Pith. sign in

REVIEW 5 major objections 6 minor 75 references

AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A chained pipeline of four foundation models re-annotates AudioSet, and every model trained on the refined labels scores higher mAP than on the original dataset.

desk verdict A useful relabeling recipe, but the evaluation loops through the same LLM pipeline that produced the labels—so the headline gains don't yet establish label quality. read the letter →

arxiv 2508.15429 v1 pith:7HNXHYBP submitted 2025-08-21 cs.SD

Yulin Sun , Qisheng Xu , Yi Su , Qian Zhu , Yong Dou , Xinwang Liu , Kele Xu This is my paper
classification cs.SD
keywords AudioSetlabelreannotationpromptchainingaudio-languagefoundationmodelsmulti-labelclassificationweaklearningontologyalignmentCLAP
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

AudioSet's weak, human-supplied labels are a known bottleneck in audio tagging, and this paper argues that label quality, not architecture, is what now limits the field's models. It proposes a fully automated reannotation chain: Qwen-Audio writes a structured description of each clip, Mistral proposes candidate tags, DeepSeek R1 maps them onto AudioSet's 527 classes, and CLAP deletes tags that do not match the audio. The resulting dataset, AudioSet-R, raises mAP across all four model families tested (AST, PANNs, SSAST, AudioMAE) when both training and evaluation use the refined labels, with the largest gain on SSAST (0.2528 to 0.2989). Evaluating on the relabeled validation set also lifts mAP for the transformer models even when training keeps original labels, which the authors read as evidence that the new validation labels are cleaner. If the gains reflect genuine label quality, AudioSet-R is a direct, scalable upgrade for the benchmark.

What carries the argument

Prompt chaining across four foundation models: each stage's output is folded into the next stage's fixed instruction template, decomposing 'tag this sound' into comprehension, prediction, alignment, and verification subtasks. Qwen-Audio performs audio comprehension via three-round prompting, producing a structured description; Mistral performs label synthesis under a count-control constraint; DeepSeek R1 performs ontology alignment by exact, fuzzy, and synonym matching against the 527 AudioSet categories; CLAP performs semantic filtering by cosine similarity between audio and text embeddings. The chain is the load-bearing mechanism: it makes label quality a function of model reasoning at eac

What would settle it

Conduct a human-listening audit of the relabeled validation set. Sample roughly 200 clips from AudioSet-R's validation split and ask independent annotators to choose whether the original AudioSet labels or the AudioSet-R labels describe each clip more accurately. If human preference for AudioSet-R is weak or absent, or if models trained on AudioSet-R lose their mAP advantage when evaluated on a fully human-verified test split, the reported gains are an artifact of circular labels — models trained and evaluated on the same pipeline's output — rather than improved label quality.

Watch

Extended reading notes

Core claim

The paper's central claim is that AudioSet's weak labels can be repaired by an automated multi-stage chain, making downstream classifiers better. Qwen-Audio writes a structured three-part summary of each clip; Mistral converts it into candidate tags; DeepSeek R1 aligns them to AudioSet's 527 categories; CLAP filters out tags whose audio-text similarity is low. Trained and evaluated on the resulting AudioSet-R, AST, CNN6, CNN14, SSAST, and AudioMAE all beat their original-AudioSet baselines, most sharply SSAST (0.2528 to 0.2989) and AST (0.0979 to 0.1310). The cross-condition — original training labels, relabeled validation — also lifts mAP for AST and SSAST, which the paper reads as evidence

Load-bearing premise

The paper treats higher mAP on the relabeled validation set as proof that the new labels are better, but that validation set was written by the same automated pipeline whose labels the models learn, so the gains could reflect models agreeing with the pipeline's own output rather than genuinely better labels — and the paper reports no human agreement check on the new validation labels.

Editorial extensions

If this is right

  • Models trained and evaluated on AudioSet-R beat the same architectures on original AudioSet, so the refined dataset can serve as a drop-in replacement for AudioSet-based training pipelines.
  • The relabeled validation set alone improves mAP for the transformer models even when training labels are unchanged, giving the field a sharper evaluation target without retraining.
  • Because the pipeline is fully automated and model-agnostic, the same chaining recipe can extend to AudioSet's unbalanced training subset or to revised versions of the ontology.
  • AudioSet-R's higher label density means downstream models receive more complete supervision, which should help with fine-grained and rare sound classes.
  • The framework applies to both supervised training from scratch and fine-tuning of self-supervised models, so the label-quality fix is orthogonal to the choice of learning paradigm.

