Pith. sign in

REVIEW 4 major objections 4 minor 25 references

A new Spanish speech benchmark, ESCUCHA, shows the best audio-language model scoring 74% against a trained human's 90%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 16:54 UTC pith:6OXHQJWL

load-bearing objection A genuinely useful Spanish in-the-wild speech benchmark with a real human–model gap, but the audio-necessity claim needs quantifiable support before the headline numbers can be fully trusted. the 4 major comments →

arxiv 2607.17812 v1 pith:6OXHQJWL submitted 2026-07-20 cs.CL

ESCUCHA: A Spanish Speech Benchmark for Heterogeneous Acoustic Conditions

Fernando L\'opez , Ana Ayala , Guillermo Segovia , Fernando Ib\'a\~nez , Ana Mart\'inez , Pablo G\'omez , Jordi Luque This is my paper
classification cs.CL
keywords Spanish speech benchmarklarge audio-language modelsin-the-wild audiopathological speechdysarthriaspeech understandingmultiple-choice question answeringaudio instruction following
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces ESCUCHA, claimed to be the first Spanish speech-understanding benchmark for large audio-language models, built from 1,000 human-curated questions over 162.9 hours of in-the-wild audio, including pathological speech from speakers with ALS and post-stroke conditions. Its purpose is to test whether audio models actually understand Spanish under heterogeneous acoustic conditions, not just transcribe normative speech. The evaluation finds a wide gap: the best model reaches 74.40% overall while a trained human annotator scores 90.10%, and most audio models sit near 50%. The paper argues this shows current audio models transfer poorly to Spanish and to non-normative speech, and that a text-only model reaching 41% points to how much of the benchmark may be answerable from language priors.

Core claim

On its own terms, ESCUCHA establishes that Spanish speech understanding under realistic conditions remains largely unsolved for large audio-language models. The benchmark is the first to combine in-the-wild Spanish audio with a multi-label taxonomy of nine perception and ten reasoning categories, including questions that require comparing two recordings, questions spoken aloud, and audio-delivered instructions with deterministic verification. Across six audio models, a cascade of speech recognition plus a text LLM, and text-only models, the best result (74.40%) falls far short of the trained human (90.10%), with a cascade (60.00%) outperforming all single audio models except the best, indica

What carries the argument

The load-bearing instrument is the benchmark itself: a four-stage construction pipeline (target refinement, wild-audio selection and question authoring, LLM-assisted review that flags items for audio necessity and distractor quality, then human and model evaluation), a multi-label perception/reasoning taxonomy, and two scoring protocols — deterministic exact-match for multiple choice and a deterministic verifier for audio-instruction-following items. The audio-necessity check is the part that is supposed to guarantee that the questions measure listening, not just reading.

Load-bearing premise

The benchmark measures speech understanding only if gold answers are not reachable from the question text, language priors, or recordings already memorized by the evaluated models; the paper's audio-necessity filter is a single LLM pass and there is no contamination analysis.

What would settle it

Conduct a transcript-only human evaluation: give 100 naive Spanish speakers the 900 MCQA questions with correct transcripts and ask them to answer without hearing the audio. If these respondents score near 74% (the best model) rather than near 25% (chance), then audio is not necessary for most questions and the central claim that ESCUCHA measures speech understanding fails. Conversely, if they score near chance, the text-only model's 41% is an artifact of language priors.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, ESCUCHA is a valid instrument for ranking LALMs on Spanish in-the-wild audio, and its results imply that high English/normative benchmark scores do not transfer to Spanish or pathological speech.
  • The 19- and 15-point drops for two models on pathological speech imply that non-normative speech remains a specific weak point for current audio models.
  • The cascade outperforming most end-to-end audio models implies that a large fraction of ESCUCHA's questions can be answered from a transcript alone, so the residual gap to the best model marks where genuine acoustic reasoning is still needed.
  • The near-chance performance of text-only models that refuse to answer (23.8% and 19.5%) and the 41% text-only score imply that language priors alone do not solve the benchmark, but they do set an upper bound on non-audio solvability.

