Pith. sign in

REVIEW 5 major objections 5 minor 35 references

MMedFD: A Real-world Healthcare Benchmark for Multi-turn Full-Duplex Automatic Speech Recognition

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read MMedFD is the first real-world Chinese healthcare speech corpus built for multi-turn full-duplex ASR, with synchronized views, role labels, and timing.

desk verdict A genuinely novel Chinese healthcare full-duplex ASR corpus, but the paper needs to fix inconsistent numbers and validate the role-separated metrics before I'd trust the benchmark. read the letter →

arxiv 2509.19817 v3 pith:PDUWCHDE submitted 2025-09-24 eess.AS

Hongzhao Chen , XiaoYang Wang , Jing Lan , Hexiao Ding , Yufeng Jiang , MingHui Yang , DanHui Xu , Jun Luo
show 4 more authors
Nga-Chun Ng Gerald W.Y. Cheng Yunlin Mao Jung Sun Yoo
This is my paper · ORCID
classification eess.AS
keywords automaticspeechrecognitionfull-duplexhealthcaredialoguemulti-turnspeakerdiarizationmedicalentityevaluationChinesecorpusbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces MMedFD, a new public corpus of real-world Chinese healthcare dialogues captured from a deployed full-duplex AI assistant, aiming to provide a benchmark where speaker overlap, interruptions, and barge-in are present rather than simulated. The dataset offers synchronized mixed-channel and user-only audio views, RTTM/CTM timing, role labels, and medical-entity annotations. The authors also propose a model-agnostic streaming pipeline for segmentation, speaker attribution, and context packaging, and demonstrate it by fine-tuning Whisper-small on role-concatenated audio, reporting WER, CER, and a new healthcare-entity-aware HC-WER metric. If the claims hold, MMedFD would become a public standard for measuring streaming ASR in Chinese medical settings and a foundation for end-to-end duplex dialogue agents.

What carries the argument

The central machinery is the corpus construction and evaluation pipeline. Each session preserves a complete mixed-channel conversation view plus a user-only view, with RTTM turn boundaries and CTM word alignments, role labels for user and agent, and medical entity annotations. A model-agnostic pipeline performs streaming segmentation (VAD), playback-aware diarization, residual-echo masking, and role-concatenated context packaging, enabling per-role ASR evaluation. The new metric HC-WER computes edit distance on medical-concept sequences (diseases, procedures, tests, medications), and LLM response quality is scored by rubric-based G-Eval and pairwise PairEval protocols.

What would settle it

Take a random sample of MMedFD sessions, have human annotators manually separate user and agent speech (or compare the released RTTM files with manual turn annotations), and compute the diarization error rate. If the user WER measured on the released 'user view' changes materially when recomputed on the manually separated user audio—or if the RTTM/user-view boundaries are frequently wrong—then the benchmark's core user metrics are contaminated and the diarization premise collapses.

Watch

Extended reading notes

Core claim

The paper presents MMedFD as the first real-world Chinese healthcare ASR corpus designed for multi-turn, full-duplex settings. Each session is captured from an in-production AI healthcare assistant that listens while speaking, preserving realistic acoustic conditions (AEC, noise suppression, gain control) and producing two synchronized views: a complete mixed conversation view and a user view that excludes non-user audio while retaining residual echo. The release includes RTTM turn boundaries, CTM word alignments, role labels for user and agent, and medical entity annotations. To make the corpus usable, the authors define a model-agnostic pipeline for streaming voice activity detection, play

Load-bearing premise

The benchmark's user-view audio and role labels come from an automatic playback-aware diarization step whose error rate is not reported, so if that diarization misattributes overlapping or echoed audio, the user-side WER and HC-WER numbers are not measuring the user's actual speech.

Editorial extensions

If this is right

  • MMedFD supplies a public, real-world testbed where full-duplex phenomena—barge-in, overlap, residual echo—are directly measurable, so streaming ASR systems can be compared on conditions that simulated datasets omit.
  • The model-agnostic pipeline standardizes how raw full-duplex audio is turned into role-separated, time-stamped ASR inputs, allowing different recognition models to be evaluated against the same preprocessed view.
  • HC-WER gives the healthcare ASR community an entity-aware metric that tracks errors on medical concepts instead of all words, making benchmark numbers more actionable for clinical applications.
  • The reported gap between agent (~1.8% WER) and user (~53% WER) speech quantifies how much harder real duplex user audio is, establishing a baseline for future work on overlap and echo robustness.
  • The LLM-judging protocol (G-Eval/PairEval) provides an end-to-end template for evaluating response quality in duplex healthcare agents, not just raw transcription.

