Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Attention2Probability claims that cross-attention weights between speech and candidate terms can be directly converted into the probability that each term occurs in the speech, replacing vector-database retrieval for terminology biasing in

desk verdict A useful new dataset and a plausible attention-based retriever, but the headline comparison against VectorDB is confounded by encoder choice, so the core advantage is not proven. read the letter →

arxiv 2508.18701 v1 pith:TWBJBROK submitted 2025-08-26 cs.CL

Yanfan Du , Jun Zhang , Bin Wang , Jin Qiu , Lu Huang , Yuan Ge , Xiaoqian Liu , Tong Xiao
show 1 more author
Jingbo Zhu
This is my paper · ORCID
classification cs.CL
keywords Attention2Probabilityterminologyretrievalspeechlanguagemodelscross-attentionbiasingASRtranslationcurriculumlearning
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

Attention2Probability (A2P) is a lightweight retriever that estimates whether a term from a predefined bank actually occurs in an audio clip, using the cross-attention already present between speech and text in a speech large language model (SLM). The paper's central claim is that these attention weights can be read as term-presence probabilities, so no vector database and no dedicated speech-text alignment training are needed. On the released bilingual test set, A2P reaches 92.57% term recall for Chinese and 86.83% for English at 8.71 ms per query, and feeding the retrieved terms into the SLM improves terminology accuracy of ASR and ST by 6–17%. The paper also releases a new speech dataset with terminology and shows that current SLMs do not fully exploit correctly retrieved terms, especially in translation.

What carries the argument

The central object is a single-layer multi-head cross-attention module that computes attention weights between frozen speech features and frozen term embeddings, converted into term-presence probabilities by masked token-level summation, length normalization, a residual connection, and a linear+sigmoid head. The load-bearing design choice is the token-level pooling: it aggregates subword attention weights over the whole term, which the ablation shows is necessary for the model to treat a term as a unit rather than as isolated tokens. Curriculum learning (word-level, then phrase-level, then real-term) is what makes this pooler trainable when authentic term-audio pairs are scarce.

What would settle it

Take the released test set, pick a term that occurs in exactly half the clips, hold the term bank fixed, and measure A2P's probability score on clips with vs without the term; if the score distribution does not separate the two groups, the sigmoid output is not a presence probability but a ranking artifact.

Watch

Extended reading notes

Core claim

Given a speech input and a bank of candidate terms, A2P extracts speech features with the frozen audio encoder of Qwen2-Audio-Instruction and term features with the same model's frozen text embeddings. A single multi-head cross-attention layer computes attention from speech to each term; the token-level weights are masked, summed over the term's tokens, normalized by term length, and combined with the term embedding through a residual connection. A linear layer and sigmoid turn the result into a per-term presence probability. The retriever is trained on positive and negative term pairs with a curriculum that moves from single words to phrases to real terms, and at inference the top-k terms b

Load-bearing premise

The frozen Qwen2-Audio-Instruction encoder and text embedding layer already share a joint cross-modal space, so a single learned cross-attention layer plus linear+sigmoid can turn attention weights into calibrated term-presence probabilities without any dedicated modality-alignment training.

Editorial extensions

If this is right

  • A2P gives a retrieval paradigm for speech terminology that does not require learning a separate speech-text alignment model.
  • At 8.71 ms per query, the retriever is fast enough for interactive prompting, though slower than VectorDB at larger term banks.
  • Adding retrieved terms to the prompt raises terminology accuracy in ASR by roughly 5–6% and in ST by about 12–13%, with BLEU/CER/WER mostly preserved or slightly improved.
  • The released dataset and code let other teams reproduce the 92.57% / 86.83% recall numbers and build on them.
  • The paper's finding that ST terminology accuracy stays far below retriever recall pinpoints a concrete weakness: the SLM fails to use many correctly retrieved terms.