Reading between the lines

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

  • The gains could partly be an artifact of circularity: the validation labels were written by the same automated pipeline whose labels the models learn, so an independent, human-labeled test split is the decisive check the paper does not report.
  • The cross-evaluation pattern is inconsistent across architectures — original-trained CNNs (CNN6, CNN14) lose mAP on the relabeled eval set while the transformers gain — suggesting the new labels are not uniformly 'better' for every model family.
  • The same decompose-then-align recipe (perception model → language model → ontology alignment) transfers to other weakly labeled multimodal benchmarks, such as video or bioacoustic tagging, wherever a fixed ontology constrains the labels.
  • The two tunable knobs — Mistral's label-count ceiling and the CLAP filtering threshold — are evaluated at single settings, so their sensitivity is untested; the reported gains could be fragile to those choices.
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 / 6 minor

Summary. The paper proposes AudioSet-R, a reannotated version of AudioSet's balanced training and evaluation subsets. Labels are generated by a three-stage pipeline: Qwen-Audio produces structured audio descriptions, Mistral predicts candidate labels, DeepSeek R1 aligns them to the 527 AudioSet classes, and CLAP filters candidates by audio-text similarity (Section 3.1-3.3). The authors evaluate four model families (AST, CNN6/CNN14, SSAST, AudioMAE) in supervised and fine-tuning settings. Their main quantitative claim, in Section 4.2 and Table 1, is that training and evaluating on AudioSet-R improves mAP relative to training and evaluating on the original AudioSet (e.g., SSAST 0.2528 → 0.2989, CNN14 0.2340 → 0.2690). The paper also claims that evaluating on the relabeled validation set improves mAP even when models are trained on original labels (e.g., AST 0.0979 → 0.1078), which is interpreted as evidence that the relabeled validation labels are of higher quality.

Significance. If the label-quality improvement were convincingly demonstrated, AudioSet-R would be a useful community resource: automatic reannotation with prompt-chained LLMs is scalable, the pipeline is modular, and the authors make code available. The three-stage decomposition and the use of CLAP as a consistency filter are reasonable first steps. However, the central evidence is weakened by an evaluation that uses validation labels produced by the same pipeline that creates the training labels, and by the absence of systematic human measurement of label accuracy. The paper also omits key pipeline hyperparameters and reports no variance over training seeds. Because these issues bear directly on the headline claim of 'higher label quality,' the contribution cannot be accepted in its present form, though the deficiencies are addressable.

major comments (5)
  1. [§4.2, Table 1; §3.3] The decisive evaluation is circular: the 'Relabel Eval' labels are produced by the same Qwen-Audio → Mistral → DeepSeek R1 + CLAP pipeline (Section 3.3) that generated the 'Relabel Train' labels. Higher mAP on this self-generated validation set may reflect agreement with the pipeline's own label distribution (including its vocabulary, label-count choices, and CLAP thresholds), not agreement with human ground truth. The paper's statement that relabeled validation has 'superior label quality' (Section 4.2) relies on the unstated assumption that pipeline output is ground truth. The manual check in Section 4.4 covers only a few clips in two categories and is not a systematic human evaluation. Please add human agreement statistics on a random sample of the relabeled validation set (e.g., per-label precision/recall against expert listening) or evaluate on an independent human-labeled benchmark
  2. [Table 1, AudioMAE row] If the relabeled validation set were uniformly higher quality, one would expect a consistent direction across models. The paper's own results contradict this: AudioMAE trained on original labels scores 0.3408 on Original Eval but only 0.3123 on Relabel Eval, while SSAST improves from 0.2528 to 0.2603 and AST from 0.0979 to 0.1078. This pattern does not support the claim that relabeled validation labels are simply better; it suggests the relabeled validation set changes the label distribution in model-dependent ways. Please provide per-class or per-label analysis to explain this inconsistency and temper the 'higher label quality' interpretation.
  3. [§3.1–§3.3] The dataset construction depends on several unreported hyperparameters: the CLAP similarity threshold in Eq. (1), the retry limit 'Qwen-try-num' in Section 3.1, the 'label count control number' in Section 3.2, and the Round 1 description length cap. These choices directly determine the final labels, so without their values the dataset is not reproducible and the stability of the reported improvements is unclear. Please report the exact values and include a sensitivity analysis, at least for the CLAP threshold and the label-count control, since those are likely to affect mAP most directly.
  4. [§4.2, Table 1] All table entries appear to be single runs. Several reported gains are small (e.g., CNN6 0.2530 → 0.2620, AST 0.0979 → 0.1078) and could be within training-seed noise. For a central claim of 'consistent and substantial performance improvements,' report at least three seeds with mean and standard deviation for each cell, or provide statistical tests over seeds. This is particularly important because the circular evaluation already makes the observed differences hard to interpret.
  5. [§4.3] The increase in average labels per clip is presented as evidence of 'semantic richness' and improved annotation density. However, more labels per clip is not by itself evidence of higher quality; if the LLM pipeline over-predicts, the relabeled set could have higher recall but lower precision. Since the prompt design includes an explicit label-count control, the reported density is partly a design choice rather than an independent quality measurement. Please reframe this analysis and pair it with per-label precision/recall against human labels.
minor comments (6)
  1. [Abstract] The phrase 'structured relabeled version of AudioSet-R' is awkward; it should be 'AudioSet-R, a structured relabeled version of AudioSet.'
  2. [§3.3, Eq. (1)] The text says 'f_a, f_b are the audio and text encoders,' but Eq. (1) uses f_t for the text encoder. Please fix the notation.
  3. [§4.1, Eq. (2)] AP_c is the area under the precision-recall curve; the text refers to p_c(r) as 'accuracy,' which is misleading. It should be precision.
  4. [§4.4] The manual evaluation is described as 'ground-truth annotations established through expert listening,' but only for a small, category-restricted sample from Human sounds and Musical instruments. Clarify how many clips were reviewed and how they were selected, or the reader cannot assess the claim.
  5. [References [35] and [36]] These two references appear to be the same 'Acoustic Prompt Tuning' paper (same authors, same venue, same pages) listed twice. Please deduplicate.
  6. [§3.3] DeepSeek R1 is described as a multimodal LLM. DeepSeek R1 is a text-based reasoning model; the cited DeepSeek-VL is the multimodal variant. Since Stage 3 only processes text labels, the description should be corrected.

