Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Efficient Speech Translation through Model Compression and Knowledge Distillation

T0 review · 3 major / 4 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 The paper reports that Qwen2-Audio-7B-Instruct can be cut to half its parameters and storage while keeping 97-100% of its speech-translation quality.

desk verdict Solid IWSLT system paper with a misleading headline: the 50% parameter reduction doesn't survive contact with quantization, and the quality numbers are selected on the test set. read the letter →

arxiv 2505.20237 v2 pith:57R4GGOR submitted 2025-05-26 cs.CL cs.SDeess.AS

Yasmin Moslem This is my paper
classification cs.CLcs.SDeess.AS
keywords speechtranslationmodelcompressionlayerpruningknowledgedistillationQLoRAquantizationQwen2-Audio
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 claims that a 7-billion-parameter speech-translation model can be reduced to about half its size in both parameter count and storage while keeping 97-100% of the full model's translation quality on English-to-German and English-to-Chinese speech translation. The route to that result combines three techniques: iteratively pruning the decoder layers whose removal least hurts translation quality, fine-tuning the pruned model on in-domain data augmented with teacher-generated knowledge-distillation translations, and then quantizing to 4-bit precision with low-rank adapters, a combination the paper calls QLoRA. The strongest setup cut the model from 8.40B to 4.12B parameters and from roughly 16.8GB to 8.65GB of storage, with compressed models scoring at or near the teacher level on BLEU, chrF/chrF++, and COMET. If the claim holds, large audio-language models do not have to be deployed at full size for practical speech translation.

What carries the argument

The load-bearing mechanism is greedy, metric-guided layer pruning applied to the decoder of an encoder-decoder transformer. Starting from the fully fine-tuned teacher, the paper removes one decoder layer at a time, choosing at each step the layer whose removal produces the smallest drop in chrF/chrF++ on the test split, and repeats until 8 of the 32 decoder layers are gone. The pruned skeleton is then restored by fine-tuning on the in-domain ACL 60/60 data with sequence-level knowledge-distillation translations from the teacher, followed by QLoRA: 4-bit quantization plus low-rank adapters, with rank 64 and alpha 128, trained on a mixture of in-domain data, distillation data, and out-of-domain CoVoST2 segments. Layer selection, not mere depth reduction, is what carries the argument: the chosen layers differ between languages and are scattered through the network, whereas removing a fixed middle block or pruning encoder and decoder layers together performs much worse.

What would settle it

Hold out a fresh evaluation set that was never used during layer selection or training, run the identical pruning-and-QLoRA pipeline, and compare the student BLEU, chrF/chrF++, and COMET against the teacher; if either language pair falls clearly below the reported 97-100% retention, the near-lossless half-size claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that compression and quality are not in opposition for this model class: a deliberately pruned, quantized, and distillation-trained student can be half the teacher's size and still translate at essentially the teacher's level. The paper's strongest evidence is its second setup, where 8 of the 32 decoder layers are removed one at a time, each step removing the layer whose absence costs least on chrF/chrF++ scores. After re-fine-tuning and QLoRA with knowledge-distillation data plus out-of-domain segments, the final German model scores BLEU 39.59, chrF++ 65.14, and COMET 59.21 against the teacher's 39.28, 65.27, and 56.32, while the Chinese model scores BLEU 56.52, chrF 50.74, and COMET 64.34 against the teacher's 58.54, 52.54, and 65.97. The author reads these numbers as 100% retention for German and 97% for Chinese, with ablations pointing to the design choices behind the result: iterative pruning beats fixed middle-layer pruning, decoder-only pruning beats pruning the encoder too, and chrF/chrF++ is a better pruning-selection metric than COMET.

Load-bearing premise

The reported quality-retention numbers assume the 100-utterance test set is an independent, untouched measure of quality, but the same test set was used to decide which model layers to remove, so the final scores may be inflated by selection on that test set.

Editorial extensions

If this is right

  • A speech-translation model based on Qwen2-Audio can be deployed at roughly half the parameter count and storage footprint without a meaningful drop in translation quality, under the paper's evaluation setup.
  • Pruning only the decoder, rather than the encoder and decoder together, is the better compression path for this model when translation quality is the target.
  • Iterative, importance-guided pruning is materially better than removing a fixed block of middle layers; after fine-tuning on English-to-German, iterative selection reached BLEU 32.16 against 30.81 for middle-layer pruning.
  • Pruning up to 12 decoder layers degrades little after fine-tuning, but pruning 16 layers causes a sharp quality drop, setting an empirical limit for this model.
  • Pruning reduces storage and accelerates inference by roughly 20-40% depending on depth, while QLoRA quantization reduces storage further but slows inference, so the best choice depends on whether storage or latency is the binding constraint.