Reading between the lines

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

  • Because the corpus includes synchronized mixed and user-only views plus RTTM timings, it can be reused without modification for speaker diarization evaluation, overlap detection, and turn-taking analysis—research the paper does not itself run.
  • The enormous user/agent WER gap suggests that most recognition error is caused by full-duplex artifacts rather than acoustic difficulty; a controlled re-reading experiment on the same transcripts would separate those factors and is a natural extension of the authors' protocol.
  • The category-specific HC-WER design is transportable: analogous entity-aware error metrics could be defined for other specialized domains (legal, customer service, eldercare) and other languages, giving the field a reusable evaluation pattern.
  • The reported session counts differ between the abstract (5,805) and the benchmark summary (1,805 dialogs); readers of any future extension should confirm which partition is officially released, since exact splits determine reproducibility.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces MMedFD, a benchmark for multi-turn full-duplex Chinese healthcare ASR, built from a deployed AI assistant's conversations. The authors describe a corpus with role labels, RTTM/CTM timing, and medical entity annotations; a preprocessing pipeline for filtering, VAD segmentation, playback-aware diarization, and quality control; and a fine-tuned Whisper-small ASR model evaluated with WER, CER, and a new HC-WER metric. They additionally evaluate LLM response quality with PairEval and G-Eval. The stated contributions are a public real-world benchmark, a model-agnostic full-duplex processing pipeline, and a multi-metric evaluation protocol. The central results are role-separated WER/CER/HC-WER numbers (Table 2) and LLM pairwise evaluations (Table 3). Two major obstacles prevent accepting the paper as is: the corpus statistics are mutually contradictory (5,805 vs. 1,805 dialogues; 39.04 h vs. 136.91 h), and the user-view diarization that underpins the headline metrics is unvalidated.

Significance. If the data and pipeline were fully released and verified, MMedFD would fill a real gap: there is no widely used public real-world Chinese healthcare ASR corpus with full-duplex interaction, synchronized TTS reference, role labels, and RTTM/CTM timing. The HC-WER concept-level metric is a reasonable extension of MC-WER and, if the extractor is made concrete and public, would be a useful evaluation instrument. The paper also deserves credit for attempting governance-first data curation (PII screening, manual verification, de-identification) and for making resources available on GitHub/HuggingFace. However, the current manuscript does not yet support the central claims. The contradictory corpus counts affect any downstream comparison, and the absence of any diarization accuracy or leakage measurement leaves the user-channel WER/HC-WER numbers uninterpretable. The LLM evaluation adds little, as nearly all pairwise outcomes are ties and the judge is one of the contestants. The significance will be high after these issues are resolved, but not as written.

