Pith. sign in

REVIEW 3 major objections 4 minor 29 references

Class-Aware Permutation-Invariant Signal-to-Distortion Ratio for Semantic Segmentation of Sound Scene with Same-Class Sources

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A class-aware permutation-invariant loss function enables label-queried source separation to handle mixtures containing multiple sources of the same class, while preserving performance on distinct-label mixtures.

desk verdict A solid, clearly-defined extension of PIT and CA-SDRi to duplicated labels in LQSS/S5, with clean math but empirical support that is narrower than the abstract suggests. read the letter →

arxiv 2601.22504 v1 pith:O6PVFN6J submitted 2026-01-30 eess.AS

Binh Thien Nguyen , Masahiro Yasuda , Daiki Takeuchi , Daisuke Niizumi , Noboru Harada This is my paper
classification eess.AS
keywords sourceseparationlabel-queriedpermutationinvarianttrainingsame-classsourcesaudiotaggingevaluationmetricsoundscenesegmentationspatial
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 tackles a blind spot in label-queried source separation: when a sound scene contains two or more sources of the same class (two speakers, two alarms), the repeated label makes the training target ambiguous, so models either merge the sources into one or emit near-identical copies. The authors propose a loss, CA-PI-SDR, that minimizes reconstruction error over all permutations that swap outputs only within the same class, resolving the ambiguity while still using label information to keep distinct classes aligned. The same class-aware permutation principle is carried into a new evaluation metric, CA-PI-SDRi, which jointly scores label prediction and separation quality for mixtures with or without duplicated labels. They also modify the audio-tagging front end to emit repeated labels. On a synthetic multi-channel test set, the proposed loss matches the distinct-label baseline and clearly improves same-class mixtures.

What carries the argument

CA-PI-SDR loss (Eq. 3): the average SDR between estimated and reference sources, minimized over the set of permutations that reorder outputs only within each class group. This constrained search removes the label-order ambiguity caused by duplicated labels while preserving the label-to-output correspondence for distinct classes. The evaluation metric CA-PI-SDRi extends the same principle to scoring, handling unequal numbers of estimated and reference sources per label via true/false positive counts.

What would settle it

Take a trained model and test it on mixtures where two sources of the same class arrive from nearly the same direction (within a few degrees) and check whether the outputs are two distinct waveforms or two near-identical copies; observing the latter would show the loss resolves label ambiguity only for well-separated same-class sources.

Watch

Extended reading notes

Core claim

The central claim is that restricting permutation-invariant training to permutations that keep class labels fixed (Eq. 3) is sufficient to resolve the ambiguity of duplicated label queries in label-queried source separation. The loss takes the minimum average SDR over the constrained permutation set, so same-class outputs can be matched to same-class references in the best way, while cross-class outputs stay aligned to their labels. The resulting model outperforms both the baseline class-aligned loss (which randomly maps same-class sources and often emits near-identical copies) and full permutation-invariant training (which ignores labels and loses distinct-label performance). The companion

Load-bearing premise

The experimental support assumes that same-class sources are always separated by at least 60 degrees in direction of arrival; if real scenes place two same-class sources close together, the reported separation gains may not hold, even though the loss and metric themselves are defined independently of that assumption.

Editorial extensions

If this is right

  • LQSS models can now be trained with queries that repeat a class label, producing genuinely distinct sources instead of a merged or duplicated output.
  • The same class-aware permutation trick works for any loss-based alignment in separation training, so it can be dropped into other label-queried extraction architectures.
  • CA-PI-SDRi gives a single number that reflects both label accuracy and separation quality on mixtures both with and without same-class sources, and it reduces to the existing metric in the distinct-label case.
  • The modified audio-tagging model can output repeated labels, removing the mutually-exclusive-label constraint that limited previous S5 systems.

Reading between the lines

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

  • The constrained-permutation idea is likely to transfer beyond S5 to sound event localization/detection and to any query-by-class extraction system where the same class can occur multiple times; one testable extension is to apply it to binaural or array-based target extraction with more than three instances.
  • The 60-degree minimum separation used in the experiments means the gains are demonstrated for spatially well-separated same-class sources; real mixtures with overlapping or near-coincident sources may be a harder regime where the loss may not prevent the model from merging them.
  • Because the metric averages over true and false predictions and the penalties are set to zero, the proposed score still penalizes missed or spurious sources through the denominator; tuning the false-positive/false-negative penalties could turn it into a tunable metric for different application trade-offs.
  • A natural next step (not in the paper) is to use the CA-PI-SDR loss as a training objective in an end-to-end joint label-prediction and separation model, which the paper notes as a future direction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper addresses the problem of spatial semantic segmentation of sound scenes when a mixture contains multiple sources of the same semantic class. It proposes (i) a class-aware permutation-invariant signal-to-distortion ratio (CA-PI-SDR) loss, Eq. (3), which trains the ResUNetK separation model by minimizing the average SDR over permutations that preserve class labels, and (ii) a corresponding evaluation metric CA-PI-SDRi, Eqs. (5)–(12), that extends CA-SDRi to handle duplicate labels. The authors also modify the M2D audio tagging model to output repeated labels. Experiments on synthesized mixtures compare the new loss with the baseline label-aligned loss (L_CA-SDR) and with conventional PIT, and evaluate the metric on full S5 systems with varying label accuracy and separation quality.

