Pith. sign in

REVIEW 3 major objections 4 minor 46 references

This paper introduces TORUS, the first benchmark that tests whether a unified audio model can answer questions about audio it generated or edited, and finds that current models largely cannot.

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

2026-08-03 01:21 UTC pith:7KKHGGFX

load-bearing objection TORUS is the first audio-native self-coherence benchmark and the main finding is plausible, but the leak-check gate never tests the actual baseline judge, so the headline gap vs the cascade is not as clean as claimed. the 3 major comments →

arxiv 2607.28896 v1 pith:7KKHGGFX submitted 2026-07-30 cs.SD cs.AIcs.CL

TORUS: A Test of Rendering-Understanding Self-Coherence for Unified Audio Models

Aryan Vijay Bhosale , Harshit Rajgarhia , Abhishek Mukherji , Dinesh Manocha This is my paper
classification cs.SD cs.AIcs.CL
keywords self-coherenceunified audio modelsaudio generationaudio editingaudio understandingbenchmarkclosed-loop evaluationaudio reasoning
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.

TORUS is the first benchmark that tests unified audio models on the loop between their two heads: after the model generates or edits audio, it must answer questions about that same audio. The driving question is whether the generation head and understanding head agree about the same clip. Across 48 three-stage tests with 432 six-option questions, the best unified model answers 50.5% correctly, a cascade of specialized generation/editing/understanding systems scores 63.2%, and chance is 16.7%. The paper's central finding is that current audio models—unified and specialized alike—show limited self-coherence, with the largest failure appearing once editing is introduced. The authors argue that self-coherence is a distinct, essential evaluation axis that existing isolated benchmarks miss.

Core claim

The paper establishes that unified audio models, despite being able to generate, edit, and understand audio, largely fail to make sense of their own outputs. On TORUS, the best unified model answers 50.5% of questions, while a cascade of specialized systems answers 63.2%, with chance at 16.7%. The gap is driven by editing stages: every evaluated system degrades once editing begins, and counterfactual edits that require preserving one attribute while changing another are especially hard. The paper also shows that objective render quality—distributional distances, word-error rate, and modality correctness—does not predict coherence, so the deficit is not simply low-quality generation; it is a

What carries the argument

The central object is the self-coherence test: three stages (generation, edit, counterfactual edit), each with a render phase and an evaluate phase. Every stage produces audio with the generation head and asks three six-option multiple-choice questions with the understanding head. Questions have sealed answer keys that state more than the prompt, the prompt is withheld at answering time, and every question must pass a coupled leak-check gate: a muted-audio solver must not answer it correctly, and an ideal-generation solver must be able to answer it from a perfect specification. The coherence metric is simply the percentage of these questions answered correctly on the model's own generated or

Load-bearing premise

The load-bearing premise is that the leak-check gate actually removes all questions that can be answered without listening; that premise rests on the two solvers used in the gate being strong enough, and a stronger text-only model could answer some of the surviving questions without hearing the audio.

What would settle it

Feed the TORUS question bank to a frontier text-only LLM with no audio at all; if it answers a large share correctly—near or above the 50% level of current unified models—then the bank leaks and the measured self-coherence gap is partly an artifact of text priors rather than audio understanding.

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

If this is right

  • If TORUS is right, future unified audio models should be required to pass a self-coherence check, not just separate understanding, generation, and editing benchmarks.
  • Training and evaluation should treat the loop as a unit: improving generation fidelity alone will not raise coherence, since render quality and self-understanding are dissociated.
  • Audio editing is the clearest bottleneck; all evaluated systems degrade when edits are introduced, so editing and counterfactual-editing research should be prioritized.
  • A cascade of specialized models currently outperforms every unified model, so claims that shared representations automatically yield coherent behavior need empirical support.
  • Objective render metrics should be reported alongside coherence, since they do not predict it.

Where Pith is reading between the lines

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

  • The protocol could be extended to measure whether models improve at self-coherence over repeated attempts, turning TORUS into a test of self-correction rather than just a static snapshot.
  • As more unified models gain native audio-in editing, the paper's self-caption chain will become unnecessary; re-running TORUS then would separate captioning fidelity from editing fidelity and likely shift the numbers.
  • The same generate-edit-understand loop could be applied to video or embodied audio scenes, where temporal causality and scene-acoustics questions are even richer.
  • A testable prediction follows: models trained with self-coherence as an explicit objective will close the gap with specialized cascades faster than models trained only on task-specific losses.

