Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation

David Han; Quoc Thinh Vo

REVIEW 3 major objections 5 minor 25 references

Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-branch ResNet-Conformer with synthetic pretraining and aggressive augmentation lifts sound-event detection F-score from 13.1 to 40.2 while cutting angular error from 36.9° to 17.5°.

desk verdict A plausible, decently-engineered SELD challenge report whose headline gains cannot be independently checked as written. read the letter →

arxiv 2507.17941 v1 pith:AX7IN64I submitted 2025-07-23 cs.SD eess.AS

Quoc Thinh Vo , David Han This is my paper

classification cs.SDeess.AS

keywords soundeventlocalizationanddetectiondistanceestimationevent-independentnetworkResNet-Conformermulti-ACCDOAmulti-scalechannelattentiondataaugmentationaudioswapping

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

This paper reports a system for sound event localization, detection, and distance estimation in real recordings, and it claims the system substantially outperforms the official baseline. The central result is an F-score of 40.2%, a direction-of-arrival error of 17.5 degrees, and a relative distance error of 0.32 on the development test set, against a baseline of 13.1%, 36.9 degrees, and 0.33. The claim is that a two-branch network with shared weights, Conformer sequence blocks, and multi-scale channel attention, trained first on synthetic data and then on real data with audio-channel swapping and random mixing, is what delivers most of that gain. The authors present three ablations supporting that attribution: without multi-phase training and augmentation the F-score is 23.1%, with synthetic initialization and standard augmentation it is 33.0%, and the full recipe reaches 40.2%.

What carries the argument

The load-bearing object is the multi-ACCDOA target with a distance coordinate: for each of up to three tracks, thirteen classes, and every time frame, the network must emit an activity flag, a unit-norm three-dimensional direction vector, and a nonnegative distance, giving 156 output units. The loss is a permutation-invariant mean-squared-error loss (MSE-ADPIT) that minimizes over all track orderings, which lets the model assign same-class overlapping events to distinct tracks. The architecture carries this target through a two-branch design with shared convolutional weights: ResNet residual blocks with rescaled connections, a multi-scale channel attention module fusing local and global channel features, Conformer blocks replacing the usual recurrent and attention layers, and max pooling after each shared layer. Two training ingredients do the heavy lifting: initializing weights on synthetic soundscapes generated by a room simulation library, then fine-tuning on real recordings with audio-channel swapping and random mix augmentation.

What would settle it

Run the official challenge scorer on the submitted predictions: if it does not return an F-score near 40.2%, an angular error near 17.5°, and a relative distance error near 0.32, the central comparison does not reproduce. A second check is to train the same network without audio-channel swapping and random mix, which the paper's ablation predicts should drop the F-score from 40.2% to about 33.0%.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a single ResNet-Conformer network with a multi-ACCDOA output can jointly predict whether a sound is present, where it comes from, and how far away it is, and that the distance-aware output does not hurt localization performance. The model is organized as two branches with shared weights: low-level features are extracted by ResNet convolutional blocks and refined by Conformer blocks that couple local and global temporal information, while a multi-scale channel attention module fuses local and global channel statistics. The output for each of three tracks and thirteen classes is a vector containing activity, a unit-norm Cartesian direction, and a distance value, trained with a permutation-invariant loss that also handles overlapping events of the same class. The authors report that this design, together with a two-phase training schedule that starts from synthetic mixtures and finetunes on augmented real data, improves the challenge's F-score from 23.1% to 40.2% and angular error from 25.3 to 17.5 degrees compared to their own no-augmentation baseline, with relative distance error essentially flat around 0.32.

Load-bearing premise

The numerical comparison is only valid if the authors' internal scoring code exactly implements the official challenge rules, and the paper describes those rules without releasing the scorer.

Editorial extensions

If this is right

In a direct comparison on the development test set, the full recipe raises the challenge F-score from the official baseline's 13.1% to 40.2% and cuts angular error from 36.9° to 17.5°.
The three reported configurations show the gain is mostly training strategy: the same architecture without multi-phase training and augmentation reaches 23.1% F-score and 25.3° angular error, while adding synthetic initialization and standard augmentation reaches 33.0% and 20.7°, and the full recipe reaches 40.2% and 17.5°.
The paper's stated conclusion is that distance estimation is the open part: the relative distance error of 0.32 is only marginally better than the baseline's 0.33, so the architecture that works for detection and direction does not yet solve distance.