Reading between the lines

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

  • Because the retriever relies only on the frozen audio encoder's existing cross-modal representation, the same attention-to-probability head could be grafted onto other jointly trained speech LLMs, so A2P is less a new model than a probe of the encoder's latent alignment.
  • The paper's own latency breakdown (top-k selection consuming ~60% of the 8.71 ms) suggests that replacing exact sorting with approximate or GPU-native top-k could close much of the speed gap to VectorDB without changing the model.
  • A calibrated presence probability could be used downstream for confidence-weighted fusion or selective prompting—e.g., only intervening when the top term's probability clears a threshold—an extension the paper does not explore.
  • One testable extension is cross-lingual retrieval: if the joint space is truly shared, the same head could estimate presence of target-language terms directly from source speech, bypassing the source-term retrieval step.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Attention2Probability (A2P), a lightweight terminology retriever for speech-to-text systems. A2P uses the frozen audio encoder and text embedding layer of Qwen2-Audio-Instruction, applies a single-layer multi-head cross-attention between speech and candidate terms, pools token-level attention weights by term length, and maps the result through a linear layer and sigmoid to estimate the probability that a term occurs in the audio. The retriever is trained with a BCE loss on positive and negative term-speech pairs, using a curriculum from word-level to phrase-level to real terms. At inference, the Top-k terms are appended to the SLM prompt. The authors also construct and release a speech dataset with terminology by synthesizing audio with MegaTTS from NER-derived terms and using LibriSpeech/Aishell-2. On a ComMT-derived test set, A2P achieves 75.55/86.83 (Top-10/Top-50) recall for English and 83.31/92.57 for Chinese, with 8.71 ms latency at 583 terms, and reports 6-17% terminology-accuracy improvements when intervening in ASR and ST tasks.

Significance. If the reported results hold under controlled comparison, A2P would offer a practical alternative to vector-database retrieval for terminology biasing in speech LLMs, potentially removing the need for explicit modality-alignment training and large-scale contrastive pretraining. The public release of a speech-terminology dataset is also a useful community contribution. However, the central claim that the cross-attention mechanism outperforms VectorDB is not currently identifiable from the experiments because the comparison changes retriever type, audio encoder, and training status simultaneously. The paper's own ablations show extreme sensitivity to the choice of audio encoder, so the headline numbers may be attributable to the Qwen2-Audio-Instruction feature space rather than the proposed architecture. The paper is honest about latency limitations and includes ablations for curriculum learning and pooling, which are strengths, but the missing controlled baseline is load-bearing.

major comments (3)
  1. [Table 2, 'Main Results'] The headline comparison conflates three variables: retriever type (cross-attention vs. cosine), encoder (Qwen2-Audio-Instruction vs. SONAR), and training status (supervised A2P vs. zero-shot VectorDB). The A2P+SONAR row is far below VectorDB+SONAR (EN Top-10/Top-50: 15.15/33.93 vs. 62.89/83.49; ZH <1%), and A2P+Qwen-Audio-Chat is also much worse than A2P+Qwen2-Audio-Instruction. Thus the EN 75.55/86.83 and ZH 83.31/92.57 numbers cannot be attributed to the cross-attention probability mechanism rather than to the Qwen2-Audio-Instruction encoder or to in-domain supervised training. Please report (a) a VectorDB baseline using Qwen2-Audio-Instruction features (with appropriate pooling) and (b) an untrained/randomly initialized A2P under the same encoder, so that the retriever architecture is independently identifiable.
  2. [Abstract and 'Cross-Modal Retriever Training'] The claim that A2P works 'without requiring any dedicated modal alignment training' is contradicted by Eq. (10), where a BCE loss supervises the cross-attention retriever to map audio features to term-presence labels. This is itself a supervised cross-modal alignment objective, even if the backbone is frozen. Please either revise the claim to 'without fine-tuning the speech encoder' or provide a genuinely zero-shot variant (e.g., raw attention without the trained linear+sigmoid) to support the stronger statement. As written, the contribution is overstated.
  3. [Table 5 and 'Experiments'] The 6-17% terminology-accuracy improvements in ASR/ST are reported on a single TTS-generated test set with no error bars, significance tests, or multiple synthesis seeds. Given the 3.3K-sentence test set and the stochastic nature of MegaTTS, these improvements may be within run-to-run variability. Similarly, Table 2 reports recall without variance estimates. Please provide confidence intervals (e.g., bootstrap over test utterances or multiple TTS seeds) or at least a clear statement of how many synthesis passes were used. Without this, the intervention claim is not statistically grounded.
minor comments (5)
  1. [Abstract / Section 'Advantages of A2P'] Typo: 'Advantages of of A2P' (duplicate 'of').
  2. [Table 4] Typo: 'pharse-level' should be 'phrase-level'.
  3. [Table 5] The caption says 'The A/B metric represents text quality versus terminology generation accuracy', but the table entries like '12.29/79.66' are not clearly defined. Please state explicitly that A is WER/CER or BLEU and B is terminology accuracy, and define how terminology accuracy is computed.
  4. [Section 'Different Sizes of Terminology Databases'] Figure 4's x-axis labels '1,2,3,4' are ambiguous; spell out the term-bank sizes (583, 1k, 5k, 10k) directly on the axis or in the caption.
  5. [Dataset description] The released dataset is not precisely specified: Table 1 lists many sources (LibriSpeech, Aishell-2, NER datasets, Niutrans, ComMT), but it is unclear which subsets are included in the public release, how MegaTTS prompts were constructed, and how negative samples are sampled per batch. Please clarify the dataset composition and release format.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: A2P is a supervised cross-attention classifier trained on held-out data; reported recalls are measured against an external benchmark-derived test set.