Significance. The proposed CA-PI-SDR loss is a natural and well-motivated extension of permutation-invariant training; it is parameter-free apart from optional penalties and reduces to the standard label-based alignment when all labels are distinct. The metric CA-PI-SDRi is a principled extension of CA-SDRi and is shown to correlate with both label prediction accuracy and separation quality. The release of source code as part of a challenge baseline is a concrete contribution to the community. The central idea is defensible, but the empirical support is limited by the constrained experimental setup (fixed minimum DOA separation) and by the lack of statistical validation, which weakens the strength of the claims as stated.

major comments (3)
  1. [Sec. 4.1] The experimental setup constrains all same-class target sources in both training and test mixtures to have directions of arrival differing by at least 60 degrees. This means the model is never tested on same-class sources that are spatially close or heavily overlapped, which is precisely the regime where the permutation ambiguity is most problematic and where model collapse is most likely. Consequently, the results in Sec. 4.3 and the conclusions in Sec. 5 overstate the practical benefit of CA-PI-SDR for real-world mixtures with arbitrary source positions. Please add experiments with smaller DOA separations (e.g., 0–30 degrees), or explicitly restrict the claim to well-separated same-class sources.
  2. [Sec. 4.3 / Fig. 4] The comparison of the three loss functions is based on single runs without any estimate of variance. The observed differences, particularly between L_CA-PI-SDR and L_CA-SDR on NoDupSet and between L_CA-PI-SDR and L_PI-SDR on DupSet, may fall within run-to-run variability. To support the central claim that the proposed loss combines the strengths of both baselines, report averages over multiple random seeds (at least three) with standard deviations or confidence intervals, or provide statistical significance tests.
  3. [Sec. 4.3, item 2] The description of the L_CA-SDR baseline is incomplete. The authors state that it performs a 'random mapping' in the case of same-class elements, but do not specify how the random permutation is sampled (e.g., independently each batch, per epoch, or fixed once). Since this baseline is the key comparison for demonstrating the benefit of the min operation in the proposed loss, the implementation details must be reported to ensure the comparison is fair and reproducible.
minor comments (4)
  1. [Figs. 4 and 5] The figures lack readable numerical axes. The x- and y-axis labels are present, but tick marks and numerical values are missing, making it impossible to judge the magnitude of the reported differences. Add proper axis scales to both figures.
  2. [Sec. 3.1] The modified AT model is described as outputting 'multiple one-hot vectors,' but the output dimensionality and how the model determines the number of sources (or handles zero-padding for mixtures with fewer than K_max sources) are not specified. This detail is needed for reproducibility of the full S5 system.
  3. [Sec. 3.3.2 / Eq. (10)] The text says that sources are paired to 'maximize the average metric,' but Eq. (10) contains a sum of SDRi values, not an average; the averaging occurs in Eq. (12) through the denominator. Please rephrase to avoid confusion.
  4. [Sec. 3.3.2] The penalty values P_FN and P_FP are set to zero with reference to [5], but the effect of non-zero penalties is not discussed. A brief comment on how these penalties affect the metric's behavior would be useful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the loss and metric are definitional extensions of standard SDR/PIT and are validated on independent data.

