Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization

Ee-Leng Tan; Jun-Wei Yeow; Santi Peksi; Woon-Seng Gan

REVIEW 4 major objections 4 minor 33 references

Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization

T0 review · 4 major / 4 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 stereo-only 3D sound pipeline cuts aggregate SELD error from 0.409 to 0.294.

desk verdict A plausible stereo SELD pipeline whose headline gains are probably real, but ITFM is under-specified, the comparison is system-level, and the distance-normalization split needs clarifying. read the letter →

arxiv 2507.00874 v1 pith:3VCLSRDX submitted 2025-07-01 eess.AS

Jun-Wei Yeow , Ee-Leng Tan , Santi Peksi , Woon-Seng Gan This is my paper

classification eess.AS

keywords stereosoundeventlocalizationanddetection3DSELDdistanceestimationmid-sideintensityfeaturesmagnitude-squaredcoherencedataaugmentationnormalizationDCASE2025Task3

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 is a challenge submission arguing that stereo-only two-channel audio is enough for practical 3D sound event localization and detection when three things are combined: perceptually motivated input features, data augmentation designed around stereo spatial cues, and normalization of the distance regression target. The authors try to establish that these ingredients, added to a lightweight 4-million-parameter CRNN, cut the aggregate SELD error from 0.409 (baseline) to 0.294 on the development test split, while raising the location-dependent F-score from 23.72 to 45.32. The practical point is that consumer stereo hardware could support direction and distance estimates for sound events, not just detect what is happening. The report also claims each ingredient pays off separately, with distance normalization alone reducing the baseline error to 0.379.

What carries the argument

The load-bearing object is the MSIC feature stack: left-right log-Mel spectrograms plus Mid-Side log-Mel spectrograms, a Mel-projected intensity component $\tilde{I}_x(t,f)=\Re\{M(t,f)S^*(t,f)\}/\left(|M(t,f)|^2+|S(t,f)|^2+\epsilon\right)$ derived from the Mid-Side cross-spectrum, and a Mel-projected magnitude-squared coherence $\hat{\gamma}(t,f)$ between the two channels. These features give the network direction and distance cues from only two channels. The second mechanism is ITFM, described as a stereo-aware time-frequency masking that pre-computes inter-channel differences and reapplies them after masking so spatial information survives; the paper does not specify the mask construction beyond that. The third mechanism is distance normalization, $d'=(d-\bar d)/\sigma_d$ followed by division by its maximum, which keeps all ACCDOA regression targets on the same $[-1,1]$ scale.

What would settle it

Replace ITFM with a standard SpecAugment mask using the same time-frequency budget in one of the submitted configurations, and check the aggregate SELD error on the development test split; if removing ITFM does not raise the error, the claimed spatial-preservation mechanism is not the source of the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that a stereo 3D SELD system built from two-channel audio can outperform the provided baseline by a wide margin through a specific recipe: log-Mel spectrograms of the left and right channels augmented with Mid-Side log-Mel features, a Mel-projected MS intensity component that plays the role of an intensity vector, and magnitude-squared coherence between the channels; stereo-aware augmentation including channel swapping, Inter-Channel-Aware Time-Frequency Masking (ITFM), and a FilterAugment-plus-frequency-shifting pipeline; and a two-step distance normalization that maps meter distances to $[-1,1]$ before regression. In the submitted configurations, all four feature-plus-augmentation combinations land in the aggregate error band $0.294$–$0.302$, with the best setup (MSI features plus FilterAugment and Frequency Shifting) reaching $F_{20^\circ/1}=45.32$, $LE_{CD}=13.2^\circ$, $RDE_{CD}=0.262$, and $E_{SELD}=0.294$, compared with the baseline's $23.72$, $20.8^\circ$, $0.347$, and $0.409$.

Load-bearing premise

The load-bearing premise is that ITFM really preserves the left-right differences that carry spatial information after time-frequency masking, since the report does not give the mask recipe or compare it with standard SpecAugment or Cutout.

