Pith. sign in

REVIEW 4 major objections 6 minor 34 references

TriA Pipeline: A Large-Scale Automatic Audio Annotation Pipeline For Audio Classification In Specific Scenarios

T0 review · 4 major / 6 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Auto-annotated audio rivals hand-labeled data

desk verdict Engineering pipeline for auto-annotating audio events at scale; gains are real but modest and partially confounded by shared annotator-evaluator architecture. read the letter →

arxiv 2607.06179 v1 pith:JKF5IONT submitted 2026-07-07 eess.AS cs.AIcs.LG

Hong Lyu , Mingru Yang , Qianhua He , Yanxiong Li , Jinxin Huang , Zhengyu Pei This is my paper
classification eess.AScs.AIcs.LG
keywords automaticaudioannotationclassificationdatapipelinedomesticsoundeventdetectionBEATsweaklabelssequentialfine-tuning
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 proposes the TriA Pipeline, a four-stage system (Standardization, Audio Activity Detection, Audio Event Detection, Filtering) that converts raw audio from streaming platforms into annotated training data for audio classification. The key idea is to use an existing model (BEATs) to label audio events automatically, then filter for quality using aesthetic and similarity metrics. The authors use this pipeline to build a 2130-hour dataset covering 431 classes, then extract a prior-knowledge-guided subset (TriA_GK) matching three domestic audio classification tasks. They show that models first fine-tuned on TriA_GK data and then on manually annotated data gain an average of 3.97% in accuracy and 3.35% in Macro-F1 over models trained on manual data alone, and that in two of three tasks, training on TriA_GK alone already matches or exceeds manual-data performance. The paper is trying to establish that automatically annotated audio can serve as effective pre-training or augmentation for specific-domain classification tasks where labeled data is scarce.

What carries the argument

The pipeline's four stages: (1) Standardization converts heterogeneous audio to a uniform format; (2) Audio Activity Detection splits long recordings using energy thresholds informed by Event Critical Time and Silent Critical Time; (3) Audio Event Detection uses the BEATs model with local and global scanning windows to assign event labels; (4) Filtering applies audiobox-aesthetics (Production Complexity, Production Quality) and CLAP text-audio similarity thresholds to remove low-quality or mislabeled segments.

What would settle it

If BEATs systematically mislabels certain domestic sound classes (e.g., confusing similar sounds like 'electric shaver' and 'electric toothbrush'), those errors propagate into the TriA dataset and would degrade rather than improve downstream model performance when TriA_GK data is used.

Watch

Extended reading notes

Core claim

The central finding is that a fully automatic pipeline—using a pre-trained audio model for labeling and aesthetic/similarity models for filtering—can produce training data that, when applied through sequential fine-tuning (auto-labeled first, manual second), consistently improves classification performance over manual data alone across three domestic audio tasks. The gains are largest where manual datasets are smallest (Kitchen20: +6.09% accuracy), suggesting the pipeline is most valuable precisely in data-scarce domains.

Load-bearing premise

The pipeline's label quality depends entirely on the BEATs model's predictions being accurate enough to serve as ground truth. The only direct validation is a 93.67% subjective accuracy check on 100 samples, which may not represent the full 2130-hour dataset or capture systematic mislabeling of specific sound classes.

Editorial extensions

If this is right

  • The pipeline could be re-targeted to any domain where a pre-trained audio model can recognize relevant classes, extending annotated data to medical, industrial, or environmental monitoring audio without manual labeling.
  • The sequential fine-tuning protocol (auto-labeled then manual) could become a standard recipe for low-resource audio tasks, reducing the amount of expensive manual annotation needed.
  • The filtering thresholds create a quality-quantity tradeoff: stricter filtering improves per-sample quality but reduces class diversity and total duration, suggesting that task-specific threshold tuning is essential.
  • The approach could be extended to multilingual or cross-cultural audio scenarios by collecting from region-specific streaming platforms, though class-ontology alignment would be needed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes TriA Pipeline, a four-stage automatic audio annotation pipeline (Standardization, Audio Activity Detection, Audio Event Detection, Filtering) that converts raw audio from streaming platforms into annotated training data. The pipeline uses BEATs (AS-2M iter3+) for audio event detection and audiobox-aesthetics/CLAP for filtering. The authors construct a 2130-hour TriA dataset covering 431 classes and partition a prior-knowledge-guided subset (TriA_GK) for three domestic audio classification tasks (DESED AC, Kitchen20, Nonspeech7k). Sequential fine-tuning on TriA_GK followed by manual data yields average relative gains of 3.97% accuracy and 3.35% Macro-F1 over manual data alone. The pipeline and dataset are publicly released.