full rationale

The paper's derivation chain is a standard learned classifier. Eq. 3 computes multi-head cross-attention between speech and term features; Eqs. 4–7 pool token weights; Eq. 8 adds a residual; Eq. 9 maps the result through Linear+Sigmoid to a presence probability; Eq. 10 trains this with binary cross-entropy on positive and negative samples. There is no equation in which the output is defined in terms of the target quantity, nor is any fitted parameter renamed as a prediction. Training data (LibriSpeech, Aishell-2, NER-derived TTS data) is disjoint from the evaluation test set (ComMT terms sourced from WMT/DAS), so the headline recall numbers are not fitted to the test distribution. The ComMT reference includes overlapping authors (Xiao, Zhu, Liu), but the labels come from external benchmarks, so this is at most a mild self-citation and is not load-bearing. The claim that no dedicated modal alignment training is required is arguably undercut by the supervised BCE objective, but that is an internal-consistency interpretation issue, not a circular reduction. The confounded VectorDB-vs-A2P comparison (different encoder, training status, and retriever type change simultaneously) is a legitimate experimental-validity criticism, but it is not a circularity: none of the compared systems' outputs reduce to the paper's own inputs by construction.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper is an empirical ML submission: it does not derive its main result from axioms. The ledger below captures the assumptions that must hold for the reported recall and downstream gains to transfer beyond the exact synthetic setup. The main load-bearing items are the pretrained encoder's alignment, the fidelity of MegaTTS audio, and the accuracy of automatically generated term labels. No invented theoretical entities are introduced. The retriever's weights are learned, not hand-set; the free parameters listed are the hand-chosen hyperparameters and evaluation choices that shape the results.

