Region-Specific Audio Tagging for Spatial Sound

Davide Berghi; Haohe Liu; Jinzheng Zhao; Junqi Zhao; Mark D. Plumbley; Qiuqiang Kong; Wenwu Wang; Xinyuan Qian; Yong Xu

REVIEW 3 major objections 6 minor 24 references

Region-Specific Audio Tagging for Spatial Sound

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper proposes region-specific audio tagging — labelling only the sound events inside a user-specified angular region or distance — and shows that six fixed 60° region-specific models outperform a single omnidirectional tagger (0.653 vs

desk verdict New task plus a benchmark, with a systematic feature study; the practical superiority claim rests only on SRSAT, and the real-data check is not a head-to-head. read the letter →

arxiv 2509.09526 v1 pith:XF7XLJHO submitted 2025-09-11 eess.AS cs.SD

Jinzheng Zhao , Yong Xu , Haohe Liu , Davide Berghi , Xinyuan Qian , Qiuqiang Kong , Junqi Zhao , Mark D. Plumbley

show 1 more author

Wenwu Wang

This is my paper

classification eess.AScs.SD

keywords region-specificaudiotaggingspatialmicrophonearraysoundeventdetectionPANNsdirectionalfeaturesSRSATbenchmark

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 a new audio tagging task: given a four-channel microphone recording and a query region (an angular range or a distance), identify only the sound events that originate inside that region. It builds a simulated dense benchmark, SRSAT, and extends three state-of-the-art taggers (PANNs, AST, PSLA) to the task by concatenating spectral, spatial, and positional features. On SRSAT, region-specific tagging with a 60° query reaches 0.473 mAP, and running six fixed 60° region-specific models and max-aggregating their outputs gives 0.653 mAP, substantially above the 0.371 mAP of an omnidirectional tagger. The paper also shows the same approach works for distance queries and finds that PANNs transfers most effectively. A sympathetic reader would take away that spatial region is a useful conditioning signal for dense sound event tagging.

What carries the argument

The region of interest is encoded as a positional feature fused with spectral and spatial features at the model input. For angular queries the paper uses the Field-of-View (FOV) feature — the maximum over directional features D(θ)=Σ_n cos(IPD_n − P_n(θ)) across angles inside the region, with −1 when the outside-view maximum exceeds the inside-view maximum — and for distance queries a learned embedding of the distance value. These positional features are concatenated on the channel dimension with log-power spectrum and inter-channel phase differences and fed to a CNN-14 PANNs backbone, letting the tagger learn semantic and spatial information simultaneously rather than cascading a beamformer

What would settle it

Train the proposed region-specific PANNs on SRSAT and evaluate on real dense spatial audio (e.g., a new recording or a re-annotated subset of STARSS23 with at least 4 simultaneous events and known positions). If six fixed 60° region-specific taggers do not exceed the omnidirectional baseline's mAP on that real data, the paper's central empirical claim is refuted.

Watch

Extended reading notes

Core claim

Audio tagging can be conditioned on the spatial region being queried. The authors define region-specific audio tagging, where a tetrahedral microphone array recording is paired with an angular range [θbegin, θend] or a distance d, and the model must tag only sound events inside that region. On their simulated SRSAT dataset, a PANNs model using log-power spectrum, inter-channel phase difference, and a directional feature achieves 0.473 mAP for a 60° query region; when six such fixed 60° region models are combined by max-aggregation, the system reaches 0.653 mAP versus 0.371 mAP for an omnidirectional tagger on the same test set. With oracle knowledge of event locations, the location-aware sys

Load-bearing premise

The SRSAT simulation (random rooms, SpatialScaper rendering, 25 events per clip) represents real spatial audio closely enough for the feature and model conclusions to transfer to physical recordings.

Editorial extensions

If this is right

Region-specific tagging lets surveillance and safety systems attend to one spatial region without a separate beamforming front end.
Running a tagger per fixed region and max-aggregating outputs improves omnidirectional tagging over a single whole-space tagger, making region division a useful inductive bias for dense sound scenes.
The FOV feature acts as a soft spatial mask; replacing it with a directional feature at the region's center preserves most of the gain at 1/72nd of the FOV computation cost.
Distance-queried tagging works but is harder than azimuth-queried tagging, indicating that learned distance embeddings are weaker positional conditioning than statistics-derived directional features.
AudioSet-pretrained PANNs transfers to the new task better than from-scratch AST or PSLA, so pretraining remains the recommended starting point.