Significance. The paper addresses a genuine problem: annotated audio data scarcity in specific domains such as domestic environments. The pipeline is practical, processes 285 hours in ~10 hours (RTF 0.03), and the released dataset (2130 hours, 431 classes) is a useful resource. The prior-knowledge-guided subset construction via ECT/SCT thresholds is a reasonable design choice. The code and dataset are publicly released, which is commendable and strengthens reproducibility. The experimental comparison across three domestic AC tasks provides useful evidence. However, the significance is tempered by the BEATs-BEATs circularity concern (annotator and downstream classifier share the same architecture and pre-training) and the modest gains on two of three tasks.

major comments (4)
  1. §2.3 and §4.1.1: The pipeline uses BEATs (AS-2M iter3+) to annotate all training data, and the downstream classifier also uses BEATs iter3+ as its backbone. This creates a specific annotator-classifier alignment: the model being evaluated is the same model family that generated its training labels. While the test sets are human-annotated (mitigating evaluation circularity), the training-time circularity means gains could partly reflect BEATs-labeled data being distributionally easier for a BEATs classifier to learn from, rather than reflecting genuine annotation quality. A control experiment with a different backbone (e.g., PANNs, AST, or another audio encoder) would substantially strengthen the central claim. Without this control, the 3.97% average gain is difficult to attribute solely to data quality.
  2. Table 4, Nonspeech7k row: TriA_Nonspeech7k alone drops accuracy from 0.9448 to 0.8938, and the sequential gain is only 0.44% (accuracy) and 0.29% (F1). No significance testing or multiple-seed evaluation is reported. A 0.44% improvement on a single seed is within plausible noise. The paper should either report multiple seeds with standard deviations or perform a statistical significance test to confirm this gain is real, especially since it is one of three data points supporting the 'average 3.97%' claim.
  3. §4.1.1: For Kitchen20, only the classification head is fine-tuned, while for DESED AC and Nonspeech7k the entire BEATs model is fully fine-tuned. This makes the three tasks non-comparable: the 6.09% gain on Kitchen20 (which drives the 3.97% average) occurs under a fundamentally different fine-tuning regime. The paper should either justify why this difference in protocol is appropriate or report results under a unified fine-tuning strategy. Averaging gains across heterogeneous protocols inflates the apparent generality of the claim.
  4. §2.5: The subjective listening test uses only 100 samples to validate annotation accuracy across 431 classes. This means roughly 0.23 samples per class on average, which cannot represent the annotation quality of the full 2130-hour dataset. The 93.67% accuracy figure is the only direct validation of annotation quality. A stratified sampling approach covering more classes with multiple samples per class, or at least reporting per-class or per-scenario accuracy, would make the validation more credible.
minor comments (6)
  1. Table 1: The CLAP similarity column shows negative values for 'Annotated w/o Filtering' (min -4.19) but positive for filtered data. Clarify whether CLAP similarity can be negative and what the scale is, or confirm these are signed distances.
  2. §2.3: The statement 'The window length for AED local detection is required to exceed the maximum ECT (3 s)' is confusing — ECT was defined as 1.2 s in §2.2. Clarify whether 3 s is a different threshold or a typo, and where the 3 s figure comes from.
  3. Table 2: TriA's PC (5.13) is much higher than all manual datasets. This is attributed to the filtering stage, but it could also reflect distributional differences (streaming platform audio vs. recorded audio). A brief discussion of this confound would help interpretation.
  4. §3.2: The mapping between AudioSet ontology classes and task-specific classes (e.g., 'Electric shaver toothbrush' → two separate AudioSet classes) is mentioned but not fully specified. A supplementary table of class mappings would aid reproducibility.
  5. The paper states 'RTF of 0.03' for 285 hours in 10 hours. Clarify whether this RTF includes only inference or also the filtering stage (audiobox-aesthetics, CLAP).
  6. Figure 2 is referenced but appears to be a class distribution plot that is difficult to read in detail. Consider providing a table of the top-N classes by sample count or duration.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises four major concerns: (1) annotator-classifier circularity due to BEATs being used both for annotation and downstream evaluation, (2) lack of significance testing or multiple-seed evaluation for the Nonspeech7k results, (3) heterogeneous fine-tuning protocols across the three tasks undermining the averaged gain claim, and (4) insufficient sample size in the subjective listening validation. We agree that all four points identify genuine weaknesses that can and should be addressed in revision. We will conduct a control experiment with an alternative backbone (AST), add multiple-seed runs with significance tests, justify or unify the fine-tuning protocols, and expand the subjective validation with stratified sampling. We do not have standing objections to any of the referee's comments.