Reading between the lines

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

Extending beyond the paper: because the Conformer blocks and attention module operate on any feature sequence, the same two-branch design could be applied to the challenge's audiovisual track by concatenating visual features; the paper reports only the audio-only version.
Extending beyond the paper: the ablation sequence does not isolate the effect of synthetic initialization from the effect of augmentation, since Model 2 changes both relative to Model 1; a controlled test would keep the augmentation recipe fixed and toggle only the synthetic pretraining.
Extending beyond the paper: the flatness of the relative distance error suggests distance errors may be dominated by scale or room-dependent bias rather than detection mistakes, and a log-distance output or a per-class calibration layer would be a direct way to test that hypothesis.

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

Short version: this is a DCASE 2024 Task 3A technical report, and for that genre it is a solid entry. The authors combine EINV2 with ResNet blocks, Conformer, MS-CAM, multi-ACCDOA output, and a multi-phase synthetic-to-real training schedule. The proposed model reports F20 of 40.2% and DOA error of 17.5 degrees against a baseline at 13.1% and 36.9 degrees. Those are large gains, and the ablation sequence (no augmentation, with augmentation, with ACS/random mix) shows a monotonic trend in the right direction. That is the main thing the paper has going for it: the engineering story is coherent, and the authors are honest that the RDE metric did not improve (0.32 vs 0.33), which they say in the conclusion.

The new piece is the specific combination and its measured performance on the 2024 development set, not any new learning principle. That is fine for a challenge system description.

The soft spots are real but proportionate. There is no code, no saved model outputs, and no error bars. Every row is a single run. The abstract says 17.7 degrees while Table 1 says 17.5; small but it makes you wonder what else was transcribed by hand. The evaluation pipeline is described only at the level of the official metric — one-second segments, micro-averaging, Hungarian matching — but not provided. It is quite likely they used the official DCASE evaluator, and the baseline numbers match the official baseline, so this is a reproducibility gap rather than evidence of a wrong result. Still, as written, an independent reader cannot verify the central claim. The citation to [17] is to an unpublished manuscript, which is a minor wart.

Overall: I trust the qualitative claim that this configuration improves on the baseline; I do not trust the exact numbers until the evaluator or outputs are released. The paper is exactly what it claims to be — a technical report — and it deserves a serious referee if submitted to a venue that expects reproducibility. I would accept it for peer review with the expectation that the authors either release the evaluation code or provide saved predictions. My own verdict would be: conditionally accept, with code release as a requirement.

Who benefits: anyone working on SELD with distance estimation, and especially participants in DCASE-style challenges. I would not cite the specific numbers in a literature review, but I would cite it as an example of a competitive system once the code is available.

Referee Report

3 major / 5 minor

Summary. The paper proposes a ResNet-Conformer network for the DCASE 2024 Task 3A sound event localization and detection (SELD) challenge, including distance estimation. The architecture builds on the EINV2 framework with shared weights, multi-scale channel attention, Conformer blocks, and a multi-phase training scheme that uses synthetic initialization, Audio Channel Swapping (ACS), and random mix augmentation. On the development test set, the authors report an F-score of 40.2%, a DOA angular error of 17.5 degrees (17.7 in the abstract), and a relative distance error (RDE) of 0.32, compared to the official baseline of 13.1%, 36.9 degrees, and 0.33. The paper is an empirical systems report, presenting ablation results for three model configurations and a discussion of the evaluation metrics.