Reading between the lines

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

  • Inference: because the same 100-utterance test split is used both to choose which layers to prune and to report final quality, the 97-100% retention figures are likely optimistic; a separate validation split for pruning decisions would give a fairer estimate.
  • Inference: the '100% retention' for German is metric-dependent: COMET actually improves over the teacher while BLEU and chrF++ are slightly lower, so a single retention percentage should be read as an average across metrics rather than uniform across all of them.
  • Inference: a useful follow-up is to run the same pipeline with a never-consulted test set and check whether chrF-guided pruning still wins over middle-layer pruning; if it does, the method is robust, and if not, the published gap is partly a test-selection artifact.
  • Inference: because the method only needs a teacher model, a pruning-selection metric, and distillation data, the same decoder-layer-pruning pipeline could transfer to other encoder-decoder audio-language models without task-specific machinery.
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 / 4 minor

Summary. This paper reports system submissions to the IWSLT 2025 Model Compression track. The authors compress Qwen2-Audio-7B-Instruct for English-to-German and English-to-Chinese speech translation using three techniques: full fine-tuning on ACL 60/60, QLoRA with 4-bit quantization plus knowledge distillation, and iterative pruning of decoder layers followed by fine-tuning, knowledge distillation, and QLoRA. The paper claims up to a 50% reduction in both model parameters and storage footprint while retaining 97-100% of the translation quality of the fully fine-tuned teacher model, with BLEU/chrF/COMET scores reported on a 100-utterance test split.

Significance. If the central efficiency and quality-retention claims were valid, the paper would offer a practical and reproducible recipe for compressing a large audio-language model, which is useful for the deployment-oriented IWSLT community. The ablation study is a genuine strength: it compares iterative pruning with middle-layer pruning, decoder-only with encoder-decoder pruning, pruning depth, immediate versus delayed recovery, and out-of-domain data size. The public code release is also commendable. However, the significance is currently limited by two load-bearing issues: the reported parameter reductions are conflated with storage savings, and the evaluation protocol selects models on the same 100-utterance test set used for final scoring. Both issues need to be resolved before the paper's headline claims can be accepted.

major comments (3)
  1. [§2.1, §3.3, §4.7] The 100-utterance test split is used both to select which decoder layers to prune (Section 3.3 states: 'we evaluate the translation performance of the pruned model on the test split of the ACL 60/60 dataset, after removing each candidate layer') and to choose the out-of-domain data size (Section 4.7, Table 8). The final quality-retention numbers in Table 2 are therefore computed on the same set used for model selection, which introduces selection bias and makes the reported 97-100% retention optimistic. A separate held-out validation set for pruning and hyperparameter decisions, or a fresh test set for final evaluation, is required to support the central claim.
  2. [Abstract; Table 2; §3.2–3.3] The claim of 'up to a 50% reduction in both model parameters and storage footprint' conflates parameter count with storage. QLoRA 4-bit quantization does not remove parameters; it stores existing weights at lower precision and adds LoRA adapters. In Table 2, the pruned model has 6.78 B parameters in 13.55 GB, but the QLoRA row reports 4.12 B parameters in 8.65 GB, which is inconsistent with a true parameter count: 8.65 GB / 2 bytes per bf16 parameter is approximately 4.32 B, so the 'Params' column appears to be storage normalized to bf16 rather than an actual parameter count. The real parameter reduction from pruning 8 of 32 decoder layers is 8.40 → 6.78 B, about 19%, not 50%. In addition, a genuinely 4-bit quantized 6.78 B-parameter model would occupy roughly 3.4 GB, not 8.65 GB; the storage accounting needs to be clarified. The abstract and the text around Tables 1 and 2 should be revised to report actual parameter reductions and storage reductions separately.
  3. [§5; Table 2] All headline conclusions rest on differences over 100 test utterances, and no confidence intervals or significance tests are reported. For example, the EN-DE BLEU difference between teacher (39.28) and student (39.59) is only 0.31 points, while COMET differs by 2.89 points; the EN-ZH BLEU difference is 58.54 versus 56.52, a 2.02-point drop. Without variance estimates, claims like 'retaining 97-100% of translation quality' cannot be distinguished from noise. Please report confidence intervals, bootstrap resampling, or a significance test on the 100-utterance test set, and avoid over-interpreting small metric differences.
minor comments (4)
  1. [§4.1] There is a typo in 'Qwen2-Audio-7B-Insruct'; it should read 'Qwen2-Audio-7B-Instruct'.
  2. [Table 7] The symbol '♂redoFT' is unconventional and unclear; please replace it with a plain label such as '+ FT [immediate]'.
  3. [Footnote 6] Footnote 6 states that 'the Qwen2-Audio model must initially be loaded in bfloat16 data type' to achieve the compression gains; the connection between dtype at load time and achievable pruning/quantization compression is not explained and should be clarified.
  4. [§3.2] The statement that QLoRA 'results in 2.41% trainable parameters of the model' should specify whether the percentage is relative to the original 8.40 B parameters or to the storage-equivalent figure reported in Table 1, since the two bases give different impressions.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline compression claim is partially self-definitional: the parameter-reduction figure is a bf16 rescaling of the storage figure, while the distillation/quality-retention chain is independent.