Circularity Check

1 steps flagged · score 7.0 of 10

The central 'label quality' evidence is measured on a validation set relabeled by the same pipeline that produced the training labels, so reported mAP gains partly measure pipeline self-consistency rather than independently verified label quality.

  1. fitted input called prediction [Section 4.2, Table 1 ('Relabel Eval' columns) and Section 3.3 (label matching/filtering pipeline)]
    "When trained on the original training set but evaluated on the relabeled validation set, the AST model’s performance improved by roughly 1 percentage point, indicating the superior label quality of the relabeled validation set. ... we integrated CLAP [56] to evaluate semantic alignment. For each candidate label, we compute its CLAP similarity score with the audio and retain only those surpassing a predefined threshold."

    The relabeled validation set is produced by the same three-stage pipeline (Qwen-Audio → Mistral → DeepSeek R1 → CLAP) that generates the relabeled training set. The paper's central evidence of 'superior label quality' is mAP measured against these pipeline-generated validation labels. High mAP on 'Relabel Eval' can therefore reflect the trained model's success at reproducing the pipeline's own label distribution, including its biases in label count, vocabulary, and CLAP similarity thresholds, rather than agreement with an independent human ground truth. The paper provides no human-agreement statistics for the relabeled validation set; Section 4.4 offers only a few illustrative examples. The conclusion 'label quality is higher' is thus supported by comparing the system to its own output, wh

full rationale

The paper's main quantitative claim is that training and evaluating on AudioSet-R improves mAP across AST, PANNs, SSAST, and AudioMAE. However, the decisive improvements are on the 'Relabel Eval' column, whose labels were generated by the same automated pipeline that produced the relabeled training set. This makes the evaluation set non-independent: models trained on pipeline-generated labels are tested on pipeline-generated labels, so performance gains partly measure self-consistency of the labeling pipeline rather than label accuracy. The paper does not provide human agreement statistics on the relabeled validation set; the only manual check is a small case study in Section 4.4. There is no formal derivation that equates mAP on the relabeled validation set with human-verifiable label quality. While the paper's relabeling framework has independent content (the prompt-chaining design and the resulting dataset are concrete artifacts), the empirical validation of 'label reliability' is substantially circular because the yardstick is an output of the same system. Self-citations in the references are not load-bearing to this argument. Overall circularity score: 7.