Editorial extensions

If this is right

Distance normalization alone lowers the baseline aggregate SELD error from 0.409 to 0.379, so part of the gain is separable from the new features and augmentation.
All four submitted systems cluster between 0.294 and 0.302 aggregate error, a relative reduction of roughly 26–28% from the baseline, suggesting the recipe is not highly sensitive to which augmentation pipeline is used.
The best system reaches a location-dependent F-score of 45.32 versus 23.72 at the $20^\circ$/$1$ m tolerance, meaning many more events are both detected and correctly placed in space.
Because the full model uses about 4 million parameters and 1.89G MACs per forward pass, these gains are claimed for a lightweight architecture that could run on modest hardware.

Reading between the lines

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

The paper does not ablate ITFM against plain SpecAugment or Cutout, so a reader should not credit the named masking mechanism with the pipeline-level gain until such a comparison is run.
The same post-mask reapplication idea could transfer to binaural or multi-microphone SELD, where standard masking can also erase inter-channel level differences that carry spatial information.
Because training mixes roughly 45 hours of channel-swapped real audio with 30,000 synthetic clips and 20 epochs of real-audio fine-tuning, the reported error is a property of the whole pipeline, not of the proposed features in isolation.
A natural next test is to swap the proposed features onto the baseline architecture without the augmentation or normalization changes, which would quantify each ingredient's marginal contribution.

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 one thing you should know: this is a competent DCASE technical report whose headline system-level gains are probably real, but whose named novel contribution is not actually verified. The paper attributes much of its improvement to ITFM, a stereo-aware time-frequency masking variant, yet gives no algorithm for it and runs no ablation. Separately, the distance-normalization step that affects the headline numbers has a data-split ambiguity worth checking before trusting Table 2.

What is new here: applying FilterAugment to 3D SELD, the MS/IV/MSC feature stack for stereo, channel swapping, and distance normalization. The experimental setup is sensible: external baseline, public STARSS23, code available, and clear metrics. That is genuine engineering value.

The soft spots are real but not fatal in themselves. First, the central comparison (Table 2 vs Table 1) changes architecture, data augmentation, features, and fine-tuning simultaneously. So 'significant improvements' cannot be attributed to any specific component. A simple ablation would fix this. Second, ITFM is described qualitatively: 'pre-computes and reapplies inter-channel differences post-masking' — no mask size, number of masks, or comparison to SpecAugment/Cutout. That is a reproducibility gap. Third, Eq. 9 states that normalization maps distances to [-1,1] but only divides by the positive max, so the lower bound is not guaranteed; this is minor compared to the split issue: the paper says 'mean and standard deviation of all distances' without saying which split. In a DCASE submission, using the dev-test split's distances to normalize targets would leak test information. If that happened, the reported ESELD gains are inflated. The paper should state clearly that only training-set statistics are used.

Net: the pipeline idea is reasonable and the engineering effort is honest, but the paper currently overclaims the evidence for ITFM and leaves a correctness-relevant ambiguity in DN. A revised version with ablations, a precise ITFM description, and an explicit normalization-split statement would be worth publishing.

I would accept it for peer review on the strength of the system-level empirical result, but the referee should demand those clarifications. For my own work, I wouldn't cite the specific numbers until the DN split is confirmed.

Referee Report

4 major / 4 minor

Summary. The manuscript is a DCASE 2025 Challenge Task 3 technical report describing a stereo 3D sound event localization and detection (SELD) system. It proposes perceptually motivated input features (mid-side intensity and magnitude-squared coherence), stereo-specific data augmentation (channel swapping, inter-channel-aware time-frequency masking, and FilterAugment plus frequency shifting), and a distance normalization preprocessing step. The system is a ResNet-biGRU CRNN with around 4 million parameters, trained on a combination of real STARSS23 stereo data, synthetic SpatialScaper data, and channel-swapped copies. Reported results on the STARSS23 development test split show an aggregated SELD error of 0.294 versus 0.409 for the baseline SELDNet.