full rationale

The derivation chain is self-contained. The CA-PI-SDR loss (Eq. 3) is defined as a constrained permutation-invariant objective over S^C_K using standard SDR (Eq. 4); it is not fitted to the data and does not assume the conclusion that duplicate-label queries can be separated. The CA-PI-SDRi metric (Eqs. 5-12) is an explicit definitional extension of the cited CA-SDRi, with TP/FN/FP counting and per-label permutation-invariant pairing; its reduction to CA-SDRi for distinct-label mixtures is a consistency property, not a disguised reuse of the target. Experimental support comes from independently generated mixtures using external toolkits and data ([4], [21], [22]), with the separation model trained under oracle labels and evaluated on held-out test sets. Self-citations to the authors' baseline [5] and task setup [4] provide the architecture and the metric being extended, but they are not invoked as proof of the new method's effectiveness, and no uniqueness theorem or ansatz is imported via self-citation. The only in-scope limitation is external validity, not circularity: Sec. 4.1 restricts same-class target sources to directions of arrival differing by at least 60 degrees, and Sec. 5 openly flags the AT model as limited; these affect how broadly the empirical gains transfer, but the proposed loss and metric themselves do not reduce to their inputs by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities or fitted scientific constants. The main load-bearing choices are the exchangeability assumption for same-class sources, the zero penalty values in the metric, and the synthetic data constraints (K_max=3, DOA separation >=60 degrees) used in the experiments.

free parameters (3)
  • Penalty values P_FN, P_FP = 0
    Set to 0 following the prior CA-SDRi metric [5]; the choice affects how false predictions are weighted in the final score, though the main comparisons remain valid for any fixed penalty.
  • Maximum number of target sources K_max = 3
    Experimental constraint inherited from DCASE25T4; the DupSet construction and the complexity of the separation/AT models depend on this value.
  • Minimum DOA separation for same-class targets = 60 degrees
    Introduced in Sec. 4.1 for generating DupSet mixtures; this makes same-class separation feasible and likely inflates apparent gains relative to unconstrained real-world mixtures.
assumptions (4)
  • standard math SDR and SDRi as defined in Eqs. (4) and (11) are valid measures of separation quality
    The loss and metric both build on the standard signal-to-distortion ratio; this is a conventional and uncontroversial starting point.
  • domain assumption Same-class sources are exchangeable, so restricting permutations to S^C_K (Eq. 2) resolves all ambiguity from duplicated labels
    This is the core modeling assumption. If same-class sources have systematic identity differences, label-only exchangeability may be suboptimal, though within-class PIT still finds the best assignment.
  • domain assumption Oracle labels used during training transfer to AT predictions at inference
    The separation model is trained with oracle labels as in [5]; the authors themselves concede in the conclusions that the AT model remains a bottleneck for full S5 systems.
  • domain assumption M2D backbone pre-trained on single-channel input remains effective after reshaping channels into the batch dimension
    The authors note this may limit spatial information extraction; the multi-channel improvement in Table 1 is empirical but may not exploit true inter-channel structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Class-Aware Permutation-Invariant Signal-to-Distortion Ratio for Semantic Segmentation of Sound Scene with Same-Class Sources." pith.science (2026). https://pith.science/paper/O6PVFN6J

@misc{pith2026260122504,
  author       = {Pith},
  title        = {Pith review of: Class-Aware Permutation-Invariant Signal-to-Distortion Ratio for Semantic Segmentation of Sound Scene with Same-Class Sources},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O6PVFN6J}},
  note         = {Machine review of arXiv:2601.22504}
}
read the original abstract

To advance immersive communication, the Detection and Classification of Acoustic Scenes and Events (DCASE) 2025 Challenge recently introduced Task 4 on Spatial Semantic Segmentation of Sound Scenes (S5). An S5 system takes a multi-channel audio mixture as input and outputs single-channel dry sources along with their corresponding class labels. Although the DCASE 2025 Challenge simplifies the task by constraining class labels in each mixture to be mutually exclusive, real-world mixtures frequently contain multiple sources from the same class. The presence of duplicated labels can significantly degrade the performance of the label-queried source separation (LQSS) model, which is the key component of many existing S5 systems, and can also limit the validity of the official evaluation metric of DCASE 2025 Task 4. To address these issues, we propose a class-aware permutation-invariant loss function that enables the LQSS model to handle queries involving duplicated labels. In addition, we redesign the S5 evaluation metric to eliminate ambiguities caused by these same-class sources. To evaluate the proposed method within the S5 system, we extend the label prediction model to support same-class labels. Experimental results demonstrate the effectiveness of the proposed methods and the robustness of the new metric on mixtures both with and without same-class sources.

Figures

Figures reproduced from arXiv: 2601.22504 by the authors.

Figure 1
Figure 1. In such cases, correctly separating these sources can be chal [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 linked inside Pith

  1. [1]

    Spatial Semantic Segmentation of Sound Scenes (S5) 1,

    INTRODUCTION Immersive communication has become more practical and widely studied in recent years [1–3], particularly with the introduction of the Immersive V oice and Audio Services (IV AS) codec [1] and the novel parametric spatial audio format, Metadata-Assisted Spatial Audio (MASA) [2]. Its core technologies include decomposing complex spatial sound s...

    arXiv 2025

  2. [2]

    BASELINE S5 SYSTEM This section describes a baseline S5 system of DCASE25T4 [5], the diagram of which is shown in Fig. 2. The input to the S5 system is Y∈R M×T , anM-channel time-domain mixture signal of length T. The output consists ofKlabels,C= (c 1, . . . , cK ), together with their corresponding separated single-channel waveforms at a refer- ence micr...