Significance. If the reported numbers are accurate and reproducible, the proposed method delivers a substantial improvement over the official DCASE 2024 Task 3A baseline in F-score and DOA error, which would be a meaningful contribution to the SELD community. The use of Conformer blocks and multi-scale channel attention within an EINV2-based architecture is a reasonable and potentially effective design. The paper also provides a clear description of the feature extraction, augmentation, training phases, and loss function, which helps readers understand the approach. However, the significance is limited by the lack of released evaluation code or model outputs, the absence of multiple runs or confidence intervals, and the fact that the reported RDE improvement is negligible and acknowledged as such by the authors. The central claim rests entirely on self-reported metrics that have not been independently verified.

major comments (3)

[Abstract vs. Table 1] There is an inconsistency in the reported DOA error: the abstract states 17.7 degrees, while Table 1 reports 17.5 degrees for the Proposed Model. Since the DOA improvement over the baseline is a central quantitative claim, this discrepancy must be resolved and a single consistent value reported.
[Section 2.5 and Table 1] The evaluation pipeline is described only at a high level, and no evaluation code, saved model outputs, or instructions for reproducing the official DCASE 2024 Task 3 scoring are provided. The comparison to the baseline in Table 1 is therefore unverifiable; if the authors' implementation of micro-averaging, one-second segments, the 20-degree threshold, RDE threshold, or Hungarian matching differs from the official evaluator, the numerical comparison is invalid. The authors should release the evaluation code or the exact model outputs, or at minimum demonstrate that their evaluator reproduces the official baseline scores.
[Table 1 and Section 3] All reported metrics are from a single training run, with no seeds, confidence intervals, or significance tests. Given the stochastic nature of the augmentations (random cutout, SpecAugment, ACS, random mix) and multi-phase training, seed-to-seed variability could be material, and the RDE improvement from 0.33 to 0.32 is likely within noise. The paper should report multiple runs or otherwise quantify uncertainty, and should temper the claim of improved performance to focus on F-score and DOA, not RDE.

minor comments (5)

[Section 2.2] The 'random mix' augmentation is described as blending original and augmented data 'with specified weights,' but the weights are not given and no reference is supplied; please provide concrete details or a citation.
[Section 2.3] The architecture description references Figure 1, but the diagram is not fully described in the text; in particular, the exact placement of the shared-weight layers, max pooling, and Conformer blocks should be spelled out so the architecture is reproducible from the text alone.
[Section 2.1] The sentence 'The audio data was sampling at 24 kHz' contains a grammatical error; it should read 'sampled at 24 kHz' or 'has a sampling rate of 24 kHz.'
[References] Reference [17] (Xue et al., 'Attention mechanism network and data augmentation for sound event localization and detection') lacks complete bibliographic information, including venue and year; please complete the entry.
[Section 3] The opening sentence of Section 3 says the network was trained using a multi-phase approach, but Model 1 is described as not using that framework; please clarify which training scheme applies to which model.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported results are empirical measurements on a held-out development test set, not derived quantities.

full rationale

This is an empirical systems paper with no derivation chain to audit for circularity. The central claims are measured SELD metrics (F20°, DOA angular error, RDE) obtained by training a ResNet-Conformer model on the DCASE 2024 Task 3A development data and evaluating on the corresponding held-out test clips. No fitted constant, closed-form equation, or self-cited uniqueness theorem is used to produce these numbers; the architecture borrows EINV2, multi-ACCDOA, and augmentation components from prior work as engineering building blocks, and the loss function is explicitly attributed to Krause et al. Equation (1). The only substantive concern is external reproducibility: the evaluation pipeline is described but not released, and the abstract's 17.7 degrees differs from Table 1's 17.5 degrees, which is a correctness or reporting risk, not a circularity. Because the metrics are measured against external ground truth rather than reconstructed from the model's own inputs or fitted targets, no circular step can be exhibited, so the appropriate score is 0.