Significance. If the reported numbers are reliable, the work demonstrates a practical and effective stereo 3D SELD system with substantial gains over the official baseline across all metrics, and the release of code is a useful contribution to the community. However, the experimental design currently does not isolate the contributions of the proposed components, the central novel augmentation (ITFM) is only described verbally, and the distance-normalization statistics are not tied to a data split. The significance of the paper therefore hinges on clarifications and additional ablations.

major comments (4)

[§7, Table 2 and §6.2] The headline claim of 'significant improvements' compares the submitted pipeline (Rows A–D) with the SELDNet baseline, but the compared systems differ simultaneously in network architecture, training data (30k synthetic samples), channel swapping, fine-tuning, and augmentation, while the proposed input features and ITFM are never ablated against a plain LR log-Mel baseline with the same architecture and data. Table 2 only contrasts MSI vs MSIC and ITFM vs FAFS; there is no row with, e.g., LR features without augmentation or with SpecAugment, so the attribution of the 0.409→0.294 gain to the proposed components is not established. Please provide ablations that keep architecture, data, and fine-tuning fixed while toggling each proposed component.
[§3.2] ITFM is the paper's named novel contribution, but Section 3.2 describes it only verbally ('pre-computes and reapplies inter-channel differences post-masking') and gives no algorithm, mask geometry, mask size, number of masks, or whether the same mask is applied across channels. Without this detail, the method cannot be replicated or compared with SpecAugment/Cutout, and the central claim that ITFM preserves spatial cues is untestable. Please provide a precise description (including pseudo-code and hyperparameters) and an ablation of ITFM against standard time-frequency masking.
[§5, Eq. (9), Tables 1–2] The distance-normalization statistics are not tied to a data split. The text says 'mean and standard deviation of all distances' without stating whether those statistics are computed from the training split only or from the full development set including the test split used for evaluation. If test-set distances are used to compute mean/std/max, the regression targets for evaluation clips are constructed using test-label information, which would bias the comparison and invalidate the reported gains. Please state explicitly that the statistics are computed on the training split only and, ideally, cross-validate the normalization parameters. Separately, Eq. (9) does not guarantee the claimed [-1, 1] range: dividing by the positive maximum of d' does not enforce a lower bound of -1.
[§6.2 and §7] All results are single runs without error bars or significance tests. The differences between the submitted configurations in Table 2 (e.g., ESELD 0.294 vs 0.300) are small, and the combination of early stopping by validation F-score and 20-epoch fine-tuning on real data makes it unclear how stable these numbers are. Please report multiple runs (or seed-averaged results) and variance, or clearly state that these are single-run challenge submissions.

minor comments (4)

[§2.2] The symbol MSC(t,f) is used both for the smoothed coherence in Eq. (7) and for its Mel-projected version in Eq. (8); please use distinct notation for the raw and projected features.
[§3.2] For Frequency Shifting, the text says 'shifting frequencies within a controlled range' but does not specify the shift range in Hz or bins or whether the shift is along the frequency axis; please add this detail.
[§6.1] Typographical and clarity issue: 'the training set containing22.5h' is missing a space, and the sentence would benefit from stating explicitly that the 30,000 synthetic samples are in addition to the real recordings rather than a replacement.
[§4] The multi-ACCDDOA output format is mentioned but not defined in the text; please add a brief description or explicitly refer the reader to [21,22] for the output format and loss computation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline gain is an external-benchmark comparison, not a consequence of the paper's own definitions.

full rationale

The paper's central claim (Section 7, Table 2: ESELD 0.294 vs baseline 0.409) is an empirical comparison against the DCASE 2025 Task 3 baseline on the external STARSS23 development test split. None of the proposed components is defined in terms of the target metric. The MS/MSIC features in Section 2 are explicit functions of the stereo input (Eqs. 1-8); the augmentation methods in Section 3 modify inputs, not labels; the CRNN in Section 4 is a standard architecture with the multi-ACCDDOA output from [21,22]; and the distance normalization in Eq. 9 is a fixed pointwise transform of the label d. Although the architecture and DN are taken from the authors' previous report [20], the paper re-evaluates DN against a no-DN baseline in Table 1, so the current gain does not rest on the self-citation. The only text-level concern is that Section 5 does not state whether mean/std/max in Eq. 9 are computed on the training split or the full dataset; if test labels were used, this would be data leakage, but that is a correctness risk and not a circularity of the derivation, so it does not change the score. ITFM is under-specified, but an underspecified augmentation is not a self-referential reduction. Overall the derivation is self-contained against external benchmarks.