major comments (5)
  1. [§3.2 vs. §4.1 and Table 1] The corpus statistics are internally contradictory. §3.2 states '5,805 dialogues, 4,926 turns and 39.04 h', with a 4,644/1,161 split. §4.1 and Table 1 state '1,805 dialogs and 10,814 turns' and '136.91 h'. These cannot both be correct; 4,926 turns for 5,805 dialogues is also structurally impossible. Since the benchmark's scale, split, and reported runtimes all depend on these numbers, the paper must present a single, verified set of statistics and explain the discrepancy. This is a load-bearing reproducibility issue.
  2. [§2.2 and §2.3, Table 2] The user-view diarization is unvalidated, which casts doubt on the central ASR results. §2.2 states that the user view 'excluded non-user segments while preserving residual echo' and that diarization was 'playback-aware' pyannote-style modeling. §2.3 then says TTS leakage from AEC failures 'was not considered user speech.' Yet no diarization error rate, no leakage measurement, and no human audit of RTTM boundaries are reported. If agent TTS frames leak into user segments, the user WER of 53.11% and HC-WER of 15.37% in Table 2 do not measure user speech; they measure a mixture. The paper needs to quantify diarization accuracy (e.g., DER, confusion matrix of role labels, fraction of user segments containing residual TTS) and, if needed, correct the user-view construction before these metrics can be trusted.
  3. [§3.3.1, Eq. (2)] HC-WER is defined in terms of a category-specific extractor π_k(·) that maps text to in-order canonical medical concept sequences, but the extractor is not specified in the paper or, apparently, released. Without the exact entity lexicon, normalization rules, and matching criterion, HC-WER cannot be reproduced or compared across systems. The phrase 'Adapting the Medical Concept WER idea' is also vague; a concrete definition or reference is needed. The paper should release the extractor code or unambiguously specify π_k for each of the four categories.
  4. [§3.3.2 and Table 3] The LLM evaluation results are effectively uninformative as reported. PairEval Tie% is 92.2–99.8% across all models, Win% is 49.0–52.3% with broad overlapping CIs, and G-Eval overall scores are 3.9–4.1 across all models. The claims of comparative response quality are therefore unsupported. Additionally, GPT-5 is both the judge and a contestant, so the evaluation is self-referential. The authors should either provide evidence that the pairwise judge is reliable (e.g., agreement with human judgments, separation on known-good/known-bad responses) or substantially temper/repurpose the response-quality evaluation.
  5. [§3.1–§3.3] The ASR evaluation lacks baselines. The paper reports metrics for a fine-tuned Whisper-small only, with no comparison to vanilla Whisper-small, Whisper-large, or any published medical ASR system. Without baselines, the reported WER/CER/HC-WER numbers do not establish the utility of MMedFD as a benchmark or the benefit of the proposed pipeline. At minimum, the paper should report the unadapted Whisper-small performance and, ideally, one or more strong reference systems, under the same normalization and evaluation protocol.
minor comments (5)
  1. [Title page/Abstract] Typo: 'AUTOMA TIC' in the title header. Also, the abstract's 5,805 sessions is inconsistent with §4.1's 1,805; the corrected number should be used consistently.
  2. [§2.2] The paragraph beginning 'Pre-processing integrates dialogue filtering...' is duplicated almost verbatim in the following paragraph. One copy should be removed.
  3. [Table 2] Column header mixes 'Hc-WER' with the text's 'HC-WER'. Unify the capitalization. Also clarify whether the 95% CIs are computed for WER/CER as well, or only for HC-WER.
  4. [§3.2] Training details are sparse: '1,000 epochs' at batch size 8 is unusual and likely a typo unless early stopping was used; please state the actual convergence criterion and number of steps. Also check formatting of '1×10 −4'.
  5. [§3.3.2] The evaluation protocol says 'consistent prompts and decoding parameters' but gives no example prompt or the exact decoding configuration. For reproducibility, at least one full prompt template and the LLM version/date should be included in an appendix or the GitHub repository.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: MMedFD is an empirical benchmark with standard held-out evaluation; flagged concerns are validation/reproducibility gaps, not input-output equivalences.

full rationale

This paper introduces a dataset and evaluates ASR/LLM systems empirically; it contains no formal derivation whose conclusion is equivalent to its premises. WER/CER/HC-WER are defined as standard edit-distance metrics against held-out reference transcripts. HC-WER explicitly adapts the prior MC-WER idea and uses a category extractor pi_k; the extractor is an annotation/parsing component, not a parameter fitted to the test predictions. Whisper-small is trained on a training split and evaluated on held-out dialogues, so the reported accuracy is not forced by construction. The self-referential aspects noted in the manuscript—GPT-5 serving as both a judged model and the judge (Sec. 3.3.2, 4.2), and the user-view diarization boundaries not being independently validated (Sec. 2.2-2.3)—are legitimate methodological and validation concerns, but they do not constitute circularity under the definition used here: no equation or fitted parameter is shown to equal its own output. The dataset-size inconsistency (5,805 sessions in the Abstract vs. 1,805 in Sec. 4.1) is likewise a reporting/reproducibility flaw, not a circular reduction. Therefore no circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim (a new benchmark corpus) rests on the data collection assumptions above; the main unstated choices are the diarization model's accuracy and the un-published details of the HC-WER extractor.

