with 1B parameters, and multilinguality with 25 source and 25 target languages, as demonstrated in submissions to IWSLT 2026 for specific language pairs.

What carries the argument

AlignAtt policy applied to the Canary model for simultaneous speech translation.

If this is right

  • Outperforms similarly sized baselines in low- and high-latency regimes in simulations.
  • Requires only 1B parameters for low computational requirements.
  • Supports 25 source and 25 target languages.
  • Submitted for Czech-English, English-German, and English-Italian translation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This could allow deployment on devices with limited resources for on-the-fly translation.
  • The approach might generalize to other offline models if the policy integration is key.
  • Real-time user studies could validate the simulation-based quality claims.

Load-bearing premise

The AlignAtt policy integrates effectively with the Canary model to deliver simultaneous translation performance without requiring major retraining or suffering quality loss.

What would settle it

Running the system in actual simultaneous conditions and measuring if quality remains higher than baselines without degradation.

Figures

Figures reproduced from arXiv: 2606.03948 by Aziz Sharipov Ortega, Dominik Mach\'a\v{c}ek.

Figure 1
Figure 1. Figure 1: English-to-Italian dev chrF vs. LongYAAL [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
read the original abstract

We implement simultaneous translation capability with the offline direct speech-to-text translation model Canary, using the state-of-the-art policy AlignAtt, and submit it to IWSLT 2026 Simultaneous Speech Translation Shared task for Czech to English and English to German and Italian. The strengths of our system are: (1) high translation quality, outperforming similarly sized baselines both in low- and high-latency regimes in computationally unaware simulations; (2) low computational requirements, as the model has only 1B parameters; (3) multilinguality -- support of 25 source and 25 target languages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper describes the CUNI submission to the IWSLT 2026 Simultaneous Speech Translation Shared Task. It implements simultaneous translation capability for Czech-to-English, English-to-German, and English-to-Italian using the offline 1B-parameter Canary direct speech-to-text model combined with the AlignAtt policy. The system is presented as achieving high translation quality while outperforming similarly sized baselines in low- and high-latency regimes under computationally unaware simulations, with additional strengths in low computational requirements and support for 25 source and 25 target languages.

Significance. If the performance claims are substantiated with experimental evidence, the work would demonstrate a practical approach to converting an existing offline multilingual speech translation model into a simultaneous system using an established policy, without apparent need for major retraining. This could be significant for efficient, low-parameter multilingual simultaneous translation deployments, particularly where computational resources are constrained.

major comments (2)
  1. [Abstract] Abstract: The claim that the system achieves 'high translation quality, outperforming similarly sized baselines both in low- and high-latency regimes in computationally unaware simulations' is presented without any reported metrics (e.g., BLEU, latency values), baseline systems, simulation details, or results tables. This renders the central empirical claim unverifiable from the manuscript.
  2. [Abstract] Abstract / System Description: No details are provided on the integration of AlignAtt with the offline-trained Canary model, including whether any adaptation or fine-tuning was required, how the policy interacts with the multilingual 25x25 capability, or whether additional decoding steps affect output quality. This leaves the assumption that the combination preserves original quality unexamined.
minor comments (1)
  1. The manuscript would benefit from including a results section or table with quantitative comparisons to baselines to support the stated strengths.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our IWSLT 2026 submission paper. We agree that the abstract requires more concrete empirical grounding and that additional clarification on the AlignAtt integration is warranted. We will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that the system achieves 'high translation quality, outperforming similarly sized baselines both in low- and high-latency regimes in computationally unaware simulations' is presented without any reported metrics (e.g., BLEU, latency values), baseline systems, simulation details, or results tables. This renders the central empirical claim unverifiable from the manuscript.

    Authors: We agree that the abstract should not stand alone without supporting numbers. The full paper contains results tables (Section 4) with BLEU and latency metrics for the three language pairs, along with comparisons to similarly sized baselines under the IWSLT simulation protocol. To address the concern, we will revise the abstract to include the key quantitative results (e.g., average BLEU and latency ranges) and explicitly reference the evaluation setup and baselines. revision: yes

  2. Referee: [Abstract] Abstract / System Description: No details are provided on the integration of AlignAtt with the offline-trained Canary model, including whether any adaptation or fine-tuning was required, how the policy interacts with the multilingual 25x25 capability, or whether additional decoding steps affect output quality. This leaves the assumption that the combination preserves original quality unexamined.

    Authors: The current manuscript briefly states that AlignAtt is applied to the offline Canary model without retraining. We will expand the system description section to clarify: (1) no fine-tuning or adaptation of Canary parameters was performed; AlignAtt is used as a plug-in policy that reads attention alignments from the existing model; (2) the 25x25 multilingual support is preserved because AlignAtt operates on cross-attention weights independently of language pair; (3) the additional decoding logic introduces negligible quality degradation, as confirmed by our offline-to-simultaneous quality comparison experiments. These details will be added in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical system description with independent experimental claims

full rationale

The paper is a system submission description for a shared task. It reports implementing simultaneous ST by combining an existing 1B-parameter offline Canary model with the AlignAtt policy, then evaluates quality and latency in simulations. No equations, derivations, fitted parameters, or predictions appear. Central claims rest on experimental results against baselines, not on any reduction to self-defined quantities or self-citation chains. Self-citations, if present, are not load-bearing for the reported performance numbers. This matches the default case of a self-contained empirical paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an applied system paper with no mathematical content, free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5633 in / 869 out tokens · 24806 ms · 2026-06-28T09:47:22.049553+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    InProceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 412–481, Vienna, Austria (in-person and online)

    Findings of the IWSLT 2025 evaluation campaign. InProceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 412–481, Vienna, Austria (in-person and online). David Ifeoluwa Adelani, Victor Agostinelli, Antonios Anastasopoulos, Luisa Bentivogli, Ondˇrej Bojar, Se- bastien Bratières, Marine Carpuat, Roldano Cattoni, ...