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

The empirical gains rest on domain assumptions about stereo features (MS intensity and MSC as distance and direction cues), on synthetic data realism, and on data-dependent label normalization. The smoothing coefficient lambda is a hand-set constant; the normalization statistics are data-derived and their split is not specified.

free parameters (4)

MSC smoothing coefficient lambda = 0.8
Set by hand following [11] in Eq. (6); the MSC feature used by all submitted systems depends on it.
distance normalization statistics (mean, std, max) = computed from STARSS23 distance labels
Eq. (9) scales regression targets using dataset statistics. The text says 'all distances' without specifying train-only, so if the test split is included this is label leakage into preprocessing.
ITFM mask hyperparameters (mask size, number of masks) = not reported
The ITFM procedure is not specified, so the masking schedule is an undisclosed free choice; without it the method cannot be replicated.
FAFS augmentation ranges (filter gain bounds, frequency shift range) = not reported
FilterAugment and frequency shifting are applied on-the-fly, but the paper gives no ranges for gains or shifts, which are needed to reproduce the augmentation pipeline.

assumptions (4)

domain assumption MSC correlates with source distance and near/focused versus diffuse/distant qualities
Section 2.2 states high MSC values signal direct, coherent sources while low values suggest diffuse or distant sources; the model is expected to learn distance from this feature.
domain assumption Channel swapping with azimuth inversion preserves reverberation and label validity for stereo
Section 3.1 assumes swapping L/R and flipping azimuth labels doubles directional events while preserving recording environment conditions.
domain assumption Synthetic SpatialScaper data converted to stereo is a valid proxy for real stereo recordings
Section 6.1 adds 30,000 synthetic stereo samples to mitigate class imbalance; if synthetic stereo does not match the real distribution, training could bias the model.
standard math STFT and cross-spectral density estimation via recursive averaging are valid for short five-second clips
Standard signal processing assumed in Section 2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization." pith.science (2026). https://pith.science/paper/3VCLSRDX

@misc{pith2026250700874,
  author       = {Pith},
  title        = {Pith review of: Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VCLSRDX}},
  note         = {Machine review of arXiv:2507.00874}
}

read the original abstract

This technical report presents our submission to Task 3 of the DCASE 2025 Challenge: Stereo Sound Event Localization and Detection (SELD) in Regular Video Content. We address the audio-only task in this report and introduce several key contributions. First, we design perceptually-motivated input features that improve event detection, sound source localization, and distance estimation. Second, we adapt augmentation strategies specifically for the intricacies of stereo audio, including channel swapping and time-frequency masking. We also incorporate the recently proposed FilterAugment technique that has yet to be explored for SELD work. Lastly, we apply a distance normalization approach during training to stabilize regression targets. Experiments on the stereo STARSS23 dataset demonstrate consistent performance gains across all SELD metrics. Code to replicate our work is available in this repository: https://github.com/itsjunwei/NTU_SNTL_Task3

Figures

Figures reproduced from arXiv: 2507.00874 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

[1]

This form of spatial intelligence can be extended into three-dimensions by integrating Sound Distance Estimation (SDE), cumulating in 3D SELD

INTRODUCTION Sound Event Localization and Detection (SELD) is a form of machine-listening that enables systems to not only understand what sounds are happening, but also where they come from [1]. This form of spatial intelligence can be extended into three-dimensions by integrating Sound Distance Estimation (SDE), cumulating in 3D SELD. The transition of ...

work page 2025
[2]

Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization

INPUT FEA TURES Let xL[n] and xR[n] denote the left and right stereo input channels, respectively, with n being the discrete-time index. The Short-Time Fourier Transform (STFT) of the c-th channel at time frame t and frequency bin f is denoted as Xc(t, f), for c ∈ {L, R}. 2.1. Mid-Side Conversion Mid-Side (MS) conversion explicitly decomposes the stereo s...

work page Pith review arXiv 2025
[3]

DA TA AUGMENTA TION We employ both waveform-level and spectrogram-level augmen- tation methods to generate meaningful variations in stereo spatial cues, thereby improving model robustness. 3.1. Waveform-level Audio channel swapping (ACS) methods have been developed for both the FOA and MIC audio formats [12, 13]. ACS-based methods are extremely effective ...

work page 2025
[4]

The architecture follows similarly to our previous DCASE submission in 2024 [20]