Where Pith is reading between the lines

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

  • If the audio-necessity filter is tightened with human transcript-only baselines on all 900 MCQA items, the true audio-only gap could be smaller or larger; the paper's reported 16-point human-model gap may predominantly reflect text-accessible reasoning rather than acoustic understanding.
  • With contamination controls against training corpora, ESCUCHA could become a standard Spanish benchmark for assistive speech technology, since the pathological subset is the first LALM reasoning benchmark over dysarthric Spanish speech.
  • Because the spoken-question subset forces the model to recover the question from the audio, it could be used to test the robustness of instruction-following to language-model priors, a property the paper does not isolate.
  • The AIF subset's 88% score exceeding the human 79% suggests that some audio-instruction items are solvable without listening, so future versions should condition on audio-necessity per AIF item; the paper acknowledges this partially.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces ESCUCHA, a Spanish speech-understanding benchmark for large audio-language models (LALMs), containing 1,000 human-curated MCQA and audio-instruction-following (AIF) questions paired with 162.9 hours of in-the-wild audio. The benchmark claims to be the first Spanish in-the-wild speech benchmark and the first to evaluate reasoning over non-normative pathological speech. A multi-label taxonomy of 9 perception and 10 reasoning categories is derived from MMAU-Pro. The authors evaluate open and closed LALMs, a Whisper-plus-LLM cascade, text-only LLMs, and a human linguist. The best model (Qwen3-Omni-30B-A3B) reaches 74.40% overall against 90.10% for the human annotator, with most audio models near 50%.

Significance. If the validity concerns are resolved, ESCUCHA would be a valuable resource: it is the first Spanish LALM benchmark built from naturally occurring audio, includes a substantial non-normative speech subset (158 questions), spans very long audio contexts, and uses a deterministic verifier for AIF, which is a methodological strength. The reported human–model gap and the fact that a cascade outperforms most end-to-end LALMs are informative and falsifiable findings. The code and annotations are promised as open-source. However, the benchmark's central claim—that it measures audio-grounded understanding—depends on evidence that answers are not reachable from text, transcripts, or language priors, and that the human reference is reliable. Currently, that evidence is missing or only partially reported.