free parameters (3)
  • ASR confidence threshold for 'unclear' marking = 0.90
    Chosen by hand in Section 2.3 to mark segments with acoustic ambiguity.
  • VAD padding and short-gap merging parameters = not specified
    Section 2.2 uses Silero VAD with padding and merging, but exact values are not given; they affect segmentation boundaries.
  • HC-WER category extractor pi_k = rule-based, unspecified
    Section 3.3.1 defines per-category extractors mapping text to canonical medical concepts; the resources/rules behind these extractors are not described, so the metric is not fully reproducible.
assumptions (3)
  • domain assumption The deployed assistant's interactions are representative of real clinical full-duplex dialogue
    Section 2.1 collects during internal beta testing; the paper assumes this reflects deployment acoustics and turn-taking behavior.
  • domain assumption Preserved AEC/ANS/AGC processing leaves the recordings suitable for ASR benchmarking
    Section 2.1 keeps this processing 'to reflect deployment acoustics'; if it introduces artifacts, model performance is not comparable to clean corpora.
  • domain assumption LLM and human screening successfully removed all PII
    Section 2.2 gates PII-positive items and redacts; the benchmark release depends on this being complete.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MMedFD: A Real-world Healthcare Benchmark for Multi-turn Full-Duplex Automatic Speech Recognition." pith.science (2026). https://pith.science/paper/PDUWCHDE

@misc{pith2026250919817,
  author       = {Pith},
  title        = {Pith review of: MMedFD: A Real-world Healthcare Benchmark for Multi-turn Full-Duplex Automatic Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDUWCHDE}},
  note         = {Machine review of arXiv:2509.19817}
}
read the original abstract

Automatic speech recognition (ASR) in clinical dialogue demands robustness to full-duplex interaction, speaker overlap, and low-latency constraints, yet open benchmarks remain scarce. We present MMedFD, the first real-world Chinese healthcare ASR corpus designed for multi-turn, full-duplex settings. Captured from a deployed AI assistant, the dataset comprises 5,805 annotated sessions with synchronized user and mixed-channel views, RTTM/CTM timing, and role labels. We introduce a model-agnostic pipeline for streaming segmentation, speaker attribution, and dialogue memory, and fine-tune Whisper-small on role-concatenated audio for long-context recognition. ASR evaluation includes WER, CER, and HC-WER, which measures concept-level accuracy across healthcare settings. LLM-generated responses are assessed using rubric-based and pairwise protocols. MMedFD establishes a reproducible framework for benchmarking streaming ASR and end-to-end duplex agents in healthcare deployment. The dataset and related resources are publicly available at https://github.com/Kinetics-JOJO/MMedFD

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 5 linked inside Pith

  1. [1]

    INTRODUCTION Automatic speech recognition (ASR) converts speech into text streams and is increasingly important in clinical work [1]. High-quality transcripts improve documentation and support downstream tasks such as entity capture, concept *Co-first author #Corresponding author linking, and note generation that benefit patient care and auditing [2, 3]. ...

    arXiv 2025

  2. [2]

    We released a real-world Chinese healthcare dialogue corpus with multi-turn structure and consented de- identification, annotated with role labels and medical entities for supervised training and evaluation

  3. [3]

    We proposed a model-agnostic full-duplex pipeline that standardizes streaming segmentation, speaker diariza- tion, context packaging, and cross-turn memory, en- abling real-time barge-in and overlap handling

  4. [4]

    We conducted a comprehensive evaluation strategy for ASR, considering healthcare precision, entity-level precision recall, and metrics for coherence of dialogue and consistency of roles

  5. [5]

    Data Acquisition We collected speech from a full-duplex healthcare assistant during internal testing (beta version)

    METHODOLOGY 2.1. Data Acquisition We collected speech from a full-duplex healthcare assistant during internal testing (beta version). The assistant continu- ously listened while generating synthesized replies. To reflect deployment acoustics, we preserved acoustic echo cancella- tion (AEC), automatic noise suppression (ANS), and auto- matic gain control (...

  6. [6]

    Experimental Setups We fine-tuned Whisper-small [14] end-to-end on the Chinese training split for automatic speech recognition in healthcare dialogue

    EXPERIMENTS 3.1. Experimental Setups We fine-tuned Whisper-small [14] end-to-end on the Chinese training split for automatic speech recognition in healthcare dialogue. Role-concatenated audio streams were constructed by merging all user and agent turns separately, allowing long- context normalization and improved terminology coverage without requiring tur...