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

3 major / 4 minor

Summary. TORUS introduces a benchmark for measuring self-coherence in unified audio models: 48 three-stage tests (generation, edit, counterfactual edit), producing 432 six-option multiple-choice questions across speech, sound, and music. Each test closes the loop by asking the model's understanding head to answer questions about audio produced by its own generation/editing head, with prompts withheld at answering time and keys sealed at construction. The paper evaluates five open unified models plus a Cascaded Baseline built from specialized generators, editors, and a frontier audio judge. The main empirical claim is that current unified models show limited self-coherence: the best unified model (Audex-30B) answers 50.5% of questions, the Cascaded Baseline answers 63.2%, and random guessing gives 16.7%, with the gap widest on editing stages. A secondary claim is that render quality (distributional metrics, WER, modality correctness) does not predict coherence, supported by a human A/B preference study.

Significance. If the measurement is sound, TORUS fills a genuine gap: it is the first audio-native self-coherence benchmark that chains a model's generation, editing, and understanding, uses sealed construction-time keys, withholds prompts at answering time, applies a coupled leak gate, and reports cluster-bootstrapped confidence intervals. The paper is transparent about several limitations and provides detailed inference settings, which supports reproducibility. The finding that unified models underperform a specialized cascade on self-coherence even when their render quality is competitive is a meaningful, falsifiable result for the field and could position self-coherence as a distinct evaluation axis alongside understanding and generation. However, the two load-bearing concerns below—leak certification limited to a single muted-audio solver and the self-caption proxy for editing—mean the current quantitative claims need additional controls before the headline numbers can be fully trusted.

major comments (3)
  1. [§3.2 (Coupled Gate); §5.1; Appendix A (Cascaded-baseline judge)] The leak-check gate certifies each question only against Audio-Flamingo-Next on one second of silence and gpt-5.6-luna on ideal specifications. It never applies the no-audio control to the models actually scored, in particular the Cascaded Baseline's gpt-audio judge (Appendix A) or the unified models' own understanding heads (e.g., Audio-Omni's Qwen-based stack). Since the headline result compares unified accuracy (50.5%) with the baseline's 63.2%, sharper text priors in gpt-audio than in Audio-Flamingo-Next would inflate the baseline and contaminate the central gap; the same applies to the unified heads, in unknown directions. Please run the muted-audio/no-audio control over all 432 items for every scored system and report both raw accuracy and the 'audio gain' (listening accuracy minus silence/no-audio accuracy). Without this, the 'leak-checked' claim and the cross-model comparison are
  2. [§4.3, Fig. 6, Table 2; §6] For the four non-native-edit arms, Stages 2 and 3 are implemented as a self-caption chain (understanding head captions the previous clip; generation head regenerates from that caption), not as audio-in editing. The paper acknowledges this in the Limitations ('may conflate captioning fidelity with editing fidelity'), but the conclusions 'all evaluated systems degrade once editing begins' and 'audio editing is the clearest open problem' (Sections 5.1 and 6) are applied to all evaluated systems. Only the Audio-Omni native arm actually edits audio. The Table 2 comparison thus mixes a native editing operation with a captioning-plus-generation proxy, so the stage-wise gap between the Cascaded Baseline (which uses real editors) and the self-caption unified models may reflect captioning fidelity rather than editing ability. Please label these arms as a proxy condition, report caption quality (e.
  3. [§3.2 and Appendix D (Coupled Gate repair loop)] The paper does not report how many questions were flagged by each gate, how many were repaired in Rounds 1 and 2, how many were admitted to the final bank despite remaining flagged, or how often human verification overrode a flag. The 'leak-checked' guarantee and the 432-question bank depend entirely on this filter; without these statistics, a reader cannot assess whether a material number of questions were answered correctly by the muted-audio solver or were judged unanswerable by the ideal-generation solver. Please add a gate-pass/flag/repair table per gate and per round, and either exclude still-flagged items from the scored bank or report their accuracy separately. This is necessary to certify the central construction claim.
minor comments (4)
  1. [§4.4, Table 2] The FAD/FD/KL metrics are computed against the Cascaded Baseline generated set (144 clips), not against ground-truth audios or the authored specifications. Please state explicitly in the caption and in Section 5.2 that these are reference-relative distances to the baseline's outputs, not absolute render-quality scores; otherwise the claim that 'render quality does not predict Coherence' can be read as measuring closeness to the baseline rather than fidelity to the prompt.
  2. [§3.2] The muted-audio solver is given one second of silence. Please clarify whether this is equivalent to a true no-audio condition (i.e., whether the model receives the same system prompt and question text with only silence as audio), and whether the gate also includes a condition with no audio input at all for models that allow it.
  3. [Appendix A] The Cascaded-baseline judge was selected by agreement with the authors' own answers on a 40-question sample. This is a reasonable sanity check, but the main text should note that this selection procedure can overfit to the authors' answer patterns and should ideally be complemented by agreement with independent human labelers.
  4. [Throughout] Typographical issues: 'propogate' (§2.2), 'intial' and 'initate' (§6), 'casacades' (§6), 'propogate' in Fig. 2 caption. The Table 1 legend uses check/cross symbols that should be defined directly in the caption rather than only in the per-cell evidence appendix.