read point-by-point responses

  1. Referee: §2.3 and §4.1.1: The pipeline uses BEATs (AS-2M iter3+) to annotate all training data, and the downstream classifier also uses BEATs iter3+ as its backbone. This creates a specific annotator-classifier alignment... A control experiment with a different backbone (e.g., PANNs, AST, or another audio encoder) would substantially strengthen the central claim.

    Authors: The referee is correct that using BEATs as both the annotator and the downstream backbone creates a potential annotator-classifier alignment that complicates attribution of the observed gains. We acknowledge this as a genuine limitation of the current experimental design. While the test sets are human-annotated (so evaluation circularity is not an issue), the training-time circularity concern is valid: gains could partly reflect BEATs-labeled data being distributionally easier for a BEATs classifier to learn from, rather than reflecting genuine annotation quality. We will address this by adding a control experiment using AST (Audio Spectrogram Transformer) as an alternative downstream backbone, fine-tuned on the same TriA_GK subsets and manual data. If the sequential fine-tuning gains persist with a different backbone, this would confirm that the improvements are attributable to data quality rather than annotator-classifier alignment. We will report these results in the revised manuscript and discuss the circularity concern explicitly. revision: yes

  2. Referee: Table 4, Nonspeech7k row: TriA_Nonspeech7k alone drops accuracy from 0.9448 to 0.8938, and the sequential gain is only 0.44% (accuracy) and 0.29% (F1). No significance testing or multiple-seed evaluation is reported. A 0.44% improvement on a single seed is within plausible noise.

    Authors: The referee is correct. A 0.44% accuracy improvement from a single seed without significance testing is insufficient to establish that the gain is real rather than noise. We will address this by running all three tasks with at least three random seeds and reporting mean ± standard deviation for both accuracy and Macro-F1. We will also perform paired t-tests (or bootstrap confidence intervals where appropriate) to assess statistical significance of the sequential fine-tuning gains over manual-data-only baselines. This is especially important for the Nonspeech7k task, where the gain is smallest. We will update Table 4 with these results and revise the discussion accordingly, including a frank assessment of which gains are statistically significant. revision: yes

  3. Referee: §4.1.1: For Kitchen20, only the classification head is fine-tuned, while for DESED AC and Nonspeech7k the entire BEATs model is fully fine-tuned. This makes the three tasks non-comparable... Averaging gains across heterogeneous protocols inflates the apparent generality of the claim.

    Authors: The referee raises a valid point. The difference in fine-tuning regime across tasks means the averaged 3.97% gain conflates results from fundamentally different protocols, which inflates the apparent generality of the claim. The original rationale was that Kitchen20 has very limited training data (480 clips), and full fine-tuning of the 90M-parameter BEATs model on such a small dataset led to severe overfitting in preliminary experiments, so only the classification head was fine-tuned. However, we agree this should be made explicit and justified, and that the averaging across heterogeneous protocols is misleading. In the revision, we will: (1) add an explicit justification for the per-task fine-tuning choice in Section 4.1.1, (2) report full fine-tuning results for Kitchen20 as well (even if overfitting occurs, this provides a fairer comparison), and (3) revise the presentation of the average gain to either compute it only over tasks with the same protocol or clearly report per-task gains with protocol annotations so readers can assess generality without a potentially misleading average. revision: yes

  4. Referee: §2.5: The subjective listening test uses only 100 samples to validate annotation accuracy across 431 classes. This means roughly 0.23 samples per class on average, which cannot represent the annotation quality of the full 2130-hour dataset. The 93.67% accuracy figure is the only direct validation of annotation quality.