  1. self definitional [Abstract; Tables 1-2; Section 3.3; footnote 5]
    "Our pruned (student) models achieve up to a 50% reduction in both model parameters and storage footprint ... Pruning 8 layers achieves a 20% reduction in model size, which increases to 50% when combined with 4-bit quantization. (Table 2: '+ QLoRA + CV ... 4.12 8.65'; footnote: 'our storage footprint calculations are based only on the size of model files (*.safetensors), including the adapter of QLoRA models.')"

    In the non-quantized rows, Params (B) is exactly Storage (GB)/2 (8.40/16.79; 6.78/13.55), i.e. the bf16 storage per parameter. The QLoRA rows continue to report a roughly halved number (4.95/9.64; 4.12/8.65). QLoRA quantizes existing weights and adds low-rank adapters; it does not delete the ~6.78B frozen parameters of the pruned model. The headline parameter reduction is therefore not an independent measurement: it is the storage reduction re-expressed in bf16-equivalent units. The claim of 'reduction in both model parameters and storage' reduces to the storage reduction by construction.

full rationale

The translation-quality derivation is not circular. The teacher is the fully fine-tuned external reference; distillation data are generated from that teacher; and all final BLEU/chrF/COMET values are direct evaluations of the student against that reference, not quantities fitted from the compression hyperparameters. The pruning and KD pipeline has independent experimental content. The circularity is confined to the headline compression metric: the tables' Params column is a bf16 rescaling of the storage footprint, so the '50% reduction in both parameters and storage' is a single measurement presented as two. A separate evaluation-protocol concern, not counted as derivation-circularity here, is that Section 3.3 selects pruned layers on the same 100-utterance test split later used to report 97-100% retention; this biases the estimate but does not make it equivalent to the method's inputs. Overall, the central quality-retention claim remains independently grounded, so the score is moderate.

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

The central claim depends on several hand-chosen hyperparameters (pruning depth, LoRA rank/alpha, OOD data size, KD oversampling) that are fitted through ablations, and on the independence of the test set, which is violated by using the same set for pruning decisions. No new entities or phenomena are introduced.