  3. [3]

    Audio tagging model Figure 3 illustrates the modified M2D AT architecture

    PROPOSED S5 SYSTEM 3.1. Audio tagging model Figure 3 illustrates the modified M2D AT architecture. Instead of predicting a multi-hot vector, we adopt the track-based approach as in [14], which outputs multiple one-hot vectors, thereby enabling re- peated label prediction. The model is trained using PIT with cross- entropy loss. We further exploit spatial ...

  4. [4]

    EXPERIMENTS AND RESULTS 4.1. Experimental setting The multi-channel mixtures were synthesized by convolving target and interfering sound sources with first-order Ambisonics room im- pulse responses (RIRs) and summing them together with multichan- nel background noise. This followed the procedure described in [4], which employed a modified version of the S...

  5. [5]

    The order of input labels is also used to align the estimated and reference sources to calculate the SDR loss function

    and modify the loss function. The order of input labels is also used to align the estimated and reference sources to calculate the SDR loss function. To handle cases where same-class sources exist (i.e., duplicated input labels), we apply PIT to map these same-class output sources to the corresponding reference sources in a manner that minimizes the avera...

  6. [6]

    We also propose an evaluation metric to address the confusion caused by duplicated labels

    CONCLUSIONS In this paper, we present a CA-PI-SDR loss function for training the LQSS model in the S5 system, with the AT model also modified to handle mixtures containing multiple sources of the same class. We also propose an evaluation metric to address the confusion caused by duplicated labels. The experimental results demonstrate the effec- tiveness o...

  7. [7]

    Immersive voice and audio services (IV AS) codec-the new 3GPP standard for immersive communication,

    Markus Multrus, Stefan Bruhn, Juan Torres, Eleni Fotopoulou, Tomas Toftg˚ard, Erik Norvell, Stefan D¨ohla, Yuan Gao, Huan- yu Su, Lasse Laaksonen, et al., “Immersive voice and audio services (IV AS) codec-the new 3GPP standard for immersive communication,” in157th AES Convention, 2024

    2024

  8. [8]

    Metadata-assisted spatial audio coding in IV AS codec,

    Adriana Vasilache, Tapani Pihlajakuja, and Mikko-Ville Laiti- nen, “Metadata-assisted spatial audio coding in IV AS codec,” in2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

Show all 29 references
  1. [9]

    3GPP IV AS codec–perspectives on development, testing and standardiza- tion,

    S Bruhn, T Toftg ˚ard, S D¨ohla, HY Su, L Laaksonen, T Moriya, S Ragot, H Ehara, M Szczerba, I Varga, et al., “3GPP IV AS codec–perspectives on development, testing and standardiza- tion,” in2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)...

    2025

  2. [10]

    De- scription and discussion on DCASE 2025 challenge task 4: Spatial semantic segmentation of sound scenes,

    Masahiro Yasuda, Binh Thien Nguyen, Noboru Harada, Ro- main Serizel, Mayank Mishra, Marc Delcroix, Shoko Araki, Daiki Takeuchi, Daisuke Niizumi, Yasunori Ohishi, et al., “De- scription and discussion on DCASE 2025 challenge task 4: Spatial semantic segmentation of sound scenes...

    2025

  3. [11]

    Baseline systems and evaluation metrics for spatial semantic segmentation of sound scenes,

    Binh Thien Nguyen, Masahiro Yasuda, Daiki Takeuchi, Daisuke Niizumi, Yasunori Ohishi, and Noboru Harada, “Baseline systems and evaluation metrics for spatial semantic segmentation of sound scenes,” in2025 33rd European Signal Processing Conference (EUSIPCO), 2025

    2025

  4. [12]

    Transformer-aided audio source separation with temporal guidance and iterative refine- ment,

    Tobias Morocutti, Florian Schmid, Jonathan Greif, Paul Primus, and Gerhard Widmer, “Transformer-aided audio source separation with temporal guidance and iterative refine- ment,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  5. [13]

    TS-TFGRIDNET: Extend- ing tfgridnet for label-queried target sound extraction via em- bedding concatentaiton,

    Fulin Wu and Zhong-Qiu Wang, “TS-TFGRIDNET: Extend- ing tfgridnet for label-queried target sound extraction via em- bedding concatentaiton,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  6. [14]

    Self-guided target sound extraction and classifi- cation through universal sound separation model and multiple clues,

    Younghoo Kwon, Dongheon Lee, Dohwan Kim, and Jung- Woo Choi, “Self-guided target sound extraction and classifi- cation through universal sound separation model and multiple clues,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  7. [15]

    Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4,

    Jongyeon Park, Joonhee Lee, Do-Hyeon Lim, Hong Kook Kim, Hyeongcheol Geum, and Jeong Eun Lim, “Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4,” Tech. Rep., DCASE2025 Challen...