NETWORK ARCHITECTURE For this year’s challenge, we employed a relatively lightweight convolutional recurrent neural network (CRNN) built on a ResNet backbone followed by bi-directional gated recurrent units (bi- GRUs). The architecture follows similarly to our previous DCASE submission in 2024 [20]. For the output, we use the multi- ACCDDOA output format ...

work page 2024
[5]

DISTANCE NORMALIZA TION The distance values in the STARSS23 dataset can range from [0.04, 7.64] in meters. If we were to directly regress these values in ACCDOA-based output format variants, the MSE loss function can very easily be biased towards further or more distant sound events [22]. Therefore, to mitigate this problem, we apply the dis- tance normal...

work page
[6]

Dataset The STARSS23 dataset consists of real-world, multi-room record- ings with annotations of event activity, spatial trajectories, and dis- tances [23]

EXPERIMENTAL METHOD 6.1. Dataset The STARSS23 dataset consists of real-world, multi-room record- ings with annotations of event activity, spatial trajectories, and dis- tances [23]. The stereo version of STARSS23 comprises of 30,000 five-second audio recordings, with the training set containing22.5h of audio data. To enrich the number of real-world direct...

work page
[7]

These include the location-dependent F-score (F20◦/1), class-dependent localization error (LE CD), and class- dependent relative distance error (RDECD)

RESULTS We employ the same validation metrics as used in DCASE 2025 Challenge Task 3. These include the location-dependent F-score (F20◦/1), class-dependent localization error (LE CD), and class- dependent relative distance error (RDECD). In addition, we also cal- culate an aggregated SELD error (ESELD) to provide an overview of the overall performance of...

work page 2025
[8]

We use perceptually-motivated input features to improve both localization and distance estimation performance

CONCLUSION This technical report details our proposed methods for the stereo- based 3D SELD task. We use perceptually-motivated input features to improve both localization and distance estimation performance. We introduce FilterAugment for the 3D SELD task, and propose a stereo-specific form of spectrogram masking augmentation. Over- all, our proposed app...

work page

Show all 33 references

[9]

Sound event localization and detection of overlapping sources using convolutional recurrent neural networks,

S. Adavanne, A. Politis, J. Nikunen, and T. Virtanen, “Sound event localization and detection of overlapping sources using convolutional recurrent neural networks,”IEEE Journal of Se- lected Topics in Signal Processing, vol. 13, no. 1, pp. 34–48, 2018

2018
[10]

Baseline models and evaluation of sound event localization and detection with distance estimation in dcase2024 challenge,

D. Diaz-Guerra, A. Politis, P. Sudarsanam, K. Shimada, D. A. Krause, K. Uchida, Y . Koyama, N. Takahashi, S. Takahashi, T. Shibuya, Y . Mitsufuji, and T. Virtanen, “Baseline models and evaluation of sound event localization and detection with distance estimation in dcase2024 c...

2024
[11]

Wearable seld dataset: Dataset for sound event localiza- tion and detection using wearable devices around head,

K. Nagatomo, M. Yasuda, K. Yatabe, S. Saito, and Y . Oikawa, “Wearable seld dataset: Dataset for sound event localiza- tion and detection using wearable devices around head,” in ICASSP 2022-2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP...

2022
[12]

Real-time sound event localization and detection: Deployment challenges on edge devices,

J. W. Yeow, E.-L. Tan, J. Bai, S. Peksi, and W.-S. Gan, “Real-time sound event localization and detection: Deployment challenges on edge devices,” arXiv preprint arXiv:2409.11700, 2024

2024 arXiv
[13]

Two vs. four-channel sound event localization and detection,

J. Wilkins, M. Fuentes, L. Bondi, S. Ghaffarzadegan, A. Abavisani, and J. P. Bello, “Two vs. four-channel sound event localization and detection,” in Proceedings of the 8th Detection and Classification of Acoustic Scenes and Events 2023 Workshop (DCASE2023), Tampere, Finland, ...