free parameters (5)
  • Number of pruned decoder layers (n) = 8
    Chosen from ablation experiments (Section 4.5) comparing 8, 10, 12, 16 layers; 8 gives the best trade-off. This is a hand-picked structural hyperparameter, not derived.
  • QLoRA rank = 64
    Set to 64 with alpha 128 and dropout 0 (Section 3.2). Selected by the author; no sensitivity analysis is given.
  • QLoRA alpha = 128
    Standard LoRA scaling factor chosen alongside rank 64.
  • Out-of-domain data size (CoVoST2) = 100k segments
    Selected from {10k, 50k, 80k, 100k} in Table 8 as the best COMET score. The exact choice is fitted to the development results.
  • Knowledge distillation oversampling factor = 10
    KD data oversampled by factor 10 in the final QLoRA stage (Section 3.3). No ablation on this factor.
assumptions (4)
  • domain assumption The 100-utterance test split is a valid, independent evaluation set for speech translation quality.
    The paper samples 100 utterances from the merged ACL 60/60 dev+eval with seed 0 and uses them as test (Section 2.1). The independence of this set is broken by using it for layer-pruning decisions (Section 3.3).
  • domain assumption chrF/chrF++ is a suitable proxy for layer importance in pruning.
    Layer importance is measured by translation performance (chrF/chrF++) on the test split (Section 3.3). This assumes the metric reflects true layer contribution in a way that transfers to other data.
  • domain assumption The ACL 60/60 dataset is representative of speech translation quality for this task.
    All in-domain training and evaluation use ACL 60/60 (784 training, 100 test). The generalization of the compression trade-off to other domains is not established.
  • domain assumption Qwen2-Audio-7B-Instruct is an appropriate base model with stable layer-wise behavior.
    The paper uses this model exclusively (Section 4.1 compares to base model and picks Instruct). The pruning and compression results may not transfer to other audio-language models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Speech Translation through Model Compression and Knowledge Distillation." pith.science (2026). https://pith.science/paper/57R4GGOR

@misc{pith2026250520237,
  author       = {Pith},
  title        = {Pith review of: Efficient Speech Translation through Model Compression and Knowledge Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/57R4GGOR}},
  note         = {Machine review of arXiv:2505.20237}
}
read the original abstract

Efficient deployment of large audio-language models for speech translation remains challenging due to their significant computational requirements. In this paper, we address this challenge through our system submissions to the "Model Compression" track at the International Conference on Spoken Language Translation (IWSLT 2025). We experiment with a combination of approaches including iterative layer pruning based on layer importance evaluation, low-rank adaptation with 4-bit quantization (QLoRA), and knowledge distillation. In our experiments, we use Qwen2-Audio-7B-Instruct for speech translation into German and Chinese. Our pruned (student) models achieve up to a 50% reduction in both model parameters and storage footprint, while retaining 97-100% of the translation quality of the in-domain (teacher) models.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

    work page

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    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 ...

    work page 2025

  4. [4]

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. 2024. http://arxiv.org/abs/2407.10759 Qwen2-Audio Technical Report . arXiv [eess.AS]

    arXiv 2024

  5. [5]

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. http://arxiv.org/abs/2311.07919 Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models . arXiv [eess.AS]

    arXiv 2023

  6. [6]

    Josep Crego and Jean Senellart. 2016. http://arxiv.org/abs/1612.06139 Neural Machine Translation from Simplified Translations . arXiv [cs.CL]

    work page Pith review arXiv 2016

  7. [7]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. http://arxiv.org/abs/2305.14314 QLoRA: Efficient Finetuning of Quantized LLMs . arXiv [cs.LG]

    arXiv 2023

  8. [8]

    Sanchit Gandhi, Patrick von Platen, and Alexander M Rush. 2023. http://arxiv.org/abs/2311.00430 Distil-Whisper: Robust knowledge distillation via large-scale pseudo labelling . arXiv [cs.CL]

    arXiv 2023

Show all 31 references
  1. [9]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. http://arxiv.org/abs/2106.09685 LoRA: Low-Rank Adaptation of Large Language Models . arXiv [cs.CL]

    2021 arXiv

  2. [10]

    Wandri Jooste, Andy Way, Rejwanul Haque, and Riccardo Superbo. 2022. https://aclanthology.org/2022.amta-upg.16 Knowledge Distillation for Sustainable Neural Machine Translation . In Proceedings of the 15th Biennial Conference of the Association for Machine Translation in the A...

    2022

  3. [11]

    Damjan Kalajdzievski. 2023. http://arxiv.org/abs/2312.03732 A rank stabilization scaling factor for fine-tuning with LoRA . arXiv [cs.CL]

    2023 arXiv

  4. [12]

    Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, Changbae Ahn, Seonghoon Yang, Sukyung Lee, Hyunbyung Park, Gyoungjin Gim, Mikyoung Cha, Hwalsuk Lee, and Sunghun Kim. 2023. http://arxiv.org/abs/2312.15...

    2023 arXiv

  5. [13]

    Yoon Kim and Alexander M Rush. 2016. https://aclanthology.org/D16-1139 Sequence-Level Knowledge Distillation . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages 1317--1327, Austin, Texas. Association for Computational Linguistics

    2016

  6. [14]

    Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Ca...