free parameters (7)
  • epsilon (epsilon) = not reported (minimal)
    Added to denominator in Eq. 7 to avoid division by zero; a hand-chosen stabilizer.
  • Top-k evaluation values = 10, 20, 30, 40, 50
    Evaluation thresholds chosen for reporting recall; no single k is justified.
  • Curriculum stage order = word-level, phrase-level, real-term
    Hand-designed training schedule; ablation shows direct real-term training fails, so this order is load-bearing for the reported results.
  • Retriever hidden size = 4096
    Set to match Qwen2-Audio-Instruction encoder dimension; hand-chosen.
  • MHA heads and dropout = 32 heads, dropout 0.1
    Hyperparameters chosen without stated tuning procedure.
  • Training hyperparameters = batch 32, lr 1e-4, 50 epochs, warmup 500, AdamW beta 0.9/0.98, weight decay 0.01
    Standard choices; no sensitivity analysis.
  • Term bank capacity per batch = 100 terms
    Limits negative samples per batch; affects training difficulty.
assumptions (6)
  • domain assumption Qwen2-Audio-Instruction's frozen audio encoder and text embeddings share a joint space sufficient for cross-attention to encode term presence (Sec. 'Multi-modal Feature Extraction').
    The method provides no dedicated alignment training and the paper states the feature space inherently encodes both modalities; the strong dependence on encoder choice (Table 2) shows this is load-bearing.
  • domain assumption MegaTTS-generated audio is a valid proxy for real speech for training and evaluation (Sec. 'Dataset').
    All training and test audio is TTS; the paper does not evaluate on real spontaneous speech, so the 'robust' claim depends on TTS fidelity.
  • domain assumption A single linear layer after pooled cross-attention can convert attention weights into calibrated presence probabilities (Eq. 9).
    The paper treats the sigmoid output as a probability without calibration analysis.
  • domain assumption Terminology ground-truth labels in the constructed datasets are accurate (Sec. 'Dataset').
    For LibriSpeech and Aishell-2, terms are randomly designated spans, so 'presence' is definitionally true; for NER-derived data, entity names are assumed to be spoken in the TTS definitions.
  • domain assumption The instruction fine-tuning on Niutrans with random terminology conditions transfers to the test set (Sec. 'Instruction Fine-Tuning').
    The SLM is fine-tuned on one corpus and evaluated on another; cross-corpus generalization is assumed.
  • standard math Cross-attention and standard optimization (AdamW, CosineAnnealingLR) converge to a useful solution (Sec. 'Retriever Training and Inference').
    Standard ML assumptions; not proven, but not controversial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System." pith.science (2026). https://pith.science/paper/TWBJBROK

@misc{pith2026250818701,
  author       = {Pith},
  title        = {Pith review of: Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TWBJBROK}},
  note         = {Machine review of arXiv:2508.18701}
}
read the original abstract

Recent advances in speech large language models (SLMs) have improved speech recognition and translation in general domains, but accurately generating domain-specific terms or neologisms remains challenging. To address this, we propose Attention2Probability: attention-driven terminology probability estimation for robust speech-to-text system, which is lightweight, flexible, and accurate. Attention2Probability converts cross-attention weights between speech and terminology into presence probabilities, and it further employs curriculum learning to enhance retrieval accuracy. Furthermore, to tackle the lack of data for speech-to-text tasks with terminology intervention, we create and release a new speech dataset with terminology to support future research in this area. Experimental results show that Attention2Probability significantly outperforms the VectorDB method on our test set. Specifically, its maximum recall rates reach 92.57% for Chinese and 86.83% for English. This high recall is achieved with a latency of only 8.71ms per query. Intervening in SLMs' recognition and translation tasks using Attention2Probability-retrieved terms improves terminology accuracy by 6-17%, while revealing that the current utilization of terminology by SLMs has limitations.

Figures

Figures reproduced from arXiv: 2508.18701 by the authors.

Figure 1
Figure 1. The overall architecture of Attention2Probability. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Schematic of token-level pooling. For example, a [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. The trend of the recall rate when k=50 as the ter [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 24 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  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]

    Bengio, Y.; Louradour, J.; Collobert, R.; and Weston, J. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, 41--48

    work page 2009

  4. [4]

    Bergmanis, T.; and Pinnis, M. 2021. Facilitating terminology translation with target lemma annotations. arXiv preprint arXiv:2101.10035

    work page Pith review arXiv 2021

  5. [5]

    Beurer-Kellner, L.; Fischer, M.; and Vechev, M. 2024. Guiding llms the right way: Fast, non-invasive constrained generation. arXiv preprint arXiv:2403.06988

    arXiv 2024

  6. [6]

    C.; Li, J.; Ghosh, S.; Balam, J.; and Ginsburg, B

    Chen, Z.; Huang, H.; Andrusenko, A.; Hrinchuk, O.; Puvvada, K. C.; Li, J.; Ghosh, S.; Balam, J.; and Ginsburg, B. 2024. Salm: Speech-augmented language model with in-context learning for speech recognition and translation. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 13521--13525. IEEE

    work page 2024

  7. [7]

    Chu, Y.; Xu, J.; Zhou, X.; Yang, Q.; Zhang, S.; Yan, Z.; Zhou, C.; and Zhou, J. 2023. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models. arXiv preprint arXiv:2311.07919

    arXiv 2023

  8. [8]