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

The central claim depends on standard SELD assumptions: the output representation, the evaluation protocol, and the transferability of synthetic pretraining. No new physical entities or fitted constants are introduced. The main unverified element is the exact evaluation implementation.

assumptions (3)

domain assumption The multi-ACCDOA output format with an added distance estimate represents the SELD task adequately (Section 2.4).
The paper extends ACCDOA to include distance following Krause et al. [18] without comparing to alternative output representations. If this representation is suboptimal, the reported metrics may understate achievable performance.
domain assumption The official DCASE 2024 evaluation metrics and scoring pipeline are used correctly (Section 2.5).
The reported F-score, angular error, and RDE depend on the exact implementation of micro-averaging, thresholds, and Hungarian matching. No evaluation code is provided to verify this.
domain assumption Synthetic data from SpatialScaper [15] and the official synthetic mixtures [19] are appropriate for pretraining (Section 2.4).
The multi-phase training assumes that pretraining on synthetic data transfers to real recordings. The paper does not ablate this assumption except by comparing Model 1 (no synthetic pretraining) with later models, but confounding factors exist.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation." pith.science (2026). https://pith.science/paper/AX7IN64I

@misc{pith2026250717941,
  author       = {Pith},
  title        = {Pith review of: Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AX7IN64I}},
  note         = {Machine review of arXiv:2507.17941}
}

read the original abstract

This technical report outlines our approach to Task 3A of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2024, focusing on Sound Event Localization and Detection (SELD). SELD provides valuable insights by estimating sound event localization and detection, aiding in various machine cognition tasks such as environmental inference, navigation, and other sound localization-related applications. This year's challenge evaluates models using either audio-only (Track A) or audiovisual (Track B) inputs on annotated recordings of real sound scenes. A notable change this year is the introduction of distance estimation, with evaluation metrics adjusted accordingly for a comprehensive assessment. Our submission is for Task A of the Challenge, which focuses on the audio-only track. Our approach utilizes log-mel spectrograms, intensity vectors, and employs multiple data augmentations. We proposed an EINV2-based [1] network architecture, achieving improved results: an F-score of 40.2%, Angular Error (DOA) of 17.7 degrees, and Relative Distance Error (RDE) of 0.32 on the test set of the Development Dataset [2 ,3].

Figures

Figures reproduced from arXiv: 2507.17941 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages

[17]

A four-stage data augmentation approach to resnet-conformer based acoustic modeling for sound event localization and de- tection,

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 de- tection,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 1251–1264, 2023

work page 2023
[1]

Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation

INTRODUCTION The goal of the Sound Event Localization and Detection (SELD) task is to detect sound events (SED) while estimating their corre- sponding direction of arrival (DOA). SELD systems have shown significant potential in diverse applications such as machine listen- ing, smart homes, navigation, and wildlife sound detection. The annual DCASE challen...

work page Pith review arXiv 2020
[2]

Features In our approach, we use audio files in Ambisonic format of the De- velopment Dataset 2024 [3]

PROPOSED METHODOLOGY 2.1. Features In our approach, we use audio files in Ambisonic format of the De- velopment Dataset 2024 [3]. The dataset comprises 7 hours and 22 minutes of real recordings, divided into 90 training clips and 78 test- ing clips. It includes 13 sound event classes: female speech, male speech, clapping, telephone, laughter, domestic sou...

work page 2024
[3]

The network was trained using a multi-phase approach involving syn- thetic, real, and augmented data

EXPERIMENTAL RESULTS Table 1 presents our results on the Development Dataset 2024. The network was trained using a multi-phase approach involving syn- thetic, real, and augmented data. Model 1, trained without the multi-phase training framework or data augmentation, achieved an F-score of 23.1%, DOA of 25.3 ◦, and RDE of 0.33. For Model 2, we initialized ...

work page 2024
[4]

CONCLUSION AND FUTURE WORK In this experiment, we implemented a ResNet-Conformer two- branch network with multi-phase training, achieving improved per- formance compared to the Baseline system in terms of F-score and Angular Error. However, our results indicate that the Relative Distance Error, a new evaluation metric introduced for this year’s challenge ...

work page
[5]

ACKNOWLEDGMENT The work on this paper is sponsored by the Office of Naval Research (Grant Number: N00014-21-1-2790)

work page
[6]

An improved event-independent network for polyphonic sound event localization and detection,

Y . Cao, T. Iqbal, Q. Kong, F. An, W. Wang, and M. D. Plumb- ley, “An improved event-independent network for polyphonic sound event localization and detection,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP). IEEE, 2021, pp. 885–889

work page 2021
[7]