Reading between the lines

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

Inference: The fixed-region result suggests a general recipe — partitioning azimuth into a small set of regions and max-aggregating tagger outputs can subsume a single global tagger — which should transfer to other array geometries if directional features are computed for the corresponding microphone pairs.
Inference: The FOV feature's hard −1 gate could be softened (e.g., log-sum-exp over directional features) to better handle reverberation and reflections; this is a testable modification on the SRSAT benchmark.
Inference: The real-data validation is limited to one model and no region-specific baselines, so the central claim that region-specific tagging beats omnidirectional tagging on dense real scenes remains to be confirmed; a real dense spatial-audio dataset with per-region labels would settle it.
Inference: The SRSAT simulation's high event density (mean 25 per 60-second clip) is what makes region-specific tagging matter — the same PANNs model gets 0.938 mAP on STARSS23 but 0.473 on SRSAT — suggesting the task's value grows with scene crowding.

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.

Desk Editor's Note

read the letter

The paper proposes region-specific audio tagging as a new task: given a four-channel recording and an angular range or distance, predict which sound events are in that region. That framing is genuinely new relative to the region-specific ASR/separation work, and the SRSAT simulated benchmark is a useful resource, especially since code and data are promised. The feature study is also reasonably thorough: LPS/IPD/DF, FOV, learned angle features, and distance features are all tried, with PANNs beating PSLA and AST in their setup. Credit where due: the authors do not oversell the method as a new architecture; it is mostly a feature-conditioned extension of existing models, which is fine for a first benchmark paper.

The soft spots are real. Every number in Tables 1–6 is a single run with no error bars or significance tests, so the headline difference in Table 3 (0.653 fixed-region vs 0.371 omnidirectional) could shift with seeds, though the gap is large enough that I doubt it would vanish. The bigger issue is that this superiority claim is only demonstrated on SRSAT, which the authors themselves generated with SpatialScaper. The STARSS23 results in Table 5 are not a comparison between region-specific and omnidirectional systems; they just show that PANNs gets high mAP on a mostly-single-event dataset. So the stress-test concern holds: there is no real-data head-to-head, and the transfer could be worse if simulated room geometry and event densities are kinder to directional features than actual recordings.

The distance-query experiment in Section 4.6 uses the ground-truth distance as the conditioning signal. That is an oracle setting, so the numbers are an upper bound, not what a user would get from a real system. The same is true for the location-aware system in Table 3, which uses known source positions; the authors do acknowledge this, but the abstract's "effectiveness" claim leans on it a bit hard.

Minor points: a cascaded beamformer-plus-PANNs baseline is mentioned as a contrast but never evaluated, which would have been a useful reference point. The FOV feature is best but set aside for computational cost; that is a reasonable engineering call, but it does mean the best-performing method is not the one used in later tables.

Overall, the central idea is sound and the benchmark is likely to be cited. The paper deserves peer review, but the revision should add repeated runs, a real-data region-specific versus omnidirectional comparison, and a non-oracle distance setting. As it stands, it is a promising feasibility study, not a fully established superiority claim.

Referee Report

3 major / 6 minor

Summary. The paper defines region-specific audio tagging: given a tetrahedral microphone array recording and a query region (angular range or distance), the system outputs tags for sound events only in that region. It introduces SRSAT, a simulated benchmark generated with SpatialScaper, and adapts PANNs, PSLA, and AST by concatenating spectral (LPS), spatial (IPD/GCCPHAT), and positional (directional, FOV, learned angle/distance) features. Experiments on SRSAT compare features and models; fixed-region and location-aware aggregation are compared with omnidirectional tagging; distance-query and STARSS23 validation are reported. The central claims are that this is the first region-specific audio tagging task/benchmark and that fixed-region/location-aware systems outperform omnidirectional tagging.

Significance. If validated, the paper opens a useful new task and provides a public benchmark and code. The synthetic SRSAT dataset with dense, overlapping events is a reasonable stress test, and the systematic feature/model comparison is informative. The paper is explicit that location-aware and distance-query settings use ground-truth positions/distances (oracle conditions), which avoids circularity. However, the main superiority claim currently rests on a single simulated dataset, and all results are single-run, so the empirical support is not yet at the level needed for a strong general claim.

major comments (3)