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

The paper introduces no new physical or conceptual entities, but it relies on several unstated modeling choices: the CLAP threshold, retry counts, label-count targets, prompt templates, and exact model versions. None of these are given values, so the central result cannot be reproduced from the text alone. The biggest burden is the assumption that the pipeline's own validation labels are a valid benchmark.

free parameters (4)
  • CLAP similarity threshold
    Used in Section 3.3 to retain candidate labels; no value or selection procedure is reported.
  • Qwen-try-num retry limit
    Fault-tolerance retry count in Section 3.1; value not stated.
  • Label count control number
    Mistral prompt instructs the model to produce a specific number of labels based on corpus statistics (2.39 and 2.55 labels per clip), but the exact target or mechanism is not given.
  • Round 1 description length cap = 50 words
    Qwen-Audio is asked for a concise description up to 50 words in Section 3.1.
assumptions (4)
  • domain assumption Qwen-Audio's output descriptions are grounded enough in the audio to serve as ground truth for later labels.
    Entire Section 3.1 assumes the captioner's structured description is reliable; no human check of descriptions is reported.
  • domain assumption The CLAP audio-text similarity is a valid proxy for content-label semantic relevance.
    Threshold filtering in Section 3.3 uses CLAP score as the arbiter of whether a candidate label belongs to the audio.
  • ad hoc to paper Agreement with pipeline-generated validation labels measures label quality.
    Table 1's main comparisons evaluate on AudioSet-R validation labels produced by the same pipeline, so this assumption is doing the load-bearing work.
  • domain assumption AudioSet's original labels are reliable enough for the figures that use them (e.g., ground truth in Figure 5).
    Expert listening is claimed for the Figure 5 cases, but the overarching evaluation otherwise alternates between trusting original labels (when they worsen results) and distrusting them (when relabeling improves results).

how reviews work

0 comments
Cite this review

Pith. "Pith review of AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation." pith.science (2026). https://pith.science/paper/7HNXHYBP

@misc{pith2026250815429,
  author       = {Pith},
  title        = {Pith review of: AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7HNXHYBP}},
  note         = {Machine review of arXiv:2508.15429}
}
read the original abstract

AudioSet is a widely used benchmark in the audio research community and has significantly advanced various audio-related tasks. However, persistent issues with label accuracy and completeness remain critical bottlenecks that limit performance in downstream applications.To address the aforementioned challenges, we propose a three-stage reannotation framework that harnesses general-purpose audio-language foundation models to systematically improve the label quality of AudioSet. The framework employs a cross-modal prompting strategy, inspired by the concept of prompt chaining, wherein prompts are sequentially composed to execute subtasks (audio comprehension, label synthesis, and semantic alignment). Leveraging this framework, we construct a high-quality, structured relabeled version of AudioSet-R. Extensive experiments conducted on representative audio classification models--including AST, PANNs, SSAST, and AudioMAE--consistently demonstrate substantial performance improvements, thereby validating the generalizability and effectiveness of the proposed approach in enhancing label reliability.The code is publicly available at: https://github.com/colaudiolab/AudioSet-R.

Figures

Figures reproduced from arXiv: 2508.15429 by the authors.

Figure 1
Figure 1. Label Hierarchy Distribution within AudioSet’s [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustrates the proposed three-stage relabeling framework for AudioSet. The left part presents the overall concept of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of ground-truth and predicted labels [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The average number of labels in the dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The detailed analysis for three-round audio content extraction. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 57 canonical work pages

  1. [1]

    Jisheng Bai, Haohe Liu, Mou Wang, Dongyuan Shi, Wenwu Wang, Mark D Plumbley, Woon-Seng Gan, and Jianfeng Chen. 2024. AudioSetCaps: Enriched Audio Captioning Dataset Generation Using Large Audio Language Models. In Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, Vol. 1. IEEE, Vancouver, Canada, 11–24. https://openrev...

    work page 2024

  2. [2]

    Oriol Barbany, Michael Huang, Xinliang Zhu, and Arnab Dhua. 2024. Lever- aging large language models for multimodal search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . IEEE, Nashville,Tennessee,USA, 1201–1210

    work page 2024

  3. [4]

    Xiuli Bi, Shizhan Tang, Bin Xiao, Weisheng Li, Xinbo Gao, and Pietro Liò. 2025. A Systematic Review of Heart Sound Detection Algorithms: Experimental Results and Insights. IEEE Transactions on Instrumentation and Measurement 74 (2025), 1–16

    work page 2025

  4. [5]

    Pengfei Cai, Yan Song, Kang Li, Haoyu Song, and Ian McLoughlin. 2024. MAT-SED: A Masked Audio Transformer with Masked-Reconstruction Based Pre-training for Sound Event Detection. In Proc. Interspeech 2024 . Interspeech, Kos Island, Greece, 557–561

    work page 2024

  5. [6]

    Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. 2022. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058 1 (2022), 1–5

    arXiv 2022

  6. [7]

    Tianshui Chen, Jianman Lin, Zhijing Yang, Chumei Qing, Yukai Shi, and Liang Lin. 2025. Contrastive decoupled representation learning and regularization for speech-preserving facial expression manipulation. International Journal of Computer Vision 133, 7 (2025), 3822–3838

    work page 2025

  7. [8]

    Xiaxu Chen, Wei Li, Chunxu Liu, Chi Xie, Xiaoyan Hu, Chengqian Ma, Feng Zhu, and Rui Zhao. 2025. On the Suitability of Reinforcement Fine-Tuning to Visual Tasks. In Proceedings of the Computer Vision and Pattern Recognition Conference. cvpr, Nashville, Tennessee, 3357–3361

    work page 2025

  8. [9]

    Xie Chen, Xunying Liu, Yongqiang Wang, Mark JF Gales, and Philip C Woodland

    work page

Show all 75 references
  1. [10]

    Rajashekar Reddy Chinthalapani, Dhairya Jigar Shah, Nobel Ang, and Ambuj Varshney. 2025. AudioCast: Enabling Ubiquitous Connectivity for Embedded Systems through Audio-broadcasting Low-power Tags. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technolog...

    2025

  2. [11]

    Shabnam Choudhury, Yash Salunkhe, Sarthak Mehrotra, and Biplab Banerjee

  3. [12]

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models. arXiv preprint arXiv:2311.07919 1 (2023), 1

    2023 arXiv

  4. [13]

    Heinrich Dinkel, Zhiyong Yan, Yongqing Wang, Junbo Zhang, and Yujun Wang

  5. [14]

    Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang

  6. [15]

    Liang Gao, Kele Xu, Huaimin Wang, and Yuxing Peng. 2022. Multi-representation knowledge distillation for audio classification. Multimedia Tools and Applications 81, 4 (2022), 5089–5112

    2022

  7. [16]

    Gemmeke, Daniel P

    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. Au- dio Set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal...

    2017

  8. [17]

    Sreyan Ghosh, Sonal Kumar, Ashish Seth, Chandra Kiran Reddy Evuru, Utkarsh Tyagi, S Sakshi, Oriol Nieto, Ramani Duraiswami, and Dinesh Manocha. 2024. GAMA: A Large Audio-Language Model with Advanced Audio Understanding and Complex Reasoning Abilities. In EMNLP. Association for...

    2024

  9. [18]

    Xun Gong, Yu Wu, Jinyu Li, Shujie Liu, Rui Zhao, Xie Chen, and Yanmin Qian

  10. [19]

    Yuan Gong, Yu-An Chung, and James Glass. 2021. AST: Audio Spectrogram Transformer. In Proc. Interspeech 2021. interspeech2021, Brno, Czechia, 571–575. https://doi.org/10.21437/Interspeech.2021-698

    2021 doi

  11. [20]

    Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. 2022. Ssast: Self- supervised audio spectrogram transformer. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. IEEE, USA, 10699–10709

    2022

  12. [21]

    Tomasz Grzywalski and Dick Botteldooren. 2024. Automatic re-labeling of Google AudioSet for improved quality of learned features and pre-training. In2024 Signal Processing: Algorithms, Architectures, Arrangements, and Applications (SPA). IEEE, Poznan, Poland, 155–160. https://...

    2024

  13. [22]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 1 (2025), 1

    2025 arXiv

  14. [23]

    Wrya Anwar Hayder. 2025. Highlighting DeepSeek-R1: Architecture, Features and Future Implications. International Journal of Computer Science and Mobile Computing 14 (2025), 1–13

    2025

  15. [24]

    Shawn Hershey, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, Devin Platt, Rif A. Saurous, Bryan Seybold, Malcolm Slaney, Ron J. Weiss, and Kevin Wilson. 2017. CNN architec- tures for large-scale audio classification. In 2...

    2017

  16. [25]

    Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Woj- ciech Galuba, Florian Metze, and Christoph Feichtenhofer. 2022. Masked Autoencoders that Listen. In Advances in Neural Information Process- ing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho...

    2022

  17. [26]

    Jiahao Ji, Lixian Zhu, Haojie Zhang, Kun Qian, Kele Xu, Zikai Song, Bin Hu, Björn W Schuller, and Yoshiharu Yamamoto. 2024. Weight light, hear right: Heart sound classification with a low-complexity model. In 2024 32nd European Signal Processing Conference (EUSIPCO). IEEE, IEE...