STARSS22: A dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,

A. Politis, K. Shimada, P. Sudarsanam, S. Adavanne, D. Krause, Y . Koyama, N. Takahashi, S. Takahashi, Y . Mitsufuji, and T. Virtanen, “STARSS22: A dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,” in Proceedings of the 8th Detection and Classification of Acoustic Scenes and Events 2022 Workshop (DCASE2022) , N...

work page 2022

Show all 25 references

[8]

Shimada, A

K. Shimada, A. Politis, P. Sudarsanam, D. A. Krause, K. Uchida, S. Adavanne, A. Hakala, Y . Koyama, N. Takahashi, S. Takahashi, T. Virtanen, and Y . Mitsufuji, “STARSS23: An audio-visual dataset of spatial recordings Detection and Classification of Acoustic Scenes and Events 2...

2024
[9]

Accdoa: Activity-coupled cartesian direction of arrival representation for sound event localization and de- tection,

K. Shimada, Y . Koyama, N. Takahashi, S. Takahashi, and Y . Mitsufuji, “Accdoa: Activity-coupled cartesian direction of arrival representation for sound event localization and de- tection,” in ICASSP 2021-2021 IEEE International Confer- ence on Acoustics, Speech and Signal Pro...

2021
[10]

Permutation invariant training for speaker-independent multi-pitch tracking,

Y . Liu and D. Wang, “Permutation invariant training for speaker-independent multi-pitch tracking,” in 2018 IEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2018, pp. 5594–5598

2018
[11]

Multi-accdoa: Localizing and detecting overlapping sounds from the same class with aux- iliary duplicating permutation invariant training,

K. Shimada, Y . Koyama, S. Takahashi, N. Takahashi, E. Tsunoo, and Y . Mitsufuji, “Multi-accdoa: Localizing and detecting overlapping sounds from the same class with aux- iliary duplicating permutation invariant training,” in ICASSP 2022-2022 IEEE International Conference on A...

2022
[12]

Sound event localization and detection using foa domain spatial aug- mentation,

L. Mazzon, M. Yasuda, Y . Koizumi, and N. Harada, “Sound event localization and detection using foa domain spatial aug- mentation,” in Proc. of the 4th Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE), 2019

2019
[13]

Polyphonic sound event detection and localization using a two-stage strategy,

Y . Cao, Q. Kong, T. Iqbal, F. An, W. Wang, and M. D. Plumb- ley, “Polyphonic sound event detection and localization using a two-stage strategy,”arXiv preprint arXiv:1905.00268, 2019

1905 arXiv
[14]

The generalized correlation method for estimation of time delay,

C. Knapp and G. Carter, “The generalized correlation method for estimation of time delay,”IEEE transactions on acoustics, speech, and signal processing , vol. 24, no. 4, pp. 320–327, 1976

1976
[15]

Random erasing data augmentation,

Z. Zhong, L. Zheng, G. Kang, S. Li, and Y . Yang, “Random erasing data augmentation,” in Proceedings of the AAAI con- ference on artificial intelligence , vol. 34, no. 07, 2020, pp. 13 001–13 008

2020
[16]

Specaugment: A simple data aug- mentation method for automatic speech recognition,

D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data aug- mentation method for automatic speech recognition,” arXiv preprint arXiv:1904.08779, 2019

1904 arXiv
[18]

Conformer: Local features coupling global representations for visual recognition,

Z. Peng, W. Huang, S. Gu, L. Xie, Y . Wang, J. Jiao, and Q. Ye, “Conformer: Local features coupling global representations for visual recognition,” inProceedings of the IEEE/CVF inter- national conference on computer vision, 2021, pp. 367–376

2021
[19]

Sound event localization based on sound intensity vector refined by dnn-based denoising and source separation,

M. Yasuda, Y . Koizumi, S. Saito, H. Uematsu, and K. Imoto, “Sound event localization based on sound intensity vector refined by dnn-based denoising and source separation,” in ICASSP 2020-2020 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) ...

2020
[20]

Spatial scaper: A library to simulate and augment soundscapes for sound event localization and detection in real- istic 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 real- istic rooms,” in IEEE International Conference on Acoustics, Speech and Signal Processing (I...