Circularity Check

0 steps flagged

No significant circularity; TORUS is an external measurement with sealed keys and a disclosed, non-fitted leak-check.

full rationale

TORUS is a construction-and-measurement benchmark rather than a derived prediction: the 432 questions are human-authored, keys are fixed at construction time, the prompt is withheld during answering, and grading is exact-letter matching with no scoring model in the loop. The central result (best unified 50.5% vs cascade 63.2%) is a direct measurement, not a consequence of a fitted parameter or a self-citation chain. The leak-check gate rests on Audio-Flamingo-Next and gpt-5.6-luna as external solvers; even though Audio-Flamingo-Next shares a coauthor with this paper, it is an independently released, code-reproducible model, so its use as a no-audio probe is real evidence rather than authority. The disclosed choice of gpt-audio as the cascade judge by agreement with authors on a 40-question sample is a model-selection heuristic that could in principle bias the baseline, but it does not make the headline result equivalent to its inputs—the judge's full-test accuracy is not fixed by that selection and the paper openly reports the protocol. The main vulnerabilities (text-prior leakage through a stronger mute-audio solver, and the self-caption chain conflating captioning with editing) are external-validity/correctness caveats, not circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

The central measurement has no fitted free parameters: coherence is an exact letter-match percentage over six-option MCQs. The load-bearing design choices are the axioms above. The main risk is that benchmark validity is delegated to the coupled-gate solvers and to the self-caption proxy for editing.

axioms (6)
  • domain assumption Human-authored seeds and LLM-expanded specifications faithfully represent what the ideal audio should contain.
    Section 3.2: Test Architect and Question Author use the specifications to write prompts and questions; the ideal-generation solver and human verification rely on this.
  • domain assumption A muted-audio solver that fails a question is sufficient to certify the question cannot be answered from text priors alone.
    Section 3.2, Coupled Gate: only one muted-audio solver, Audio-Flamingo-Next, is used; the certification is contingent on that solver's blind spots.
  • domain assumption An LLM given ideal specifications is a valid oracle for whether a question is answerable from perfect audio.
    Section 3.2, Ideal Generation Solver: gpt-5.6-luna flags questions it cannot answer from the specs; no independent oracle is used.
  • domain assumption Exact letter-match accuracy on six-option MCQs is a valid operationalization of self-coherence between the two heads.
    Section 4.4 defines Coherence as % correct with exact letter match; this equates self-coherence with MCQ accuracy.
  • ad hoc to paper The self-caption chain can stand in for native audio editing in models without an audio-in edit head.
    Section 4.3 describes the decomposition; Section 6 admits it 'may conflate captioning fidelity with editing fidelity', which directly affects the editing-stage conclusions.
  • domain assumption gpt-audio is a valid proxy for human understanding in the Cascaded Baseline.
    Appendix A: gpt-audio was selected for 75.0% agreement with the authors' answers on a 40-question sample, against 47.5% for the alternative; no larger validation is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 15698 in / 11029 out tokens · 115508 ms · 2026-08-03T01:21:17.549958+00:00 · methodology

0 comments
read the original abstract

Unified audio models capable of audio understanding, audio generation and, increasingly, audio editing are proliferating rapidly. Yet a basic question about them remains unanswered: do the two heads of a unified model agree about the same audio? Current practice evaluates each capability in isolation on specialized benchmarks, and never asks whether a model can make sense of its own generations. We present TORUS, the first self-coherence test for audio-native unified models. TORUS comprises 48 three-stage self-coherence tests carrying 432 six-option questions spanning speech, sound and music across five task families. We holistically evaluate five open unified models alongside a Cascaded Baseline that combines state-of-the-art specialized generation, editing and understanding models. The best unified model answers 50.5% of questions against the Cascaded Baseline's 63.2% and a 16.7% chance floor. Models struggle on audio editing. Among the evaluated audio models (specialized and unified), we observe limited self-coherence, and thus position self-coherence as an essential test for future audio systems.