    2024

  18. [27]

    Sampath Kini K, Siddhartha Bose, Babita Tyagi, Priyanka Sharma, and Amit Sharma. 2024. Prompt Engineering: An efficient way for Content Creation using Prompt Chaining. In 2024 IEEE 4th International Conference on ICT in Business Industry & Government (ICTBIG). IEEE, Indore, In...

    2024

  19. [28]

    Atharva Kale, Tanay Yaduka, Tauhid Shaikh, Gagandeep Kaur, Latika Pinjarkar, Poorva Agrawal, Rutuja Rajendra Patil, Suhashini Chaurasia, and Harmeet Kaur Khanuja. 2024. Unveiling the Power of AI Prompt Engineering: A Compre- hensive Exploration. In 2024 10th International Conf...

    2024

  20. [29]

    Arik Kershenbaum, Çağlar Akçay, Lakshmi Babu-Saheer, Alex Barnhill, Paul Best, Jules Cauzinille, Dena Clink, Angela Dassow, Emmanuel Dufourq, Jonathan Growcott, et al. 2025. Automatic detection for bioacoustic research: a practical guide from and for biologists and computer sc...

    2025

  21. [30]

    Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. 2020. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28 (2020), 2880–2894

    2020

  22. [31]

    Étienne Labbé, Thomas Pellegrini, and Julien Pinquier. 2024. Conette: An effi- cient audio captioning system leveraging multiple datasets with task embedding. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32 (2024), 3785–3794

    2024

  23. [32]

    Yuhsin Lai, Yicheng Hsu, Jiaxuan Fan, Jiahsin Lin, Hsuanyu Shih, Zongen Wu, Tingyu Liu, and MingSian Bai. 2024. A multichannel audio tagging and local- ization system for home surveillance. The Journal of the Acoustical Society of America 156, 4_Supplement (2024), A25–A25

    2024

  24. [33]

    Lanzendörfer, Constantin Pinkl, Nathanaël Perraudin, and Roger Watten- hofer

    Luca A. Lanzendörfer, Constantin Pinkl, Nathanaël Perraudin, and Roger Watten- hofer. 2025. Bootstrapping Language-Audio Pre-training for Music Captioning. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Indore, In...

    2025

  25. [34]

    Da Li, Boqing Zhu, Sen Yang, Kele Xu, Ming Yi, Yukai He, and Huaimin Wang

  26. [36]

    Plumbley, Huy Phan, and Em- manouil Benetos

    Jinhua Liang, Xubo Liu, Wenwu Wang, Mark D. Plumbley, Huy Phan, and Em- manouil Benetos. 2025. Acoustic Prompt Tuning: Empowering Large Language Models With Audition Capabilities. IEEE Transactions on Audio, Speech and Language Processing 33 (2025), 949–961. https://doi.org/10...

    2025 doi

  27. [37]

    Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. 2024. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525 1 (2024), 949–962

    2024 arXiv

  28. [38]

    Manjunath Mulimani and Annamaria Mesaros. 2024. Class-incremental learn- ing for multi-label audio classification. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, IEEE, Seoul, Korea, 916–920

    2024

  29. [39]

    Manjunath Mulimani and Annamaria Mesaros. 2025. Domain-Incremental Learn- ing for Audio Classification. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, IEEE, Hyderabad, India, 1–5

    2025

  30. [40]

    ACM Transactions on Asian and Low-Resource Language Information Processing 22, 11 (2023), 1–20

    Multi-task pre-training language model for semantic network completion. ACM Transactions on Asian and Low-Resource Language Information Processing 22, 11 (2023), 1–20

    2023

  31. [41]

    Aurian Quelennec, Pierre Chouteau, Geoffroy Peeters, and Slim Essid. 2025. Masked Latent Prediction and Classification for Self-Supervised Audio Represen- tation Learning. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IE...

    2025

  32. [42]

    Florian Schmid, Tobias Morocutti, Francesco Foscarin, Jan Schlüter, Paul Primus, and Gerhard Widmer. 2025. Effective pre-training of audio transformers for sound event detection. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICAS...