  7. [7]

    Benchmark Description MMedFD is a benchmark for Chinese healthcare spoken di- alogue constructed from live user–agent interactions under full-duplex conditions

    RESULTS 4.1. Benchmark Description MMedFD is a benchmark for Chinese healthcare spoken di- alogue constructed from live user–agent interactions under full-duplex conditions. The corpus contains 136.9 hours of role-labeled, turn-segmented transcripts across 1,805 dialogs and 10,814 turns. As shown in Table 1, most existing health- care corpora collected in...

  8. [8]

    LLM-judged results for healthcare queries using PairEval and G-Eval with a consistent GPT-5 judge

    CONCLUSION MMedFD offers a reproducible real-world benchmark for full-duplex Chinese healthcare dialogue, combining governance- compliant data, synchronized acoustic views, and grounded Table 3. LLM-judged results for healthcare queries using PairEval and G-Eval with a consistent GPT-5 judge. PairEval (Pairwise Judge) Model Win % (95% CI) Tie% Any-Failure...

Show all 35 references
  1. [9]

    P0051278, Jung Sun Yoo) and by the Research Grants Council of the Hong Kong Special Administrative Region, China (General Re- search Fund, Project No

    ACKNOWLEDGMENT This work was supported by an internal grant from The Hong Kong Polytechnic University (Project No. P0051278, Jung Sun Yoo) and by the Research Grants Council of the Hong Kong Special Administrative Region, China (General Re- search Fund, Project No. PolyU 15101...

  2. [10]

    The sound of healthcare: Improving medical transcription asr accuracy with large language models,

    Ayo Adedeji et al., “The sound of healthcare: Improving medical transcription asr accuracy with large language models,”arXiv preprint arXiv:2402.07658, 2024

    2024 arXiv

  3. [11]

    Medical dialogue system: A survey of cat- egories, methods, evaluation and challenges,

    X. Shi et al., “Medical dialogue system: A survey of cat- egories, methods, evaluation and challenges,” inFind- ings of the Association for Computational Linguistics (Findings of ACL), 2024, pp. 2840–2861

    2024

  4. [12]

    Multimed: Multilingual medi- cal speech recognition via attention encoder decoder,

    Khai Le-Duc et al., “Multimed: Multilingual medi- cal speech recognition via attention encoder decoder,” 2024

    2024

  5. [13]

    The AI doctor is in: A survey of task-oriented dialogue systems for healthcare appli- cations,

    M. Valizadeh et al., “The AI doctor is in: A survey of task-oriented dialogue systems for healthcare appli- cations,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 2022

    2022

  6. [14]

    A full-duplex speech dialogue scheme based on large language model,

    P. Wang et al., “A full-duplex speech dialogue scheme based on large language model,”Advances in Neural Information Processing Systems, vol. 37, pp. 13372– 13403, 2024

    2024

  7. [15]

    Primock57: A dataset of primary care mock consultations,

    Alex Papadopoulos Korfiatis et al., “Primock57: A dataset of primary care mock consultations,” inACL, Dublin, Ireland, 2022, pp. 588–598, Association for Computational Linguistics

    2022

  8. [16]

    Mtalk-bench: Evaluating speech-to- speech models in multi-turn dialogues via arena-style and rubrics protocols,

    Yuhao Du et al., “Mtalk-bench: Evaluating speech-to- speech models in multi-turn dialogues via arena-style and rubrics protocols,” 2025

    2025

  9. [17]

    Eberhard H ¨ansler and Gerhard Schmidt,Acoustic Echo and Noise Control: A Practical Approach, Wiley-IEEE Press, Hoboken, NJ, 2004

    2004

  10. [18]

    KWS15 keyword search evaluation plan,

    NIST, “KWS15 keyword search evaluation plan,” Tech. Rep., National Institute of Standards and Technology (NIST), 2015, Appendix C: RTTM File Format Spec- ification

    2015

  11. [19]

    The bigscience roots corpus: A 1.6tb composite multilingual dataset,

    Hugo Laurenc ¸on et al., “The bigscience roots corpus: A 1.6tb composite multilingual dataset,” inNeurIPS, 2022