    2025

  2. [2]

    InProceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026), San Diego, California, US

    Speech translation and metrics in 2026: Findings of the iwslt campaign. InProceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026), San Diego, California, US. Association for Computational Linguistics. Antonios Anastasopoulos and 1 others

    2026

  3. [3]

    InProceedings of the 23rd In- ternational Conference on Spoken Language Trans- lation (IWSLT 2026)

    Speech translation and metrics in 2026: Findings of the IWSLT campaign. InProceedings of the 23rd In- ternational Conference on Spoken Language Trans- lation (IWSLT 2026). Association for Computational Linguistics. 5 Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruom- ing Pang, Wei Li, and Colin Raffel

    2026

  4. [4]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

    Seam- lessm4t: Massively multilingual & multimodal ma- chine translation.Preprint, arXiv:2308.11596. Marco Gaido, Sara Papi, Mauro Cettolo, Matteo Ne- gri, and Luisa Bentivogli

    work page arXiv

  5. [5]

    Preprint, arXiv:2512.17648

    Simulstream: Open-source toolkit for evaluation and demonstra- tion of streaming speech-to-text translation systems. Preprint, arXiv:2512.17648. Nuno M. Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and André F. T. Mar- tins

    work page arXiv

  6. [6]

    InText, Speech, and Dialogue: 24th International Conference, TSD 2021, Olomouc, Czech Republic, September 6–9, 2021, Proceedings, page 293–304, Berlin, Heidelberg

    Parczech 3.0: A large czech speech corpus with rich metadata. InText, Speech, and Dialogue: 24th International Conference, TSD 2021, Olomouc, Czech Republic, September 6–9, 2021, Proceedings, page 293–304, Berlin, Heidelberg. Springer-Verlag. Roman Koshkin, Je Haesung, Lianbo Liu, Hao Shi, Meng Zhao, Yusuke Fujita, and Yui Sudo

    2021

  7. [7]

    Dominik Macháˇcek and Peter Polák

    High-fidelity simultaneous speech-to- speech translation.Preprint, arXiv:2502.03382. Dominik Macháˇcek and Peter Polák

    work page arXiv

  8. [8]

    Jan Niehues and 1 others

    InProceed- ings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 389–398, Vienna, Austria (in-person and online). Jan Niehues and 1 others

    2025

  9. [9]

    Sara Papi, Peter Polák, Dominik Macháˇcek, and Ondˇrej Bojar

    Does simultaneous speech translation need simultaneous models? InFindings of the Asso- ciation for Computational Linguistics: EMNLP 2022, pages 141–153, Abu Dhabi, United Arab Emirates. Sara Papi, Peter Polák, Dominik Macháˇcek, and Ondˇrej Bojar. 2025a. How “real” is your real-time simulta- neous speech-to-text translation system?Transac- tions of the As...

    2022

  10. [10]

    Alignatt: Using attention-based audio-translation alignments as a guide for simultaneous speech translation.arXiv preprint arXiv:2305.11408,

    Alig- natt: Using attention-based audio-translation align- ments as a guide for simultaneous speech translation. Preprint, arXiv:2305.11408. Sara Papi, Maike Züfle, Marco Gaido, Beatrice Savoldi, Danni Liu, Ioannis Douros, Luisa Bentivogli, and Jan Niehues. 2025b. Mcif: Multimodal crosslingual instruction-following benchmark from scientific talks. Preprin...

    work page arXiv

  11. [11]

    Maja Popovi´c

    Better late than never: Meta-evaluation of latency metrics for simultaneous speech-to-text translation.Preprint, arXiv:2509.17349. Maja Popovi´c

    work page arXiv

  12. [12]

    Canary-1b-v2 & parakeet-tdt-0.6 b-v3: Efficient and high-performance models for multilingual asr and ast.arXiv preprint arXiv:2509.14128,

    Canary-1b-v2 & parakeet-tdt-0.6b-v3: Efficient and 6 high-performance models for multilingual asr and ast. Preprint, arXiv:2509.14128. Sukanta Sen, Ond ˇrej Bojar, and Barry Haddow

    work page arXiv

  13. [13]

    Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, Jin Xu, Jingren Zhou, and Junyang Lin

    Simultaneous translation for unsegmented input: A sliding window approach.Preprint, arXiv:2210.09754. Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, Jin Xu, Jingren Zhou, and Junyang Lin

    work page arXiv

  14. [14]

    Qwen3-ASR Technical Report

    Qwen3-asr technical report. Preprint, arXiv:2601.21337. Silero Team

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Qwen3 technical report.Preprint, arXiv:2505.09388. 7

    work page internal anchor Pith review Pith/arXiv arXiv