    Authors: The referee is correct that 100 samples across 431 classes (approximately 0.23 samples per class) is insufficient to represent annotation quality across the full dataset. We will expand the subjective validation in the revision. Specifically, we will adopt a stratified sampling approach: for each of the three downstream task scenarios (DESED, Kitchen20, Nonspeech7k), we will sample at least 10 clips per class with multiple annotators, and report per-class and per-scenario annotation accuracy. For the broader TriA dataset, we will sample from the most populous classes (e.g., top 50 classes by sample count) with at least 5 samples per class. This will provide a more credible and granular validation of annotation quality. We will report the expanded results in a revised Section 2.5 and include per-scenario breakdowns. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the pipeline's annotation model (BEATs) and downstream classifier (BEATs) share architecture, but the central claim is validated against externally human-annotated test sets, not against the annotator's own labels.

full rationale

The paper uses BEATs (AS-2M iter3+) both to auto-annotate the TriA training data (§2.3) and as the downstream classification backbone (§4.1.1). This creates a methodological concern—BEATs-labeled training data may be distributionally easier for a BEATs classifier to learn from—but it does not constitute circularity in the derivation-chain sense. The central claim (TriA_GK improves downstream AC performance by 3.97% accuracy / 3.35% Macro-F1) is evaluated on three independently human-annotated test sets: DESED_real test set, Kitchen20 fold-5, and Nonspeech7k test set (§4.1.2). The evaluation metric is computed against external ground truth, not against BEATs' own predictions. The annotator and evaluator sharing the same model family is a confound for interpreting gains (a correctness/external-validity concern), but the paper's claimed derivation chain—pipeline produces data → data is used to train model → model is tested on human labels—does not reduce to its inputs by construction. No equation or definition makes the prediction equivalent to a fit. No self-citation chain is load-bearing for the central result. The 93.67% subjective accuracy on 100 samples (§2.5) is a small-sample validation, not a circular definition. Score 1 reflects the shared-architecture concern without rising to a constructed reduction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The pipeline introduces no new entities (particles, forces, dimensions, etc.). It combines existing models (BEATs, CLAP, audiobox-aesthetics, auditok) into a sequential processing pipeline. The free parameters are operational thresholds determined through subjective tests or heuristic reasoning, not fitted to downstream task performance. The axioms are domain assumptions about the reliability of existing models as annotation tools, which is the central load-bearing premise of the work.

free parameters (7)
  • ECT (Event Critical Time) = 1.2 s
    Minimum duration for sound activity segments, determined through subjective listening tests on domestic scenario data (§2.2). Fitted to the target domain.
  • SCT (Silent Critical Time) = 2.0 s
    Maximum duration for silent segments, determined through subjective listening tests (§2.2). Fitted to the target domain.
  • AED local detection window length = 5 s
    Set to exceed maximum ECT (3 s) for complete event detection (§2.3). Chosen based on heuristic reasoning about event completeness.
  • AED local detection window shift = 3 s
    Window shift for local detection scanning (§2.3). No derivation provided.
  • AED confidence threshold = 0.6
    Event confidence threshold for BEATs detection (§2.3). Higher values increase reliability but reduce data volume; the specific value is not justified against a validation metric.
  • Filtering thresholds (Filtering1: PC, PQ, CLAP) = 1.8, 5.5, 5.0
    Thresholds for audiobox-aesthetics PC, PQ, and CLAP similarity (§2.4, Table 1). Applied to produce the TriA dataset.
  • Filtering thresholds (Filtering2: PC, PQ, CLAP) = 2.24, 5.85, 7.43
    Stricter thresholds used for comparison in mini-batch evaluation (Table 1). The relationship between Filtering1/Filtering2 and the final TriA dataset thresholds is not clearly specified.
assumptions (4)
  • domain assumption BEATs model (AS-2M fine-tuned iter3+) produces sufficiently accurate audio event annotations to serve as training labels
    The entire pipeline's annotation quality depends on this assumption (§2.3). Validated only by 100-sample subjective test with 93.67% accuracy.
  • domain assumption CLAP similarity between audio segments and text labels is a reliable proxy for annotation correctness
    Used as a filtering criterion in §2.4. The paper does not independently validate that high CLAP similarity correlates with annotation accuracy.
  • domain assumption audiobox-aesthetics PC and PQ scores are meaningful quality indicators for training data
    Used as filtering indicators in §2.4. The paper states these are 'relatively objective' but does not validate their correlation with downstream task performance.
  • domain assumption Audio collected from Bilibili and Douyin streaming platforms is representative of domestic acoustic environments
    The TriA dataset is built from streaming platform audio (§3.1) and evaluated on domestic AC tasks. The paper does not analyze potential domain mismatch between streaming platform audio and real domestic recordings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TriA Pipeline: A Large-Scale Automatic Audio Annotation Pipeline For Audio Classification In Specific Scenarios." pith.science (2026). https://pith.science/paper/JKF5IONT