[§4.3, §4.5, Tables 3 and 5] The central claim that fixed-region/location-aware systems are superior to omnidirectional tagging rests only on SRSAT, a simulated dataset generated with SpatialScaper (§3.1). The STARSS23 validation in §4.5/Table 5 reports a single PANNs mAP (0.938 vs 0.473) and does not specify whether this is the region-specific or omnidirectional model, nor does it include a fixed-region vs omnidirectional comparison, a beamformer+PANNs baseline, or repeated runs. Because SRSAT uses simulated rooms, known source positions, and a mean of 2.2 simultaneous events (Table 5), the gains in Table 3 (0.653 vs 0.371 mAP) may not transfer to real array recordings. A fair real-data head-to-head on STARSS23 (or another real dataset) with the same protocol is needed to support the practical superiority claim.
[Tables 1–6] All mAP/EER numbers are single runs without error bars, confidence intervals, or significance tests. In Table 1 the differences among FOV (0.485), DF (0.479), and GCCPHAT with DF (0.479) are small (≤0.006); similarly Table 3's OD variants range 0.370–0.376. Without multiple seeds or a significance test, the feature and model rankings cannot be distinguished from noise. Please report mean±std over at least 3 seeds and, where rankings matter, a paired test.
[§4.6 and §4.3 (oracle conditions)] The distance-query experiment (§4.6) randomly selects a sound event and uses its ground-truth distance as the query; the location-aware system (§4.3) uses ground-truth positions and filters overlapping regions. These are oracle upper bounds, not deployable settings, and should be labeled as such in the abstract/conclusion. They demonstrate feasibility and headroom but do not by themselves establish that a practical system can estimate region queries from the array signal. Please add a non-oracle variant (e.g., estimated distance/azimuth from a localizer) or explicitly delimit the claim.

minor comments (6)

[§3.1 vs Table 5] Section 3.1 says the mean number of sound events is 25 per clip, while Table 5 reports Avg. No. Events 2.233 for SRSAT. Clarify that Table 5 statistics are per frame, not per clip.
[Eq. (8), §2.3.3] The dimension of F is written as R^{T*F}; since F_in/F_out are computed per time-frequency bin via max over θ, clarify whether F is a scalar field of shape T×F and state the dimensions of F_in/F_out.
[§3.2] 'we use thesmall224 version' appears to be a typo; please correct.
[§4.3] The overlap filter 'two regions overlapping by 30°' is ambiguous; specify the exact rule used to drop redundant regions in the location-aware system.
[§4.5, Table 5] Specify the model configuration and whether the STARSS23 mAP is for region-specific or omnidirectional tagging; otherwise the comparison is hard to interpret.
[§4.1] Since FOV gives the best mAP in Table 1, justify more explicitly why DF is used for the rest of the paper beyond computational cost, and consider reporting FR/LA with FOV as well.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central claims are supervised evaluations on held-out clips, with no fitted-input-as-prediction or self-citation chain.

full rationale

The paper's derivation chain is empirical rather than analytic. The task (Section 2.1) defines region-specific tagging from an independent region query p and ground-truth event positions; no quantity is defined in terms of the model output it is supposed to support. Features in Section 2.3 are standard spectral/spatial/positional features drawn from prior work, but they are used as inputs to a learned model and compared on a held-out test set (Section 4.1, Table 1), so no fitted parameter is renamed as a prediction. The main superiority claim (Section 4.3, Table 3) compares omnidirectional, fixed-region, and location-aware systems on the SRSAT test split; the location-aware system uses ground-truth positions and is explicitly labeled as having 'prior knowledge of the locations of all sound sources,' i.e. an oracle, not a hidden circular prediction. No uniqueness theorem is imported, and no ansatz is smuggled solely via self-citation: the authors' own prior works (e.g. [17], [18]) are used only for spatial-feature motivation, not to justify the central result. The STARSS23 check in Section 4.5 is a dataset-difficulty comparison rather than a head-to-head region-specific vs omnidirectional evaluation; that is an external-validity limitation, explicitly visible in the table as a single mAP row, but it is not circular reasoning. The manuscript's only self-noted limitation is the 13-class scope in Section 5, which is a scope statement, not a circular step. Therefore no circular step can be exhibited with the required specificity, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper rests on standard supervised audio ML practice and on a few domain assumptions about array geometry and simulation realism. No new physical entities are introduced. Main ledger items are hand-chosen experimental settings (region width, FOV resolution, overlap threshold, simulation event distribution) that directly affect headline numbers.

free parameters (5)

Angular region width = 60 degrees (also 180, 300 in Table 4)
All main region experiments use a 60-degree query region; Table 4 shows mAP drops from 0.473 to 0.370 as range increases, so headline numbers depend on this choice.
DF extraction angle = middle angle of region
For DF and learned angle features the query angle is the center of the 60-degree region, a hand choice that affects feature quality.
FOV resolution = 5 degrees
FOV feature computes DF at 5-degree resolution, 72x cost of DF; its performance and cost depend on this resolution.
Overlap threshold for LA filtering = 30 degrees
Location-aware system drops regions overlapping by 30 degrees to reduce cost; this changes the number of regions evaluated.
SRSAT event count distribution = mean 25, std 3 per clip (Section 3.1); Table 5 reports 2.233 average per frame
The simulated dataset difficulty is set by this distribution; the text and table are inconsistent, making exact replication ambiguous.