major comments (4)
  1. [§II.B (stage 3) and §IV, Table IV] The audio-necessity check is not quantified. The pipeline says 'Gemma-4-31B-IT' flags items for manual review, but the paper never reports how many items were flagged, rejected, or confirmed as audio-required. This is load-bearing because the text-only Qwen3-4B-Instruct-2507 reaches 41.00% and the Whisper-Large-v3 + Qwen3 cascade reaches 60.00% on MCQA, far above the 25.61% random baseline. These numbers indicate that a substantial fraction of items may be solvable from question structure, language priors, or transcript alone. To support the 'speech understanding' claim, the authors should report the per-item outcome of the audio-necessity check, provide a transcript-only human baseline, and/or show a per-item breakdown of which questions genuinely require acoustic information.
  2. [§IV, Table IV and §V (Limitations)] The human ceiling is based on a single trained linguist, and no inter-annotator reliability is reported. With a random baseline of 25.61% on MCQA, a single annotator's 91.33% on MCQA provides no estimate of question ambiguity or gold-label confidence. The paper should report agreement over multiple annotators at least on a subset, along with per-item gold confidence or adjudication procedures. Additionally, the finite sample size means that small differences in Table IV (e.g., 49.20 vs 49.30 or 50.22 vs 50.60) are likely within noise; binomial confidence intervals or a significance test should be reported before claiming that one model outperforms another near these values.
  3. [§V (Limitations) and Table IV, AIF column] The AIF subset is acknowledged to be partially answerable from the written instruction alone, yet AIF scores are reported as headline results, including Qwen3-Omni-30B-A3B at 88.00%, exceeding the human 79.00%. This undermines the claim that AIF exclusively measures audio instruction following. The paper should quantify the fraction of the 100 AIF items that can be satisfied without the audio, score models under both conditions (spoken instruction vs. written instruction in the prompt), and report AIF accuracy excluding leakage items. Without this, the AIF column overstates audio-based instruction-following ability.
  4. [§V (Limitations) and Table IV] The Gemini-2.5-Flash evaluation is explicitly biased: a payload limit forced exclusion of items above a size threshold, yet its overall score (52.80%) is placed in Table IV alongside models evaluated on the full set. This makes the comparison misleading, especially for long-audio items where other models show different behavior. The authors should either exclude Gemini-2.5-Flash from the main table, report it separately as a subset evaluation, or provide a matched comparison on the same subset with other models. The current presentation risks being interpreted as a fair full-benchmark result.
minor comments (4)
  1. [Throughout] Model names are inconsistent: 'Audio-Flamingo-3-HF' vs 'Audio Flamingo 3', 'V oxtral-Mini' with a spacing artifact, and 'Qwen3-Omni-30B-A3B' vs 'Qwen3-Omni-30B-A3B' in the abstract. Please standardize.
  2. [Figure 3 caption] The caption reads 'AIF questions only with perception labels, were excluded'—the comma is ungrammatical; suggest 'AIF questions, which only have perception labels, were excluded.'
  3. [§II.B] The four-stage pipeline is clear, but the description of Stage 3 says the LLM 'returns the identifiers of items it flags,' yet no examples or statistics of flagged items are given. A sample of flagged items and their resolutions would strengthen the description.
  4. [§II.D and Table I] The spoken-questions subset is generated with OmniVoice TTS, but the paper does not state whether the TTS audio was verified for naturalness or whether any spoken questions were discarded. A brief note would help.

Circularity Check

0 steps flagged

No significant circularity: ESCUCHA is an empirical benchmark; reported accuracies are measured outputs, not fitted constants, and the MMAU-Pro taxonomy borrowing is a methodological choice, not a self-justifying derivation.

full rationale

ESCUCHA's central contributions are the benchmark itself and the measured performance of models on it. These results are empirical observations, not predictions derived from fitted parameters or from the benchmark's own definitions. The only self-referential element is the taxonomy, which the paper explicitly states is derived from MMAU-Pro [15], a paper co-authored by one of the current authors. However, this borrowing is a design choice for labeling questions; it does not determine any model score, nor is it used to infer the human–model gap. The paper also reports an LLM-assisted audio-necessity review, but it does not claim a quantitative output from that review as a result; it is part of the curation pipeline. The admitted AIF limitation (some items answerable from the written instruction) and the text-only baseline (Qwen3 at 41%) are validity concerns about what the benchmark measures, not circular reasoning: the benchmark scores remain independent measurements. No equation or construction step reduces to its own input, and no load-bearing claim is justified solely by a self-citation. Thus the paper is not circular.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No free parameters or invented entities. The benchmark's validity rests on five untested assumptions: gold annotations reliable, YouTube representative of in-the-wild audio, LLM review effective, self-reported diagnoses accurate, and TTS faithful. None are machine-checked or externally benchmarked.

axioms (5)
  • domain assumption Human-authored gold answers and perception/reasoning labels are correct ground truth.
    No inter-annotator agreement or adjudication statistics are reported; all benchmark scores hinge on the correctness of the curated labels.
  • domain assumption YouTube recordings selected 'from the wild' are representative of real acoustic conditions for Spanish speech.
    Audio selection favors spontaneous speech and public recordings, but no acoustic diversity analysis beyond duration is provided in Section II-B.
  • ad hoc to paper LLM-assisted review with Gemma-4-31B-IT reliably flags annotation errors.
    The four review checks in Section II-B are performed by a single LLM with no reported calibration against human experts or reliability metrics.
  • domain assumption Self-reported diagnoses and video metadata correctly identify pathological speech.
    The non-normative subset relies on self-reports and metadata rather than clinical assessment; authors acknowledge this is not clinical usage.
  • domain assumption Spoken questions and audio instructions generated by OmniVoice TTS are faithful and intelligible in Spanish.
    No human listening test or verification of TTS output accuracy is reported in Section II-D.