2023
[14]

Convolutional neural networks with binaural representations and background subtraction for acoustic scene classification

Y . Han, J. Park, and K. Lee, “Convolutional neural networks with binaural representations and background subtraction for acoustic scene classification.” in DCASE, 2017, pp. 46–50

2017
[15]

Two-stage sound event localization and detec- tion using intensity vector and generalized cross-correlation,

Y . Cao, T. Iqbal, Q. Kong, M. Galindo, W. Wang, and M. Plumbley, “Two-stage sound event localization and detec- tion using intensity vector and generalized cross-correlation,” DCASE2019 Challenge, Tech. Rep, 2019

2019
[16]

Crnn-based multiple doa estimation using acoustic intensity features for ambisonics recordings,

L. Perotin, R. Serizel, E. Vincent, and A. Gu´erin, “Crnn-based multiple doa estimation using acoustic intensity features for ambisonics recordings,” IEEE Journal of Selected Topics in Signal Processing, vol. 13, no. 1, pp. 22–33, 2019

2019
[17]

Supervised learning-based sound source distance es- timation using multivariate features,

K. Zhagyparova, R. Zhagypar, A. Zollanvari, and M. T. Akhtar, “Supervised learning-based sound source distance es- timation using multivariate features,” in2021 IEEE Region 10 Symposium (TENSYMP). IEEE, 2021, pp. 1–5

2021
[18]

Coherent-to-diffuse power ratio estimation for dereverberation,

A. Schwarz and W. Kellermann, “Coherent-to-diffuse power ratio estimation for dereverberation,” IEEE/ACM Transac- tions on Audio, Speech, and Language Processing , vol. 23, no. 6, pp. 1006–1018, 2015

2015
[19]

Blind estima- tion of the coherent-to-diffuse energy ratio from noisy speech signals,

M. Jeub, C. Nelke, C. Beaugeant, and P. Vary, “Blind estima- tion of the coherent-to-diffuse energy ratio from noisy speech signals,” in 2011 19th European Signal Processing Confer- ence. IEEE, 2011, pp. 1347–1351

2011
[20]

First order ambisonics domain spatial augmentation for dnn-based direction of arrival estimation,

L. Mazzon, Y . Koizumi, M. Yasuda, and N. Harada, “First order ambisonics domain spatial augmentation for dnn-based direction of arrival estimation,” arXiv preprint arXiv:1910.04388, 2019

1910 arXiv
[21]

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

2023
[22]

An experimental study on sound event localization and detection under realistic testing conditions,

S. Niu, J. Du, Q. Wang, L. Chai, H. Wu, Z. Nian, L. Sun, Y . Fang, J. Pan, and C.-H. Lee, “An experimental study on sound event localization and detection under realistic testing conditions,” in ICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal P...

2023
[23]

Filteraugment: An acoustic environmental data augmentation method,

H. Nam, S.-H. Kim, and Y .-H. Park, “Filteraugment: An acoustic environmental data augmentation method,” in ICASSP 2022-2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 4308–4312

2022
[24]

Salsa-lite: A fast and effective feature for polyphonic sound event localization and detection with microphone arrays,

T. N. T. Nguyen, D. L. Jones, K. N. Watcharasupat, H. Phan, and W.-S. Gan, “Salsa-lite: A fast and effective feature for polyphonic sound event localization and detection with microphone arrays,” in ICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Sign...

2022
[25]

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
[26]

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
[27]

Improved regularization of convolutional neural networks with cutout,

T. DeVries and G. W. Taylor, “Improved regularization of convolutional neural networks with cutout,” arXiv preprint arXiv:1708.04552, 2017

2017 arXiv
[28]

Squeeze-and-excite resnet-conformers for sound event local- ization, detection, and distance estimation for dcase2024 chal- lenge,

J. W. Yeow, E.-L. Tan, J. Bai, S. Peksi, and W.-S. Gan, “Squeeze-and-excite resnet-conformers for sound event local- ization, detection, and distance estimation for dcase2024 chal- lenge,” DCASE2024 Challenge, Tech. Rep., 2024

2024
[29]

Multi-accdoa: Localiz- ing and detecting overlapping sounds from the same class with auxiliary duplicating permutation invariant training,

K. Shimada, Y . Koyama, S. Takahashi, N. Takahashi, E. Tsunoo, and Y . Mitsufuji, “Multi-accdoa: Localiz- ing and detecting overlapping sounds from the same class with auxiliary duplicating permutation invariant training,” in ICASSP 2022-2022 IEEE international conference on a...