Figures

Figures reproduced from arXiv: 2607.28896 by Abhishek Mukherji, Aryan Vijay Bhosale, Dinesh Manocha, Harshit Rajgarhia.

Figure 1
Figure 1. Figure 1: Left: Self-evaluation and objective metrics of models. Larger radius = better, rim = perfect. Right: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmarking pipeline and evaluation strategy for unified audio models that are capable of audio [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The test-generation pipeline: human-authored seed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Blind pairwise A/B study of human preference [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Coherence by task family (left) and audio type (right). Whiskers: 95% cluster-bootstrap CIs over tests; [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Benchmarking pipeline for a self-coherence test in TORUS for unified models that lack native audio [PITH_FULL_IMAGE:figures/full_fig_p013_6.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

46 extracted references · 6 linked inside Pith

  1. [1]

    2025 , url =

    Sakshi, S and Tyagi, Utkarsh and Kumar, Sonal and Seth, Ashish and Selvakumar, Ramaneswaran and Nieto, Oriol and Duraiswami, Ramani and Ghosh, Sreyan and Manocha, Dinesh , booktitle =. 2025 , url =

    2025

  2. [2]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Kumar, Sonal and Sedl. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    2026

  3. [3]

    and Liu, Jian and Kong, Qiuqiang , howpublished =

    He, Haolin and Sun, Renhe and Dai, Zheqi and Du, Xingjian and Wu, Chunyat and Zhou, Jiayi and Li, Xiquan and Chen, Yun and Chen, Xie and Duan, Zhiyao and Wang, Weiqiang and Plumbley, Mark D. and Liu, Jian and Kong, Qiuqiang , howpublished =. Audio-Dependent Question Answering (. 2026 , url =

    2026

  4. [4]

    Proceedings of the Fourteenth International Conference on Learning Representations (ICLR) , year =

    Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models , author =. Proceedings of the Fourteenth International Conference on Learning Representations (ICLR) , year =

  5. [5]

    and Sakshi, S

    Ghosh, Sreyan and Seth, Ashish and Kumar, Sonal and Tyagi, Utkarsh and Evuru, Chandra Kiran and Ramaneswaran, S. and Sakshi, S. and Nieto, Oriol and Duraiswami, Ramani and Manocha, Dinesh , booktitle =. 2024 , url =

    2024

  6. [6]

    2025 , url =

    Ma, Ziyang and Ma, Yinghao and Zhu, Yanqiao and Yang, Chen and Chao, Yi-Wen and Xu, Ruiyang and Chen, Wenxi and Chen, Yuanzhe and Chen, Zhuo and Cong, Jian and Li, Kai and Li, Keliang and Li, Siyou and Li, Xinfeng and Li, Xiquan and Lian, Zheng and Liang, Yuzhe and Liu, Minghao and Niu, Zhikang and Wang, Tianrui and Wang, Yuping and Wang, Yuxuan and Wu, Y...

    2025

  7. [7]

    2026 , url =

    Wang, Dingdong and Li, Junan and Wu, Jincenzi and Yang, Dongchao and Chen, Xueyuan and Zhang, Tianhua and Meng, Helen , booktitle =. 2026 , url =

    2026

  8. [8]

    2026 , eprint =

    All That Glitters Is Not Audio: Rethinking Text Priors and Audio Reliance in Audio-Language Evaluation , author =. 2026 , eprint =

    2026

  9. [9]

    2026 , doi =

    Wang, Hui and Liu, Cheng and Chen, Junyang and Liu, Haoze and Jia, Yuhang and Zhao, Shiwan and Zhou, Jiaming and Sun, Haoqin and Bu, Hui and Qin, Yong , journal =. 2026 , doi =

    2026

  10. [10]

    2025 , doi =

    He, Yuhang and Jain, Yash and Liu, Xubo and Markham, Andrew and Vineet, Vibhav , booktitle =. 2025 , doi =

    2025

  11. [11]

    2606.01804 , archivePrefix =

    Zhang, Hanlin and Tan, Daxin and Tao, Dehua and Chen, Xiao and Tan, Haochen and Song, Linqi , year =. 2606.01804 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    2026 , eprint =