pith-pipeline@v1.3.0-alltime-deepseek · 9712 in / 11672 out tokens · 120487 ms · 2026-08-01T16:54:16.021455+00:00 · methodology

0 comments
read the original abstract

As large audio language models (LALMs) advance, robust evaluation frameworks have become essential. In this context, Spanish speech understanding under realistic acoustic conditions has received particularly little attention. We introduce ESCUCHA, the first Spanish speech understanding benchmark designed to evaluate LALMs across heterogeneous acoustic conditions and reasoning abilities. ESCUCHA comprises 1,000 human-curated questions paired with audio, totaling 162.9 hours sourced directly ``from the wild'' rather than drawn from existing datasets, with durations ranging from a few seconds to over 80 minutes. The benchmark emphasizes reasoning, spanning 9 perceptual and 10 reasoning categories, and it captures linguistic diversity through multiple Spanish accents and non-normative speech. ESCUCHA further includes multi-audio questions, spoken questions, and audio instructions, and it flags which questions support open-ended evaluation. Benchmarking several state-of-the-art multimodal and speech models reveals substantial performance gaps relative to trained humans.

Figures

Figures reproduced from arXiv: 2607.17812 by Ana Ayala, Ana Mart\'inez, Fernando Ib\'a\~nez, Fernando L\'opez, Guillermo Segovia, Jordi Luque, Pablo G\'omez.

Figure 1
Figure 1. Figure 1: Example of chained cross-category reasoning. The question requires three steps: recognizing two numerical values (time since Irene’s diagnosis and the life expectancy she cites), mapping them onto a shared timeline, and comparing them. The most frequent failure is selecting (C) “No puede saberse,” which conflates the absence of an explicit comparison in the audio with genuine unknowability. TABLE I ESCUCHA… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline for constructing the ESCUCHA benchmark. The process comprises four stages: (i) target setting and refinement of the perception and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Perception × reasoning co-occurrence across the 900 MCQA questions. AIF questions only with perception labels, were excluded. Each cell counts question label pairs; questions carrying multiple perception or reasoning labels contribute to several cells, so cell totals exceed the question count and row and column sums are not disjoint. ALS (114 questions) and post-stroke speech (44), annotated by sex (123 ma… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of audio duration per question (log scale). The dashed [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

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

25 extracted references · 8 linked inside Pith

  1. [1]

    Listen, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. Glass, “Listen, think, and understand,” inProc. ICLR, 2024

    2024

  2. [2]

    video-salmonn: Speech-enhanced audio-visual large language models,

    G. Sun, W. Yu, C. Tang, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, Y . Wang, and C. Zhang, “video-salmonn: Speech-enhanced audio-visual large language models,” inProc. ICML, 2024

    2024

  3. [3]

    GAMA: A large audio-language model with advanced audio understanding and complex reasoning abilities,

    S. Ghosh, S. Kumar, A. Seth, C. K. R. Evuru, U. Tyagi, S. Sakshi, O. Nieto, R. Duraiswami, and D. Manocha, “GAMA: A large audio-language model with advanced audio understanding and complex reasoning abilities,” inProc. EMNLP. Miami, Florida, USA: ACL, 2024, pp. 6288–6313. [Online]. Available: https: //aclanthology.org/2024.emnlp-main.361/

    2024

  4. [4]

    Qwen3-omni technical report,

    J. Xu, Z. Guo, H. Hu, Y . Chu, X. Wang, J. He, Y . Wang, X. Shi, T. He, X. Zhu, Y . Lv, Y . Wang, D. Guo, H. Wang, L. Ma, P. Zhang, X. Zhang, H. Hao, Z. Guo, B. Yang, B. Zhang, Z. Ma, X. Wei, S. Bai, K. Chen, X. Liu, P. Wang, M. Yang, D. Liu, X. Ren, B. Zheng, R. Men, F. Zhou, B. Yu, J. Yang, L. Yu, J. Zhou, and J. Lin, “Qwen3-omni technical report,” 2025...