@misc{pith2026260706179,
  author       = {Pith},
  title        = {Pith review of: TriA Pipeline: A Large-Scale Automatic Audio Annotation Pipeline For Audio Classification In Specific Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKF5IONT}},
  note         = {Machine review of arXiv:2607.06179}
}
abstract

There are some datasets of varying scales for audio classification (AC) applied to different tasks. However, annotated data is limited for most scenarios, such as domestic environments. To address this challenge, we propose an $\textbf{A}$utomatic $\textbf{A}$udio $\textbf{A}$nnotation Pipeline--TriA Pipeline, which can efficiently convert audio from various scenarios into high-quality training data with audio event annotations. A TriA dataset was constructed with the TriA Pipeline, over 2130 hours of audio covering 431 audio classes. Furthermore, we partitioned a prior-knowledge-guided subset (TriA$_{\mathrm{GK}}$) from TriA and conduct comparative experiments on three domestic AC tasks. Comparing the result on manually annotated data only and that on manually annotated data combines TriA$_{\mathrm{GK}}$, TriA$_{\mathrm{GK}}$ could achieve average relative gains of 3.97% in accuracy and 3.35% in Macro-F1, validating the effectiveness of TriA$_{\mathrm{GK}}$ and the TriA Pipeline.

Figures

Figures reproduced from arXiv: 2607.06179 by the authors.

Figure 1
Figure 1. Overview of the TriA Pipeline. AAD and AED denote Audio Activity Detection and Audio Event Detection. 2. TriA Pipeline This section details the TriA Pipeline and its evaluation on mini￾batch data. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Class distribution of TriA [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    TriA Pipeline: A Large-Scale Automatic Audio Annotation Pipeline For Audio Classification In Specific Scenarios

    Introduction Audio Classification (AC) enables the recognition of various en- vironmental sound events, serving as a core component in appli- cations ranging from multimedia content analysis [1] and audio captioning [2] to bio-acoustic monitoring [3, 4]. Ultimately, the efficacy of these AC systems heavily depends on the availability of large-scale audio ...

    work page Pith review arXiv 2026

  2. [2]

    As illustrated in Figure 1, the TriA Pipeline consists of four main stages: Standardization, Audio Activity Detection (AAD), Audio Event Detection (AED), and Filtering

    TriA Pipeline This section details the TriA Pipeline and its evaluation on mini- batch data. As illustrated in Figure 1, the TriA Pipeline consists of four main stages: Standardization, Audio Activity Detection (AAD), Audio Event Detection (AED), and Filtering. 2.1. Standardization This step follows the same procedure as Emilia-Pipe and aims to standardiz...

    work page

  3. [3]

    TriA Dataset 3.1. Statistics and Analysis Over 8706 hours of original audio data were collected from streaming platforms including Bilibili and Douyin, covering di- verse topics such as daily life, entertainment media, and tech- nology. After processing with the TriA Pipeline, the TriA dataset was constructed. The TriA dataset contains over 2130 hours of ...

    work page

  4. [4]

    For each task, three experiments are conducted

    Experiments This section evaluates the effectiveness of TriAGK for three do- mestic AC tasks: DESED AC, Kitchen20, and Nonspeech7k. For each task, three experiments are conducted. The baseline model is trained with different datasets, the first on the manu- ally annotated data, the second on the TriAGK, and the third on both, the TriAGK first and then the...

    work page

  5. [5]

    It efficiently converts audio collected from various streaming platforms into high- quality training data with event annotations

    Conclusion In this paper, we propose the TriA Pipeline, a large-scale automatic audio annotation pipeline. It efficiently converts audio collected from various streaming platforms into high- quality training data with event annotations. Based on the TriA Pipeline, the TriA dataset is constructed, which contains over 2130 hours of audio data covering 431 a...

    work page

  6. [6]

    Acknowledgments This work was partly supported by the national natural science foundation of China (62371195, 62111530145), and the ex- change project of the 10th Meeting of China-Croatia Science and Technology Cooperation Committee (10-34)

    work page

  7. [7]