    Ma, Ziyang and Yan, Ruiqi and Xu, Ruiyang and Fang, Jie and Niu, Zhikang and Chao, Yi-Wen and Tu, Wenming and Wang, Tianrui and. 2026 , eprint =

    2026

  13. [13]

    2023 , url =

    Wang, Yuancheng and Ju, Zeqian and Tan, Xu and He, Lei and Wu, Zhizheng and Bian, Jiang and Zhao, Sheng , booktitle =. 2023 , url =

    2023

  14. [14]

    Zero-Shot Unsupervised and Text-Based Audio Editing Using

    Manor, Hila and Michaeli, Tomer , booktitle =. Zero-Shot Unsupervised and Text-Based Audio Editing Using. 2024 , url =

    2024

  15. [15]

    2026 , url =

    Xie, Wulin and Zhang, Yi-Fan and Fu, Chaoyou and Shi, Yang and Nie, Bingyan and Chen, Hongkai and Zhang, Zhang and Wang, Liang and Tan, Tieniu , booktitle =. 2026 , url =

    2026

  16. [16]

    2512.17196 , archivePrefix =

    Liu, Kai and Chen, Leyang and Li, Wenbo and Chen, Zhikai and Wang, Zhixin and Pei, Renjing and Kong, Linghe and Zhang, Yulun , year =. 2512.17196 , archivePrefix =

    arXiv

  17. [17]

    2505.10483 , archivePrefix =

    Li, Yi and Wang, Haonan and Zhang, Qixiang and Xiao, Boyu and Hu, Chenchang and Wang, Hualiang and Li, Xiaomeng , year =. 2505.10483 , archivePrefix =

    Pith/arXiv arXiv

  18. [18]

    2026 , eprint =

    Beyond Accuracy: Benchmarking Cross-Task Consistency in Unified Multimodal Models , author =. 2026 , eprint =

    2026

  19. [19]

    2601.14728 , archivePrefix =

    Kuan, Chun-Yi and Chang, Kai-Wei and Lee, Hung-yi , year =. 2601.14728 , archivePrefix =

    arXiv

  20. [20]

    2026 , eprint =

    Quantifying the Gap between Understanding and Generation within Unified Multimodal Models , author =. 2026 , eprint =

    2026

  21. [21]

    2602.04683 , archivePrefix =

    Yang, Dongchao and Wang, Yuanyuan and Chong, Dading and Liu, Songxiang and Wu, Xixin and Meng, Helen , year =. 2602.04683 , archivePrefix =

    arXiv

  22. [22]

    , booktitle =

    Yang, Dongchao and Tian, Jinchuan and Tan, Xu and Huang, Rongjie and Liu, Songxiang and Guo, Haohan and Chang, Xuankai and Shi, Jiatong and Zhao, Sheng and Bian, Jiang and Zhao, Zhou and Wu, Xixin and Meng, Helen M. , booktitle =. 2024 , url =

    2024

  23. [23]

    2026 , url =

    Tian, Jinchuan and Lee, Sang-gil and Kong, Zhifeng and Ghosh, Sreyan and Goel, Arushi and Yang, Chao-Han Huck and Dai, Wenliang and Liu, Zihan and Ye, Hanrong and Watanabe, Shinji and Shoeybi, Mohammad and Catanzaro, Bryan and Valle, Rafael and Ping, Wei , booktitle =. 2026 , url =

    2026

  24. [24]

    2026 , eprint =

    Unified Audio Intelligence Without Regressing on Text Intelligence , author =. 2026 , eprint =

    2026

  25. [25]

    2026 , url =

    Tian, Zeyue and Yang, Binxin and Liu, Zhaoyang and Zhang, Jiexuan and Yuan, Ruibin and Yin, Hubery and Chen, Qifeng and Li, Chen and Lyu, Jing and Xue, Wei and Guo, Yike , booktitle =. 2026 , url =

    2026

  26. [26]

    Unified-

    Lu, Jiasen and Clark, Christopher and Lee, Sangho and Zhang, Zichen and Khosla, Savya and Marten, Ryan and Hoiem, Derek and Kembhavi, Aniruddha , booktitle =. Unified-. 2024 , url =

    2024

  27. [27]

    2509.17765 , archivePrefix =

    Xu, Jin and Guo, Zhifang and Hu, Hangrui and Chu, Yunfei and Wang, Xiong and He, Jinzheng and Wang, Yuxuan and Shi, Xian and He, Ting and Zhu, Xinfa and others , year =. 2509.17765 , archivePrefix =

    Pith/arXiv arXiv

  28. [28]