2022
[30]

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
[31]

Starss23: An audio-visual dataset of spatial recordings of real scenes with spatiotemporal an- notations of sound events,

K. Shimada, A. Politis, P. Sudarsanam, D. A. Krause, K. Uchida, S. Adavanne, A. Hakala, Y . Koyama, N. Taka- hashi, S. Takahashi, et al., “Starss23: An audio-visual dataset of spatial recordings of real scenes with spatiotemporal an- notations of sound events,” Advances in Neu...

2024
[32]

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
[33]

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] [1]

This form of spatial intelligence can be extended into three-dimensions by integrating Sound Distance Estimation (SDE), cumulating in 3D SELD

INTRODUCTION Sound Event Localization and Detection (SELD) is a form of machine-listening that enables systems to not only understand what sounds are happening, but also where they come from [1]. This form of spatial intelligence can be extended into three-dimensions by integrating Sound Distance Estimation (SDE), cumulating in 3D SELD. The transition of ...

work page 2025

[2] [2]

Improving Stereo 3D Sound Event Localization and Detection: Perceptual Features, Stereo-specific Data Augmentation, and Distance Normalization

INPUT FEA TURES Let xL[n] and xR[n] denote the left and right stereo input channels, respectively, with n being the discrete-time index. The Short-Time Fourier Transform (STFT) of the c-th channel at time frame t and frequency bin f is denoted as Xc(t, f), for c ∈ {L, R}. 2.1. Mid-Side Conversion Mid-Side (MS) conversion explicitly decomposes the stereo s...

work page Pith review arXiv 2025

[3] [3]

DA TA AUGMENTA TION We employ both waveform-level and spectrogram-level augmen- tation methods to generate meaningful variations in stereo spatial cues, thereby improving model robustness. 3.1. Waveform-level Audio channel swapping (ACS) methods have been developed for both the FOA and MIC audio formats [12, 13]. ACS-based methods are extremely effective ...

work page 2025

[4] [4]

The architecture follows similarly to our previous DCASE submission in 2024 [20]

NETWORK ARCHITECTURE For this year’s challenge, we employed a relatively lightweight convolutional recurrent neural network (CRNN) built on a ResNet backbone followed by bi-directional gated recurrent units (bi- GRUs). The architecture follows similarly to our previous DCASE submission in 2024 [20]. For the output, we use the multi- ACCDDOA output format ...

work page 2024

[5] [5]

DISTANCE NORMALIZA TION The distance values in the STARSS23 dataset can range from [0.04, 7.64] in meters. If we were to directly regress these values in ACCDOA-based output format variants, the MSE loss function can very easily be biased towards further or more distant sound events [22]. Therefore, to mitigate this problem, we apply the dis- tance normal...

work page

[6] [6]

Dataset The STARSS23 dataset consists of real-world, multi-room record- ings with annotations of event activity, spatial trajectories, and dis- tances [23]

EXPERIMENTAL METHOD 6.1. Dataset The STARSS23 dataset consists of real-world, multi-room record- ings with annotations of event activity, spatial trajectories, and dis- tances [23]. The stereo version of STARSS23 comprises of 30,000 five-second audio recordings, with the training set containing22.5h of audio data. To enrich the number of real-world direct...

work page

[7] [7]

These include the location-dependent F-score (F20◦/1), class-dependent localization error (LE CD), and class- dependent relative distance error (RDECD)

RESULTS We employ the same validation metrics as used in DCASE 2025 Challenge Task 3. These include the location-dependent F-score (F20◦/1), class-dependent localization error (LE CD), and class- dependent relative distance error (RDECD). In addition, we also cal- culate an aggregated SELD error (ESELD) to provide an overview of the overall performance of...

work page 2025

[8] [8]

We use perceptually-motivated input features to improve both localization and distance estimation performance

CONCLUSION This technical report details our proposed methods for the stereo- based 3D SELD task. We use perceptually-motivated input features to improve both localization and distance estimation performance. We introduce FilterAugment for the 3D SELD task, and propose a stereo-specific form of spectrogram masking augmentation. Over- all, our proposed app...

work page