    Generative AI Use Disclosure We used GPT-5.2 to assist in polishing the manuscript

    work page

  8. [8]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in2017 IEEE inter- national conference on acoustics, speech and signal processing (ICASSP), 2017, pp. 776–780

    work page 2017

Show all 34 references
  1. [9]

    Clotho: An audio cap- tioning dataset,

    K. Drossos, S. Lipping, and T. Virtanen, “Clotho: An audio cap- tioning dataset,” inICASSP 2020-2020 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 736–740

    2020

  2. [10]

    Bioa- coustics data analysis–a taxonomy, survey and open challenges,

    R. R. Kvsn, J. Montgomery, S. Garg, and M. Charleston, “Bioa- coustics data analysis–a taxonomy, survey and open challenges,” IEEE Access, vol. 8, pp. 57 684–57 708, 2020

    2020

  3. [11]

    Comparison of feature extraction methods for sound-based clas- sification of honey bee activity,

    A. Terenzi, N. Ortolani, I. Nolasco, E. Benetos, and S. Cecchi, “Comparison of feature extraction methods for sound-based clas- sification of honey bee activity,”IEEE/ACM transactions on au- dio, speech, and language processing, vol. 30, pp. 112–122, 2021

    2021

  4. [12]

    Fsd50k: an open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “Fsd50k: an open dataset of human-labeled sound events,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 30, pp. 829–852, 2021

    2021

  5. [13]

    ESC: Dataset for Environmental Sound Classi- fication,

    K. J. Piczak, “ESC: Dataset for Environmental Sound Classi- fication,” inProceedings of the 23rd Annual ACM Conference on Multimedia. ACM Press, 2015, pp. 1015–1018. [Online]. Available: http://dl.acm.org/citation.cfm?doid=2733373.2806390

    2015

  6. [14]

    Few-shot class-incremental audio classification using pseudo-incrementally trained embedding learner and continually updated stochastic classifier,

    Y . Li, W. Cao, J. Tan, Q. Li, and G. Chen, “Few-shot class-incremental audio classification using pseudo-incrementally trained embedding learner and continually updated stochastic classifier,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 3880–3895, 2025

    2025

  7. [15]

    Benchmark for kitchen20, a daily life dataset for audio-based human action recognition,

    M. Moreaux, M. G. Ortiz, I. Ferran ´e, and F. Lerasle, “Benchmark for kitchen20, a daily life dataset for audio-based human action recognition,” in2019 International Conference on Content-Based Multimedia Indexing (CBMI), 2019, pp. 1–6

    2019

  8. [16]

    Chime-home: A dataset for sound source recognition in a do- mestic environment,

    P. Foster, S. Sigtia, S. Krstulovic, J. Barker, and M. D. Plumbley, “Chime-home: A dataset for sound source recognition in a do- mestic environment,” in2015 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2015, pp. 1–5

    2015

  9. [17]

    Few-shot open-set audio classification using attention information-fused prototypes,

    Y . Li, J. Tan, Q. Li, G. Chen, S. Huang, and T. Virtanen, “Few-shot open-set audio classification using attention information-fused prototypes,”IEEE Transactions on Audio, Speech and Language Processing, vol. 34, pp. 1929–1943, 2026

    1929

  10. [18]

    Panns: Large-scale pretrained audio neural networks for audio pattern recognition,

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumb- ley, “Panns: Large-scale pretrained audio neural networks for audio pattern recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

    2020

  11. [19]

    Beats: audio pre-training with acoustic tok- enizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “Beats: audio pre-training with acoustic tok- enizers,” inProceedings of the 40th International Conference on Machine Learning, 2023, pp. 5178–5193

    2023

  12. [20]

    Atst: Audio representation learning with teacher-student transformer,

    X. LI and X. Li, “Atst: Audio representation learning with teacher-student transformer,” inProc. Interspeech 2022, 2022, pp. 4172–4176

    2022

  13. [21]

    Streaming audio transformers for online audio tagging,

    H. Dinkel, Z. Yan, Y . Wang, J. Zhang, Y . Wang, and B. Wang, “Streaming audio transformers for online audio tagging,” inProc. Interspeech 2024, 2024, pp. 1145–1149

    2024

  14. [22]

    Sound event detection in domestic environments with weakly labeled data and soundscape synthesis,

    N. Turpault, R. Serizel, A. Parag Shah, and J. Salamon, “Sound event detection in domestic environments with weakly labeled data and soundscape synthesis,” inWorkshop on Detection and Classification of Acoustic Scenes and Events, 2019

    2019

  15. [23]

    Domestic activities clustering from audio recordings using con- volutional capsule autoencoder network,

    Z. Lin, Y . Li, Z. Huang, W. Zhang, Y . Tan, Y . Chen, and Q. He, “Domestic activities clustering from audio recordings using con- volutional capsule autoencoder network,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2...