    Coupland, N. 2014. Language change, social change, sociolinguistic change: A meta-commentary. Journal of Sociolinguistics, 18(2)

    work page 2014

Show all 42 references
  1. [9]

    Crego, J.; Kim, J.; Klein, G.; Rebollo, A.; Yang, K.; Senellart, J.; Akhanov, E.; Brunelle, P.; Coquard, A.; Deng, Y.; et al. 2016. Systran's pure neural machine translation systems. arXiv preprint arXiv:1610.05540

    2016 arXiv

  2. [10]

    Ding, N.; Xu, G.; Chen, Y.; Wang, X.; Han, X.; Xie, P.; Zheng, H.-T.; and Liu, Z. 2021. Few-nerd: A few-shot named entity recognition dataset. arXiv preprint arXiv:2105.07464

    2021 arXiv

  3. [11]

    Dinu, G.; Mathur, P.; Federico, M.; and Al-Onaizan, Y. 2019. Training neural machine translation to apply terminology constraints. arXiv preprint arXiv:1906.01105

    2019 arXiv

  4. [12]

    Du, J.; Na, X.; Liu, X.; and Bu, H. 2018. Aishell-2: Transforming mandarin asr research into industrial scale. arXiv preprint arXiv:1808.10583

    2018 arXiv

  5. [13]

    Du, X.; Jia, Y.; and Zan, H. 2022. MRC-based medical NER with multi-task learning and multi-strategies. In China National Conference on Chinese Computational Linguistics, 149--162. Springer

    2022

  6. [14]

    Elizalde, B.; Deshmukh, S.; Al Ismail, M.; and Wang, H. 2023. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5. IEEE

    2023

  7. [15]

    Feng, P.; Ma, Z.; Chen, W.; Li, Y.; Wang, S.; Yu, K.; and Chen, X. 2025. Enhancing Speech-to-Speech Dialogue Modeling with End-to-End Retrieval-Augmented Generation. arXiv preprint arXiv:2505.00028

    2025

  8. [16]

    Gao, Y.; Xiong, Y.; Gao, X.; Jia, K.; Pan, J.; Bi, Y.; Dai, Y.; Sun, J.; Wang, H.; and Wang, H. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2(1)

    2023 arXiv

  9. [17]

    Gong, X.; Lv, A.; Wang, Z.; and Qian, Y. 2024. Contextual Biasing Speech Recognition in Speech-enhanced Large Language Model. Proc. Interspeech. ISCA, 257--261

    2024

  10. [18]

    Gong, X.; Lv, A.; Wang, Z.; Zhu, H.; and Qian, Y. 2025. BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM. arXiv preprint arXiv:2505.19179

    2025 arXiv

  11. [19]

    Hasler, E.; De Gispert, A.; Iglesias, G.; and Byrne, B. 2018. Neural machine translation decoding with terminology constraints. arXiv preprint arXiv:1805.03750

    2018 arXiv

  12. [20]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  13. [21]

    Hokamp, C.; and Liu, Q. 2017. Lexically constrained decoding for sequence generation using grid beam search. arXiv preprint arXiv:1704.07138

    2017 arXiv

  14. [22]

    Jiang, Z.; Ren, Y.; Li, R.; Ji, S.; Zhang, B.; Ye, Z.; Zhang, C.; Jionghao, B.; Yang, X.; Zuo, J.; et al. 2025. Megatts 3: Sparse alignment enhanced latent diffusion transformer for zero-shot speech synthesis. arXiv preprint arXiv:2502.18924

    2025 arXiv

  15. [23]

    Kim, S.; Hori, T.; and Watanabe, S. 2017. Joint CTC-attention based end-to-end speech recognition using multi-task learning. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), 4835--4839. IEEE

    2017

  16. [24]

    L.; and Fuegen, C

    Lakomkin, E.; Wu, C.; Fathullah, Y.; Kalinli, O.; Seltzer, M. L.; and Fuegen, C. 2024. End-to-end speech recognition contextualization with large language models. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 12406--12410. IEEE

    2024

  17. [25]