    2021

  7. [15]

    Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, K Keutzer, D Klein, and Joseph E Gonzalez. 2020. https://dl.acm.org/doi/10.5555/3524938.3525491 Train large, then compress: Rethinking model size for efficient training and inference of transformers . In Proceedings of the 37th ...

    2020

  8. [16]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://aclanthology.org/P02-1040 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--318, P...

    2002

  9. [17]

    David Peer, Sebastian Stabinger, Stefan Engl, and Antonio Rodríguez-Sánchez. 2022. https://www.sciencedirect.com/science/article/pii/S0167865522000885 Greedy-layer pruning: Speeding up transformer models for natural language processing . Pattern Recognit. Lett., 157:76--82

    2022

  10. [18]

    Yifan Peng, Jaesong Lee, and Shinji Watanabe. 2023 a . https://ieeexplore.ieee.org/document/10096662 I3D: Transformer architectures with input-dependent dynamic depth for speech recognition . In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal P...

    2023

  11. [19]

    Yifan Peng, Yui Sudo, Shakeel Muhammad, and Shinji Watanabe. 2023 b . http://dx.doi.org/10.21437/Interspeech.2023-1213 DPHuBERT: Joint Distillation and Pruning of Self-Supervised Speech Models . In Proceedings of the 24th Annual Conference of the International Speech Communica...

    2023 doi

  12. [20]

    Maja Popović. 2017. https://aclanthology.org/W17-4770 chr F ++: words helping character n-grams . In Proceedings of the Second Conference on Machine Translation , pages 612--618, Copenhagen, Denmark. Association for Computational Linguistics

    2017

  13. [21]

    Matt Post. 2018. https://aclanthology.org/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

  14. [22]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. https://aclanthology.org/2020.emnlp-main.213 COMET : A Neural Framework for MT Evaluation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 2685--2702, On...

    2020

  15. [23]

    Pedram Rostami and Mohammad Javad Dousti. 2024. http://arxiv.org/abs/2411.06506 CULL-MT: Compression using language and layer pruning for machine translation . arXiv [cs.CL]

    2024 arXiv

  16. [24]

    Hassan Sajjad, Fahim Dalvi, Nadir Durrani, and Preslav Nakov. 2023. https://dl.acm.org/doi/10.1016/j.csl.2022.101429 On the effect of dropping layers of pre-trained transformer models . Comput. Speech Lang., 77(101429):101429

    2023

  17. [25]

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

    2023

  18. [26]

    Marcos Treviso, Ji-Ung Lee, Tianchu Ji, Betty van Aken, Qingqing Cao, Manuel R Ciosici, Michael Hassid, Kenneth Heafield, Sara Hooker, Colin Raffel, Pedro H Martins, André F T Martins, Jessica Zosa Forde, Peter Milder, Edwin Simpson, Noam Slonim, Jesse Dodge, Emma Strubell, Ni...

    2023 doi

  19. [27]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. http://arxiv.org/abs/1706.03762 Attention Is All You Need . In Advances in Neural Information Processing Systems (NIPS 2017) , volume 30. Curran A...

    2017 arXiv

  20. [28]

    Changhan Wang, Anne Wu, and Juan Pino. 2021. http://dx.doi.org/10.21437/Interspeech.2021-2027 CoVoST 2 and Massively Multilingual Speech-to-Text Translation . In Proceedings of the 22nd Annual Conference of the International Speech Communication Association, InterSpeech 2021 ,...

    2021 doi

  21. [29]

    Haoyu Wang, Siyuan Wang, Wei-Qiang Zhang, Suo Hongbin, and Yulong Wan. 2023. http://dx.doi.org/10.21437/Interspeech.2023-1442 Task-Agnostic Structured Pruning of Speech Representation Models . In Proceedings of the 24th Annual Conference of the International Speech Communicati...

    2023 doi

  22. [30]

    Haryo Akbarianto Wibowo, Haiyue Song, Hideki Tanaka, Masao Utiyama, Alham Fikri Aji, and Raj Dabre. 2025. http://arxiv.org/abs/2503.06291 IteRABRe: Iterative recovery-aided block reduction . arXiv [cs.CL]

    2025 arXiv

  23. [31]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

Pith tools

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