    2022

  12. [20]

    Silero vad: Pre-trained enterprise- grade voice activity detector,

    Silero Team, “Silero vad: Pre-trained enterprise- grade voice activity detector,”https://github. com/snakers4/silero-vad, 2021

    2021

  13. [21]

    pyannote.audio: neural build- ing blocks for speaker diarization,

    Herv ´e Bredin et al., “pyannote.audio: neural build- ing blocks for speaker diarization,”arXiv preprint arXiv:1911.01255, 2019

    1911 arXiv

  14. [22]

    Openasr21 challenge evaluation plan,

    NIST, “Openasr21 challenge evaluation plan,” Tech. Rep., National Institute of Standards and Technology (NIST), 2021, Defines STM/CTM usage and scoring protocols

    2021

  15. [23]

    Robust speech recognition via large-scale weak supervision,

    Alec Radford et al., “Robust speech recognition via large-scale weak supervision,” inProceedings of the 40th International Conference on Machine Learning (ICML). 2023, vol. 202 ofProceedings of Machine Learning Research, pp. 28492–28518, PMLR

    2023

  16. [24]

    Paireval: Open-domain dialogue evaluation with pairwise comparison,

    ChaeHun Park et al., “Paireval: Open-domain dialogue evaluation with pairwise comparison,” inProceedings of the 2nd Conference on Language Modeling (COLM), 2024

    2024

  17. [25]

    G-eval: NLG evaluation using GPT- 4 with better human alignment,

    Yang Liu et al., “G-eval: NLG evaluation using GPT- 4 with better human alignment,” inProceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 2023

    2023

  18. [26]

    Vietmed: A dataset and benchmark for automatic speech recognition of vietnamese in the medical domain,

    Khai Le-Duc, “Vietmed: A dataset and benchmark for automatic speech recognition of vietnamese in the medical domain,” inProceedings of LREC-COLING 2024, Torino, Italia, 2024, pp. 17365–17370, ELRA and ICCL

    2024

  19. [27]

    A dataset of simulated patient- physician medical interviews with a focus on respiratory cases,

    Faiha Fareez et al., “A dataset of simulated patient- physician medical interviews with a focus on respiratory cases,”Scientific Data, vol. 9, no. 313, 2022

    2022

  20. [28]

    mymedicon: End-to-end burmese automatic speech recognition for medical conversa- tions,

    Hay Man Htun et al., “mymedicon: End-to-end burmese automatic speech recognition for medical conversa- tions,” inProceedings of LREC-COLING 2024, Torino, Italia, 2024, pp. 12032–12039, ELRA and ICCL

    2024

  21. [29]

    Afrispeech-200: Pan-african ac- cented speech dataset for clinical and general domain asr,

    Tobi Olatunji et al., “Afrispeech-200: Pan-african ac- cented speech dataset for clinical and general domain asr,”Transactions of the Association for Computational Linguistics, vol. 11, pp. 1669–1685, 2023

    2023

  22. [30]

    Spokenwoz: A large-scale speech-text benchmark for spoken task-oriented dia- logue agents,

    Shuzheng Si et al., “Spokenwoz: A large-scale speech-text benchmark for spoken task-oriented dia- logue agents,” 2023

    2023

  23. [31]

    V oxdialogue: Can spoken dialogue systems understand information beyond words?,

    Xize Cheng et al., “V oxdialogue: Can spoken dialogue systems understand information beyond words?,” in International Conference on Learning Representations (ICLR), 2025

    2025

  24. [32]

    Gpt-5 system card,

    OpenAI, “Gpt-5 system card,” Tech. Rep., OpenAI, Aug. 2025, Accessed 2025-09-17

    2025

  25. [33]

    Claude opus 4.1 system card addendum,

    Anthropic, “Claude opus 4.1 system card addendum,” Tech. Rep., Anthropic PBC, Aug. 2025, Accessed 2025- 09-17

    2025

  26. [34]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next-generation agentic capabilities,

    Gemini Team, “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next-generation agentic capabilities,”arXiv preprint arXiv:2507.06261, 2025

    2025 arXiv

  27. [35]

    Qwen3 technical report,

    An Yang et al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.