2024
[21]

At- tentional feature fusion,

Y . Dai, F. Gieseke, S. Oehmcke, Y . Wu, and K. Barnard, “At- tentional feature fusion,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2021, pp. 3560–3569

2021
[22]

Attention mechanism network and data augmentation for sound event localization and detec- tion

L. Xue, H. Liu, and Y . Zhou, “Attention mechanism network and data augmentation for sound event localization and detec- tion.”
[23]

Sound event detec- tion and localization with distance estimation,

D. A. Krause, A. Politis, and A. Mesaros, “Sound event detec- tion and localization with distance estimation,”arXiv preprint arXiv:2403.11827, 2024

2024 arXiv
[24]

DCASE2024 Task 3 synthetic seld mixtures for baseline training,

D. A. Krause and A. Politis, “DCASE2024 Task 3 synthetic seld mixtures for baseline training,” Apr. 2024. [Online]. Available: https://doi.org/10.5281/zenodo.10932241

2024 doi
[25]

Overview and evaluation of sound event localization and detection in dcase 2019,

A. Politis, A. Mesaros, S. Adavanne, T. Heittola, and T. Vir- tanen, “Overview and evaluation of sound event localization and detection in dcase 2019,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, vol. 29, pp. 684–698, 2020

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [17]

A four-stage data augmentation approach to resnet-conformer based acoustic modeling for sound event localization and de- tection,

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 de- tection,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 1251–1264, 2023

work page 2023

[2] [1]

Resnet-conformer network with shared weights and attention mechanism for sound event localization, detection, and distance estimation

INTRODUCTION The goal of the Sound Event Localization and Detection (SELD) task is to detect sound events (SED) while estimating their corre- sponding direction of arrival (DOA). SELD systems have shown significant potential in diverse applications such as machine listen- ing, smart homes, navigation, and wildlife sound detection. The annual DCASE challen...

work page Pith review arXiv 2020

[3] [2]

Features In our approach, we use audio files in Ambisonic format of the De- velopment Dataset 2024 [3]

PROPOSED METHODOLOGY 2.1. Features In our approach, we use audio files in Ambisonic format of the De- velopment Dataset 2024 [3]. The dataset comprises 7 hours and 22 minutes of real recordings, divided into 90 training clips and 78 test- ing clips. It includes 13 sound event classes: female speech, male speech, clapping, telephone, laughter, domestic sou...

work page 2024

[4] [3]

The network was trained using a multi-phase approach involving syn- thetic, real, and augmented data

EXPERIMENTAL RESULTS Table 1 presents our results on the Development Dataset 2024. The network was trained using a multi-phase approach involving syn- thetic, real, and augmented data. Model 1, trained without the multi-phase training framework or data augmentation, achieved an F-score of 23.1%, DOA of 25.3 ◦, and RDE of 0.33. For Model 2, we initialized ...

work page 2024

[5] [4]

CONCLUSION AND FUTURE WORK In this experiment, we implemented a ResNet-Conformer two- branch network with multi-phase training, achieving improved per- formance compared to the Baseline system in terms of F-score and Angular Error. However, our results indicate that the Relative Distance Error, a new evaluation metric introduced for this year’s challenge ...

work page

[6] [5]

ACKNOWLEDGMENT The work on this paper is sponsored by the Office of Naval Research (Grant Number: N00014-21-1-2790)

work page

[7] [6]

An improved event-independent network for polyphonic sound event localization and detection,

Y . Cao, T. Iqbal, Q. Kong, F. An, W. Wang, and M. D. Plumb- ley, “An improved event-independent network for polyphonic sound event localization and detection,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP). IEEE, 2021, pp. 885–889

work page 2021

[8] [7]

STARSS22: A dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,

A. Politis, K. Shimada, P. Sudarsanam, S. Adavanne, D. Krause, Y . Koyama, N. Takahashi, S. Takahashi, Y . Mitsufuji, and T. Virtanen, “STARSS22: A dataset of spatial recordings of real scenes with spatiotemporal annotations of sound events,” in Proceedings of the 8th Detection and Classification of Acoustic Scenes and Events 2022 Workshop (DCASE2022) , N...

work page 2022