    2021

  16. [24]

    A Multi-Task Learning Frame- work for Sound Event Detection using High-level Acoustic Char- acteristics of Sounds,

    T. Khandelwal and R. K. Das, “A Multi-Task Learning Frame- work for Sound Event Detection using High-level Acoustic Char- acteristics of Sounds,” inInterspeech 2023, 2023, pp. 1214–1218

    2023

  17. [25]

    Leverag- ing audio-tagging assisted sound event detection using weakified strong labels and frequency dynamic convolutions,

    T. Khandelwal, R. K. Das, A. Koh, and E. S. Chng, “Leverag- ing audio-tagging assisted sound event detection using weakified strong labels and frequency dynamic convolutions,” in2023 IEEE Statistical Signal Processing Workshop (SSP), 2023, pp. 329–333

    2023

  18. [26]

    Htad: A home-tasks activities dataset with wrist-accelerometer and audio features,

    E. Garcia-Ceja, V . Thambawita, S. A. Hicks, D. Jha, P. Jakob- sen, H. L. Hammer, P. Halvorsen, and M. A. Riegler, “Htad: A home-tasks activities dataset with wrist-accelerometer and audio features,” inInternational Conference on Multimedia Modeling. Springer, 2021, pp. 196–205

    2021

  19. [27]

    The cirdo corpus: comprehensive audio/video database of domestic falls of elderly people,

    M. Vacher, S. Bouakaz, M.-E. Bobillier-Chaumon, F. Aman, R. A. Khan, S. Bekkadja, F. Portet, E. Guillou, S. Rossato, and B. Lecouteux, “The cirdo corpus: comprehensive audio/video database of domestic falls of elderly people,” in10th Interna- tional Conference on Language Reso...

    2016

  20. [28]

    Bi-modal multiperspective per- cussive (bimp) dataset for visual and audio human fall detection,

    J. Dibble and M. C. Bazzocchi, “Bi-modal multiperspective per- cussive (bimp) dataset for visual and audio human fall detection,” IEEE Access, 2025

    2025

  21. [29]

    Nonspeech7k dataset: Clas- sification and analysis of human non-speech sound,

    M. M. Rashid, G. Li, and C. Du, “Nonspeech7k dataset: Clas- sification and analysis of human non-speech sound,”IET Signal Processing, vol. 17, no. 6, p. e12233, 2023

    2023

  22. [30]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shiet al., “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” inIEEE Spoken Language Technology Workshop (SLT), 2024, pp. 885–890

    2024

  23. [31]

    Nvspeech: An integrated and scalable pipeline for human-like speech modeling with paralinguistic vocalizations,

    H. Liao, Q. Ni, Y . Wang, Y . Lu, H. Zhan, P. Xie, Q. Zhang, and Z. Wu, “Nvspeech: An integrated and scalable pipeline for human-like speech modeling with paralinguistic vocalizations,” arXiv preprint arXiv:2508.04195, 2025

    2025 arXiv

  24. [32]

    A scalable pipeline for enabling non-verbal speech generation and understanding,

    R. Ye, Y . Zhou, R. Yu, Z. Lin, K. Li, X. Li, X. Liu, G. Zeng, and Z. Wu, “A scalable pipeline for enabling non-verbal speech generation and understanding,”arXiv preprint arXiv:2508.05385, 2025

    2025

  25. [33]

    Meta audiobox aesthetics: Unified automatic qual- ity assessment for speech, music, and sound,

    A. Tjandra, Y .-C. Wu, B. Guo, J. Hoffman, B. Ellis, A. Vyas, B. Shi, S. Chen, M. Le, N. Zacharov, C. Wood, A. Lee, and W.-N. Hsu, “Meta audiobox aesthetics: Unified automatic qual- ity assessment for speech, music, and sound,”arXiv preprint arXiv:2502.05139, 2025

    2025 arXiv

  26. [34]

    Clap learning audio concepts from natural language supervision,

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

    2023

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.