    Le, D.; Jain, M.; Keren, G.; Kim, S.; Shi, Y.; Mahadeokar, J.; Chan, J.; Shangguan, Y.; Fuegen, C.; Kalinli, O.; et al. 2021. Contextualized streaming end-to-end speech recognition with trie-based deep biasing and shallow fusion. arXiv preprint arXiv:2104.02194

    2021 arXiv

  18. [26]

    Levow, G.-A. 2006. The third international Chinese language processing bakeoff: Word segmentation and named entity recognition. In Proceedings of the Fifth SIGHAN workshop on Chinese language processing, 108--117

    2006

  19. [27]

    u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \

    Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; K \"u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \"a schel, T.; et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33: 9459--9474

    2020

  20. [28]

    Luo, Y.; Zheng, T.; Mu, Y.; Li, B.; Zhang, Q.; Gao, Y.; Xu, Z.; Feng, P.; Liu, X.; Xiao, T.; et al. 2025. Beyond Decoder-only: Large Language Models Can be Good Encoders for Machine Translation. arXiv preprint arXiv:2503.06594

    2025 arXiv

  21. [29]

    M.; and Senellart, J

    Michon, E.; Crego, J. M.; and Senellart, J. 2020. Integrating domain terminology into neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics, 3925--3937

    2020

  22. [30]

    Morgan, R. 2025. Hermeneutical disarmament. The Philosophical Quarterly, 75(3): 1071--1093

    2025

  23. [31]

    Pan, X.; Zhang, B.; May, J.; Nothman, J.; Knight, K.; and Ji, H. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th annual meeting of the association for computational linguistics (volume 1: long papers), 1946--1958

    2017

  24. [32]

    Panayotov, V.; Chen, G.; Povey, D.; and Khudanpur, S. 2015. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), 5206--5210. IEEE

    2015

  25. [33]

    Post, M.; and Vilar, D. 2018. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. arXiv preprint arXiv:1804.06609

    2018 arXiv

  26. [34]

    Rajbhandari, S.; Rasley, J.; Ruwase, O.; and He, Y. 2020. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, 1--16. IEEE

    2020

  27. [35]

    Sun, C.; Liu, B.; Cui, Z.; Qi, A.; Zhang, T.-h.; Zhou, D.; and Lu, L. 2025. SEAL: Speech Embedding Alignment Learning for Speech Large Language Model with Retrieval-Augmented Generation. arXiv preprint arXiv:2502.02603

    2025

  28. [36]

    Tang, C.; Yu, W.; Sun, G.; Chen, X.; Tan, T.; Li, W.; Lu, L.; Ma, Z.; and Zhang, C. 2023. Salmonn: Towards generic hearing abilities for large language models. arXiv preprint arXiv:2310.13289

    2023 arXiv

  29. [37]

    Wang, X.; Chen, Y.; and Zhu, W. 2021. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9): 4555--4576

    2021

  30. [38]

    Wiltschko, M. 2008. The syntax of non-inflectional plural marking. Natural language & linguistic theory, 26(3): 639--694

    2008

  31. [39]

    Wu, S.; Tang, J.; Yang, C.; Zhang, P.; Yang, B.; Li, J.; Yao, J.; Zhang, M.; and Su, J. 2025. Locate-and-Focus: Enhancing Terminology Translation in Speech Language Models. arXiv preprint arXiv:2507.18263

    2025 arXiv

  32. [40]

    Xu, C.; Hu, B.; Li, Y.; Zhang, Y.; Ju, Q.; Xiao, T.; Zhu, J.; et al. 2021. Stacked acoustic-and-textual encoding: Integrating the pre-trained models into speech translation encoders. arXiv preprint arXiv:2105.05752

    2021 arXiv

  33. [41]

    Yang, X.; Kang, W.; Yao, Z.; Yang, Y.; Guo, L.; Kuang, F.; Lin, L.; and Povey, D. 2024. PromptASR for contextualized ASR with controllable style. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 10536--10540. IEEE

    2024

  34. [42]

    Zheng, Y.; Zhang, R.; Zhang, J.; Ye, Y.; Luo, Z.; Feng, Z.; and Ma, Y. 2024. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.