    2025

  33. [43]

    Ankit Shah, Fuyu Tang, Zelin Ye, Rita Singh, and Bhiksha Raj. 2024. Importance of negative sampling in weak label learning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, IEEE, Seoul, Korea, 7530–7534

    2024

  34. [44]

    Yanjie Sun, Kele Xu, Yong Dou, and Tian Gao. 2024. Self-Supervised Learning- Based General Fine-tuning Framework For Audio Classification and Event De- tection. In 2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, IEEE, Niagra Falls, Canada, 1–6

    2024

  35. [45]

    Yanjie Sun, Kele Xu, Chaorun Liu, Yong Dou, and Kun Qian. 2023. Automatic audio augmentation for requests sub-challenge. In Proceedings of the 31st ACM International Conference on Multimedia. Association for Computing Machinery, New York, NY, USA, 9482–9486

    2023

  36. [46]

    Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuo-Yiin Chang, and Tara Sainath. 2019. Deep Learning for Audio Signal Processing. IEEE Journal of Selected Topics in Signal Processing 13, 2 (2019), 206–219. https://doi.org/10.1109/ JSTSP.2019.2908700

    2019

  37. [47]

    Yash Suryawanshi, Vedanshi Shah, Shyam Randar, and Amit Joshi. 2025. Au- dio meets text: a loss-enhanced journey with manifold mixup and re-ranking. Knowledge and Information Systems 67, 3 (2025), 2195–2231

    2025

  38. [48]

    Hiren Thakkar and A Manimaran. 2023. Comprehensive Examination of Instruction-Based Language Models: A Comparative Analysis of Mistral-7B and Llama-2-7B. In 2023 International Conference on Emerging Research in Computa- tional Science (ICERCS). IEEE, Coimbatore, India, 1–6. ht...

    2023

  39. [49]

    Yuan Tseng, Layne Berry, Yi-Ting Chen, I-Hsiang Chiu, Hsuan-Hao Lin, Max Liu, Puyuan Peng, Yi-Jen Shih, Hung-Yu Wang, Haibin Wu, Po-Yao Huang, Chun-Mao Lai, Shang-Wen Li, David Harwath, Yu Tsao, Abdelrahman Mohamed, Chi-Luen Feng, and Hung-Yi Lee. 2024. AV-SUPERB: A Multi-Task...

    2024

  40. [50]

    Mingsheng Tu, Tianjiao Wan*, Qisheng Xu*, Xinhao Jiang, Kele Xu, and Cheng Yang. 2024. Higher-Order Vision-Language Alignment for Social Media Pre- diction. In Proceedings of the 32nd ACM International Conference on Multimedia. Association for Computing Machinery, New York, NY...

    2024

  41. [51]

    Ludovic Tuncay, Etienne Labbé, and Thomas Pellegrini. 2025. Hierarchical Label Propagation: A Model-Size-Dependent Performance Booster for AudioSet Tagging. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, IEEE, Hyder...

    2025

  42. [52]

    Yanjie Sun, Kele Xu, Chaorun Liu, Yong Dou, Huaimin Wang, Bo Ding, and Qinghua Pan. 2024. Automated data augmentation for audio classification. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32 (2024), 2716–2728

    2024

  43. [53]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. InProceedings of the 36th International Conference on Neural Information Processin...

    2022

  44. [54]

    Tongshuang Wu, Ellen Jiang, Aaron Donsbach, Jeff Gray, Alejandra Molina, Michael Terry, and Carrie J Cai. 2022. PromptChainer: Chaining Large Lan- guage Model Prompts through Visual Programming. In Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Sys...

    2022

  45. [55]

    Tongshuang Wu, Michael Terry, and Carrie Jun Cai. 2022. AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts. InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI ’22). Association for Computing Ma...

    2022

  46. [56]

    Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. 2023. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech an...

    2023

  47. [57]

    Kele Xu, Hengxing Cai, Xi Liu, Zhifeng Gao, and Bingbing Zhang. 2017. North atlantic right whale call detection with very deep convolutional neural networks. The Journal of the Acoustical Society of America 141, 5_Supplement (2017), 3944– 3945

    2017

  48. [58]

    Siyin Wang, Wenyi Yu, Yudong Yang, Changli Tang, Yixuan Li, Jimin Zhuang, Xianzhao Chen, Xiaohai Tian, Jun Zhang, Guangzhi Sun, et al. 2025. Enabling auditory large language models for automatic speech quality evaluation. In ICASSP 2025-2025 IEEE International Conference on Ac...