    Pith/arXiv arXiv 2025

  5. [5]

    Kimi-audio technical report,

    D. Ding, Z. Ju, Y . Leng, S. Liu, T. Liu, Z. Shang, K. Shen, W. Song, X. Tan, H. Tanget al., “Kimi-audio technical report,”arXiv preprint arXiv:2504.18425, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,

    S. Ghosh, A. Goel, J. Kim, S. Kumar, Z. Kong, S. gil Lee, C.-H. H. Yang, R. Duraiswami, D. Manocha, R. Valle, and B. Catanzaro, “Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2025. [Online]. Available: https://openreview.net/forum?id=FjByDpDVIO

    2025

  7. [7]

    Dynamic-superb: Towards a dynamic, collaborative, and comprehensive instruction-tuning benchmark for speech,

    C.-y. Huang, K.-H. Lu, S.-H. Wang, C.-Y . Hsiao, C.-Y . Kuan, H. Wu, S. Arora, K.-W. Chang, J. Shi, Y . Penget al., “Dynamic-superb: Towards a dynamic, collaborative, and comprehensive instruction-tuning benchmark for speech,” inProc. ICASSP. IEEE, 2024, pp. 12 136– 12 140

    2024

  8. [8]

    Air-bench: Benchmarking large audio-language models via generative comprehension,

    Q. Yang, J. Xu, W. Liu, Y . Chu, Z. Jiang, X. Zhou, Y . Leng, Y . Lv, Z. Zhao, C. Zhouet al., “Air-bench: Benchmarking large audio-language models via generative comprehension,” inProc. ACL. Bangkok, Thailand: ACL, Aug. 2024, pp. 1979–1998. [Online]. Available: https://aclanthology.org/2024.acl-long.109/

    2024

  9. [9]

    Audiobench: A universal benchmark for audio large language models,

    B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “Audiobench: A universal benchmark for audio large language models,” inProc. NAACL:HLT, Albuquerque, New Mexico, 2025, pp. 4297–4316

    2025

  10. [10]

    Dynamic-superb phase-2: A collaboratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,

    C.-y. Huang, W.-C. Chen, S.-w. Yang, A. T. Liu, C.-A. Li, Y .-X. Lin, W.- C. Tseng, A. Diwan, Y .-J. Shih, J. Shiet al., “Dynamic-superb phase-2: A collaboratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,” inProc. ICLR, 2025

    2025

  11. [11]

    MMAU: A massive multi- task audio understanding and reasoning benchmark,

    S. Sakshi, U. Tyagi, S. Kumar, A. Seth, R. Selvakumar, O. Nieto, R. Duraiswami, S. Ghosh, and D. Manocha, “MMAU: A massive multi- task audio understanding and reasoning benchmark,” inProc. ICLR, 2024

    2024

  12. [12]

    Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix,

    Z. Ma, Y . Ma, Y . Zhu, C. Yang, Y .-W. Chao, R. Xu, W. Chen, Y . Chen, Z. Chen, J. Conget al., “Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS) Datasets and Benchmarks Track, 2025

    2025

  13. [13]

    SAKURA: On the Multi- hop Reasoning of Large Audio-Language Models Based on Speech and Audio Information,

    C.-K. Yang, N. Ho, Y .-T. Piao, and H. yi Lee, “SAKURA: On the Multi- hop Reasoning of Large Audio-Language Models Based on Speech and Audio Information,” inInterspeech, 2025, pp. 1788–1792