assumptions (3)

domain assumption Far-field plane-wave microphone-array phase model: P_n(theta) = 2*pi*f*phi_n*cos(theta)*fs/c
Equation (5) is used to build DF/FOV features; assumes known array geometry, far-field sources, and constant sound speed, which may fail in near-field or reverberant conditions.
domain assumption SpatialScaper simulation is representative of real spatial audio
SRSAT is generated with SpatialScaper using random rooms; conclusions about feature and model superiority are assumed to transfer, but real-data evidence is thin.
domain assumption STARSS23 labels can be converted to region tags without additional alignment error
STARSS23 provides event positions; converting them to region ground truth assumes exact position-to-region mapping.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Region-Specific Audio Tagging for Spatial Sound." pith.science (2026). https://pith.science/paper/XF7XLJHO

@misc{pith2026250909526,
  author       = {Pith},
  title        = {Pith review of: Region-Specific Audio Tagging for Spatial Sound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XF7XLJHO}},
  note         = {Machine review of arXiv:2509.09526}
}

read the original abstract

Audio tagging aims to label sound events appearing in an audio recording. In this paper, we propose region-specific audio tagging, a new task which labels sound events in a given region for spatial audio recorded by a microphone array. The region can be specified as an angular space or a distance from the microphone. We first study the performance of different combinations of spectral, spatial, and position features. Then we extend state-of-the-art audio tagging systems such as pre-trained audio neural networks (PANNs) and audio spectrogram transformer (AST) to the proposed region-specific audio tagging task. Experimental results on both the simulated and the real datasets show the feasibility of the proposed task and the effectiveness of the proposed method. Further experiments show that incorporating the directional features is beneficial for omnidirectional tagging.

Figures

Figures reproduced from arXiv: 2509.09526 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 4 linked inside Pith

[1]

Detection and classification of acoustic scenes and events,

D. Stowell, D. Giannoulis, E. Benetos, M. Lagrange, and M. D. Plumbley, “Detection and classification of acoustic scenes and events,”IEEE Transactions on Multimedia, vol. 17, no. 10, pp. 1733–1746, 2015

2015
[2]

Content-based classifica- tion, search, and retrieval of audio,

E. Wold, T. Blum, D. Keislar, and J. Wheaten, “Content-based classifica- tion, search, and retrieval of audio,”IEEE Multimedia, vol. 3, no. 3, pp. 27–36, 1996

1996
[3]

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

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “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
[4]

AudioSet: 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, “AudioSet: An ontology and human-labeled dataset for audio events,” inInternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017, pp. 776–780

2017
[5]

PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,

Y . Gong, Y .-A. Chung, and J. Glass, “PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3292–3306, 2021

2021
[6]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

2017
[7]

AST: Audio spectrogram trans- former,

Y . Gong, Y .-A. Chung, and J. Glass, “AST: Audio spectrogram trans- former,”arXiv preprint arXiv:2104.01778, 2021

arXiv 2021
[8]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

Show all 24 references

[9]

Convolutional gated recurrent neural network incorporating spatial features for audio tagging,

Y . Xu, Q. Kong, Q. Huang, W. Wang, and M. D. Plumbley, “Convolutional gated recurrent neural network incorporating spatial features for audio tagging,” inInternational Joint Conference on Neural Networks (IJCNN). IEEE, 2017, pp. 3461–3466

2017
[10]

ATGNN: Audio tagging graph neural network,

S. Singh, C. J. Steinmetz, E. Benetos, H. Phan, and D. Stowell, “ATGNN: Audio tagging graph neural network,”IEEE Signal Processing Letters, 2024

2024
[11]

Multi- channel overlapped speech recognition with location guided speech extraction network,

Z. Chen, X. Xiao, T. Yoshioka, H. Erdogan, J. Li, and Y . Gong, “Multi- channel overlapped speech recognition with location guided speech extraction network,” inSpoken Language Technology Workshop (SLT). IEEE, 2018, pp. 558–565

2018
[12]

Multi-modal multi-channel target speech separation,

R. Gu, S.-X. Zhang, Y . Xu, L. Chen, Y . Zou, and D. Yu, “Multi-modal multi-channel target speech separation,”IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 3, pp. 530–541, 2020

2020
[13]