    2025

  49. [59]

    Kele Xu, Boqing Zhu, Qiuqiang Kong, Haibo Mi, Bo Ding, Dezhi Wang, and Huaimin Wang. 2019. General audio tagging with ensembling convolutional neural networks and statistical features. The Journal of the Acoustical Society of America 145, 6 (2019), EL521–EL527

    2019

  50. [60]

    Qisheng Xu, Jingfei Jiang, Kele Xu, Yong Dou, Caili Gao, Boqing Zhu, Kang You, and Qian Zhu. 2023. Self-supervised learning-for underwater acoustic signal classification with mixup. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17 (2023), 3530–3542

    2023

  51. [61]

    Kang You, Kele Xu, Ming Feng, and Boqing Zhu. 2022. Underwater acoustic classification using masked modeling-based swin transformer. The Journal of the Acoustical Society of America 152, 4_Supplement (2022), A296–A296

    2022

  52. [62]

    Kang You, Kele Xu, Boqing Zhu, Ming Feng, Dawei Feng, Bo Liu, Tian Gao, and Bo Ding. 2022. Masked modeling-based audio representation for ACM multimedia 2022 computational paralinguistics challenge. In Proceedings of the 30th ACM International Conference on Multimedia. Associa...

    2022

  53. [63]

    Khalid Zaman, Melike Sah, Cem Direkoglu, and Masashi Unoki. 2023. A Survey of Audio Classification Using Deep Learning. IEEE Access 11 (2023), 106620–106649. https://doi.org/10.1109/ACCESS.2023.3318015

    2023

  54. [64]

    Kele Xu, Qisheng Xu, Kang You, Boqing Zhu, Ming Feng, Dawei Feng, and Bo Liu

  55. [65]

    The Journal of the Acoustical Society of America 154, 1 (2023), 5–15

    Self-supervised learning–based underwater acoustical signal classification via mask modeling. The Journal of the Acoustical Society of America 154, 1 (2023), 5–15

    2023

  56. [66]

    Ziheng Zhou, Jinxing Zhou, Wei Qian, Shengeng Tang, Xiaojun Chang, and Dan Guo. 2025. Dense audio-visual event localization under cross-modal consistency and multi-temporal granularity collaboration. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, Vol. 39. ...

    2025

  57. [67]

    Boqing Zhu, Kele Xu, Qiuqiang Kong, Huaimin Wang, and Yuxing Peng. 2020. Audio tagging by cross filtering noisy labels. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28 (2020), 2073–2083

    2020

  58. [68]

    Ge Zhu, Jordan Darefsky, and Zhiyao Duan. 2024. Cacophony: An Improved Contrastive Audio-Text Model. IEEE/ACM Transactions on Audio, Speech, and Language Processing 32 (2024), 4867–4879. https://doi.org/10.1109/TASLP.2024. 3485170

    2024 doi

  59. [69]

    Qian Zhu, Qisheng Xu, Boqing Zhu, Zijian Gao, Lingbin Zeng, and Kele Xu

  60. [71]

    Jing-Xuan Zhang, Genshun Wan, Jianqing Gao, and Zhen-Hua Ling. 2025. Audio- visual representation learning via knowledge distillation from speech foundation models. Pattern Recognition 162 (2025), 111432

    2025

  61. [72]

    Zhenghao Zhou, Yongjie Liu, and Chen Cao. 2025. Advancing Audio-Based Text Generation with Imbalance Preference Optimization. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. AAAI Press, Philadelphia, Pennsylvania, 26120–26128

    2025

  62. [77]

    In ICASSP 2025-2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP)

    SSAST-Adapter: A Parameter-efficient Incremental Learning Algorithm for Underwater Acoustic Target Recognition. In ICASSP 2025-2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, IEEE, Hyderabad, India, 1–5

    2025

  63. [2016]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing 24, 11 (2016), 2146–2157

    Efficient training and evaluation of recurrent neural network language models for automatic speech recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 24, 11 (2016), 2146–2157

    2016

  64. [2022]

    In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Pseudo Strong Labels for Large Scale Weakly Supervised Audio Tagging. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Singapore, Singapore, 336–340. https://doi.org/10. 1109/ICASSP43922.2022.9746431

    2022

  65. [2023]

    In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    CLAP Learning Audio Concepts from Natural Language Supervision. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Rhodes Island, Greece, 1–5. https://doi.org/10.1109/ ICASSP49357.2023.10095889

    2023

  66. [2024]

    IEEE/ACM transactions on audio, speech, and language processing 32 (2024), 1803–1815

    Advanced long-content speech recognition with factorized neural trans- ducer. IEEE/ACM transactions on audio, speech, and language processing 32 (2024), 1803–1815

    2024

  67. [2025]

    InProceedings of the Computer Vision and Pattern Recognition Conference

    REJEPA: A Novel Joint-Embedding Predictive Architecture for Efficient Remote Sensing Image Retrieval. InProceedings of the Computer Vision and Pattern Recognition Conference. cvpr, Nashville, Tennessee, 2373–2382

Pith tools

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