    2025

  14. [14]

    MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark,

    D. Wang, J. Wu, J. Li, D. Yang, X. Chen, T. Zhang, and H. Meng, “MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark,”arXiv preprint arXiv:2506.04779, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Mmau-pro: A challenging and comprehensive benchmark for holistic evaluation of audio general intelligence,

    S. Kumar, ˇS. Sedl ´aˇcek, V . Lokegaonkar, F. L ´opez, W. Yu, N. Anand, H. Ryu, L. Chen, M. Pli ˇcka, M. Hlav ´aˇceket al., “Mmau-pro: A challenging and comprehensive benchmark for holistic evaluation of audio general intelligence,” inProc. AAAI Conf. Artif. Intell. (AAAI), vol. 40, no. 27, 2026, pp. 22 688–22 697

    2026

  16. [16]

    The torgo database of acoustic and articulatory speech from speakers with dysarthria,

    F. Rudzicz, A. K. Namasivayam, and T. Wolff, “The torgo database of acoustic and articulatory speech from speakers with dysarthria,”Lang. Resources Eval., vol. 46, no. 4, pp. 523–541, 2012

    2012

  17. [17]

    Dysarthric speech database for universal access research

    H. Kim, M. Hasegawa-Johnson, A. Perlman, J. R. Gunderson, T. S. Huang, K. L. Watkin, S. Frameet al., “Dysarthric speech database for universal access research.” inInterspeech, vol. 2008, 2008, pp. 1741– 1744

    2008

  18. [18]

    Globeaudio: A multilingual multicultural bench- mark for naturalistic evaluation of large audio-language models,

    R. Tan and W. Zhang, “Globeaudio: A multilingual multicultural bench- mark for naturalistic evaluation of large audio-language models,”arXiv preprint arXiv:2606.08194, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Fleurs-slu: A massively multilingual benchmark for spoken language understanding,

    F. D. Schmidt, I. Vuli ´c, G. Glava ˇs, and D. I. Adelani, “Fleurs-slu: A massively multilingual benchmark for spoken language understanding,” inProc. Conf. Lang. Model. (COLM), 2025

    2025

  20. [20]

    Iberobench: A benchmark for llm evaluation in iberian languages,

    I. Baucells, J. Aula-Blasco, I. de Dios-Flores, S. P. Su ´arez, N. Perez, A. Salles, S. S. Docio, J. Falc ˜ao, J. J. Saiz, R. Sep ´ulveda-Torreset al., “Iberobench: A benchmark for llm evaluation in iberian languages,” inProc. Int. Conf. Comput. Linguistics (COLING), 2025, pp. 10 491– 10 519

    2025

  21. [21]

    When audio-language models fail to leverage mul- timodal context for dysarthric speech recognition,

    P. Moure, N. Pokel, B. Bounajma, Y . Gao, R. Boehringer, L. Cheng, and S.-C. Liu, “When audio-language models fail to leverage mul- timodal context for dysarthric speech recognition,”arXiv preprint arXiv:2605.02782, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    Omnivoice: Towards omnilingual zero-shot text-to-speech with diffusion language models,

    H. Zhu, L. Ye, W. Kang, Z. Yao, L. Guo, F. Kuang, Z. Han, W. Zhuang, L. Lin, and D. Povey, “Omnivoice: Towards omnilingual zero-shot text-to-speech with diffusion language models,”arXiv preprint arXiv:2604.00688, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    V oxtral,

    A. H. Liu, A. Ehrenberg, A. Lo, C. Denoix, C. Barreau, G. Lample, J.-M. Delignon, K. R. Chandu, P. von Platen, P. R. Muddireddyet al., “V oxtral,”arXiv preprint arXiv:2507.13264, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” inProc. ICML. PMLR, 2023, pp. 28 492–28 518

    2023

  25. [25]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025