Rezero: Region-customizable sound extraction,

R. Gu and Y . Luo, “Rezero: Region-customizable sound extraction,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2576–2589, 2024

2024
[14]

An alternative approach to linearly constrained adaptive beamforming,

L. Griffiths and C. Jim, “An alternative approach to linearly constrained adaptive beamforming,”IEEE Transactions on antennas and propagation, vol. 30, no. 1, pp. 27–34, 1982

1982
[15]

Deep audio zooming: Beamwidth-controllable neural beamformer,

M. Yu and D. Yu, “Deep audio zooming: Beamwidth-controllable neural beamformer,”arXiv preprint arXiv:2311.13075, 2023

2023 arXiv
[16]

Audio inputs for active speaker detection and localization via microphone array,

D. Berghi and P. J. Jackson, “Audio inputs for active speaker detection and localization via microphone array,” inWorkshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2023, pp. 1–5

2023
[17]

Audio-visual speaker tracking: Progress, challenges, and future directions,

J. Zhao, Y . Xu, X. Qian, D. Berghi, P. Wu, M. Cui, J. Sun, P. J. Jackson, and W. Wang, “Audio-visual speaker tracking: Progress, challenges, and future directions,”arXiv preprint arXiv:2310.14778, 2023

2023 arXiv
[18]

Attention- based end-to-end differentiable particle filter for audio speaker tracking,

J. Zhao, Y . Xu, X. Qian, H. Liu, M. D. Plumbley, and W. Wang, “Attention- based end-to-end differentiable particle filter for audio speaker tracking,” IEEE Open Journal of Signal Processing, vol. 5, pp. 449–458, 2024

2024
[19]

On spatial features for supervised speech separation and its application to beamforming and robust ASR,

Z.-Q. Wang and D. Wang, “On spatial features for supervised speech separation and its application to beamforming and robust ASR,” in International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5709–5713

2018
[20]

STARSS23: An audio-visual dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,

K. Shimada, A. Politis, P. Sudarsanam, D. A. Krause, K. Uchida, S. Adavanne, A. Hakala, Y . Koyama, N. Takahashi, S. Takahashiet al., “STARSS23: An audio-visual dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,”Advances in Neural Inf...

2024
[21]

Spatial Scaper: a library to simulate and augment soundscapes for sound event localization and detection in realistic rooms,

I. R. Roman, C. Ick, S. Ding, A. S. Roman, B. McFee, and J. P. Bello, “Spatial Scaper: a library to simulate and augment soundscapes for sound event localization and detection in realistic rooms,” inInternational Conference on Acoustics, Speech and Signal Processing (ICASSP). ...

2024
[22]

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 Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 829–852, 2021

2021
[23]

A four- stage data augmentation approach to ResNet-Conformer based acoustic modeling for sound event localization and detection,

Q. Wang, J. Du, H.-X. Wu, J. Pan, F. Ma, and C.-H. Lee, “A four- stage data augmentation approach to ResNet-Conformer based acoustic modeling for sound event localization and detection,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1251–1264, 2023

2023
[24]

SALSA: Spatial cue-augmented log-spectrogram features for polyphonic sound event localization and detection,

T. N. T. Nguyen, K. N. Watcharasupat, N. K. Nguyen, D. L. Jones, and W.-S. Gan, “SALSA: Spatial cue-augmented log-spectrogram features for polyphonic sound event localization and detection,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 1749–1762, 2022

2022

Pith tools

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

[1] [1]

Detection and classification of acoustic scenes and events,

D. Stowell, D. Giannoulis, E. Benetos, M. Lagrange, and M. D. Plumbley, “Detection and classification of acoustic scenes and events,”IEEE Transactions on Multimedia, vol. 17, no. 10, pp. 1733–1746, 2015

2015

[2] [2]

Content-based classifica- tion, search, and retrieval of audio,

E. Wold, T. Blum, D. Keislar, and J. Wheaten, “Content-based classifica- tion, search, and retrieval of audio,”IEEE Multimedia, vol. 3, no. 3, pp. 27–36, 1996

1996

[3] [3]

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

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “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

[4] [4]

AudioSet: 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, “AudioSet: An ontology and human-labeled dataset for audio events,” inInternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017, pp. 776–780

2017

[5] [5]

PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,

Y . Gong, Y .-A. Chung, and J. Glass, “PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3292–3306, 2021

2021

[6] [6]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

2017

[7] [7]

AST: Audio spectrogram trans- former,

Y . Gong, Y .-A. Chung, and J. Glass, “AST: Audio spectrogram trans- former,”arXiv preprint arXiv:2104.01778, 2021

arXiv 2021

[8] [8]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

arXiv 2010