    2503.20215 , archivePrefix =

    Xu, Jin and Guo, Zhifang and He, Jinzheng and Hu, Hangrui and He, Ting and Bai, Shuai and Chen, Keqin and Wang, Jialin and Fan, Yang and Dang, Kai and Zhang, Bin and Wang, Xiong and Chu, Yunfei and Lin, Junyang , year =. 2503.20215 , archivePrefix =

    Pith/arXiv arXiv

  29. [29]

    2412.06660 , archivePrefix =

    Liu, Shansong and Hussain, Atin Sakkeer and Wu, Qilong and Sun, Chenshuo and Shan, Ying , year =. 2412.06660 , archivePrefix =

    Pith/arXiv arXiv

  30. [30]

    2026 , url =

    Hung, Chia-Yu and Majumder, Navonil and Kong, Zhifeng and Mehrish, Ambuj and Bagherzadeh, Amir Ali and Li, Chuan and Valle, Rafael and Catanzaro, Bryan and Poria, Soujanya , booktitle =. 2026 , url =

    2026

  31. [31]

    2602.00744 , archivePrefix =

    Gong, Junmin and Song, Yulin and Zhao, Wenxiao and Wang, Sen and Xu, Shengyuan and Guo, Jing and Yang, Xuerui , year =. 2602.00744 , archivePrefix =

    arXiv

  32. [32]

    2506.00045 , archivePrefix =

    Gong, Junmin and Zhao, Sean and Wang, Sen and Xu, Shengyuan and Guo, Joe , year =. 2506.00045 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    2511.03601 , archivePrefix =

    Yan, Chao and Wu, Boyong and Yang, Peng and Tan, Pengfei and Hu, Guoqiang and Xie, Li and Zhang, Yuxin and Zhang, Xiangyu (Tony) and Tian, Fei and Yang, Xuerui and Zhang, Xiangyu and Jiang, Daxin and Zhou, Shuchang and Yu, Gang , year =. 2511.03601 , archivePrefix =

    arXiv

  34. [34]

    2512.20339 , archivePrefix =

    Tao, Ye and Wu, Wen and Zhang, Chao and Wu, Mengyue and Wang, Shuai and Xu, Xuenan , year =. 2512.20339 , archivePrefix =

    arXiv

  35. [35]

    2026 , eprint =

    Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music , author =. 2026 , eprint =

    2026

  36. [36]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Robust Speech Recognition via Large-Scale Weak Supervision , author =. Proceedings of the 40th International Conference on Machine Learning , series =. 2023 , url =

    2023

  37. [37]

    Kilgour, Kevin and Zuluaga, Mauricio and Roblek, Dominik and Sharifi, Matthew , booktitle =. Fr. 2019 , doi =

    2019

  38. [38]

    Hershey, Shawn and Chaudhuri, Sourish and Ellis, Daniel P. W. and Gemmeke, Jort F. and Jansen, Aren and Moore, R. Channing and Plakal, Manoj and Platt, Devin and Saurous, Rif A. and Seybold, Bryan and Slaney, Malcolm and Weiss, Ron J. and Wilson, Kevin W. , booktitle =. 2017 , doi =

    2017

  39. [39]

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages =

    Audio Set: An Ontology and Human-Labeled Dataset for Audio Events , author =. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages =. 2017 , doi =

    2017

  40. [40]

    Efficient Training of Audio Transformers with Patchout , author =. Proc. Interspeech 2022 , pages =. 2022 , doi =

    2022

  41. [41]

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

    Large-Scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation , author =. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

  42. [42]

    Advances in Neural Information Processing Systems 36 (NeurIPS) , year =

    Simple and Controllable Music Generation , author =. Advances in Neural Information Processing Systems 36 (NeurIPS) , year =

  43. [43]

    , booktitle =

    Liu, Haohe and Chen, Zehua and Yuan, Yi and Mei, Xinhao and Liu, Xubo and Mandic, Danilo and Wang, Wenwu and Plumbley, Mark D. , booktitle =. 2023 , url =

    2023

  44. [44]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Codec Does Matter: Exploring the Semantic Shortcoming of Codec for Audio Language Model , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    2025

  45. [45]

    and Zhang, Hao and Stoica, Ion , booktitle =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , booktitle =. Efficient Memory Management for Large Language Model Serving with. 2023 , doi =

    2023

  46. [46]

    2025 , url =

    Vaessen, Nik , howpublished =. 2025 , url =

    2025