    2025

  8. [16]

    REDUX: An iterative strategy for semantic source separation,

    Vasileios Stergioulis and Gerasimos Potamianos, “REDUX: An iterative strategy for semantic source separation,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  9. [17]

    SJTU-AUDIOCC system for DCASE 2025 Challenge Task 4: Spatial semantic segmentation of sound scenes,

    Xin Zhou, Hongyu Wang, Chenda Li, Bing Han, Xinhu Zheng, and Yanmin Qian, “SJTU-AUDIOCC system for DCASE 2025 Challenge Task 4: Spatial semantic segmentation of sound scenes,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  10. [18]

    A hybrid S5 system based on neural blind source separation,

    Yuto Nozaki, Shun Sakurai, Yoshiaki Bando, Kohei Saijo, Keisuke Imoto, and Masaki Onishi, “A hybrid S5 system based on neural blind source separation,” Tech. Rep., DCASE2025 Challenge, June 2025

    2025

  11. [19]

    Permutation invariant training of deep models for speaker- independent multi-talker speech separation,

    Dong Yu, Morten Kolbæk, Zheng-Hua Tan, and Jesper Jensen, “Permutation invariant training of deep models for speaker- independent multi-talker speech separation,” in2017 IEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2017, pp. 241–245

    2017

  12. [20]

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

    Yin Cao, Turab Iqbal, Qiuqiang Kong, Fengyan An, Wenwu Wang, and Mark D Plumbley, “An improved event- independent network for polyphonic sound event localization and detection,” in2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 202...

    2021

  13. [21]

    Zero-and few-shot sound event localization and detection,

    Kazuki Shimada, Kengo Uchida, Yuichiro Koyama, Takashi Shibuya, Shusuke Takahashi, Yuki Mitsufuji, and Tatsuya Kawahara, “Zero-and few-shot sound event localization and detection,” in2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE...

    2024

  14. [22]

    Cross-attention inspired selective state space models for tar- get sound extraction,

    Donghang Wu, Yiwen Wang, Xihong Wu, and Tianshu Qu, “Cross-attention inspired selective state space models for tar- get sound extraction,” in2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  15. [23]

    Real-time tar- get sound extraction,

    Bandhav Veluri, Justin Chan, Malek Itani, Tuochao Chen, Takuya Yoshioka, and Shyamnath Gollakota, “Real-time tar- get sound extraction,” in2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  16. [24]

    Soundbeam meets M2D: Target sound extrac- tion with audio foundation model,

    Carlos Hernandez-Olivan, Marc Delcroix, Tsubasa Ochiai, Daisuke Niizumi, Naohiro Tawara, Tomohiro Nakatani, and Shoko Araki, “Soundbeam meets M2D: Target sound extrac- tion with audio foundation model,” in2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Pro...

    2025

  17. [25]

    Universal source separation with weakly labelled data,

    Qiuqiang Kong, Ke Chen, Haohe Liu, Xingjian Du, Taylor Berg-Kirkpatrick, Shlomo Dubnov, and Mark D Plumbley, “Universal source separation with weakly labelled data,”arXiv preprint arXiv:2305.07447, 2023

    2023 arXiv

  18. [26]

    Masked modeling duo: Towards a universal audio pre-training framework,

    Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino, “Masked modeling duo: Towards a universal audio pre-training framework,”IEEE/ACM Trans. on Audio, Speech, and Lang. Process., 2024

    2024

  19. [27]

    Spatial scaper: a library to simulate and augment soundscapes for sound event localiza- tion and detection in realistic rooms,

    Iran R Roman, Christopher Ick, Sivan Ding, Adrian S Roman, Brian McFee, and Juan P Bello, “Spatial scaper: a library to simulate and augment soundscapes for sound event localiza- tion and detection in realistic rooms,” in2024 IEEE Interna- tional Conference on Acoustics, Speec...

    2024

  20. [28]

    Semantic hearing: Programming acoustic scenes with binaural hearables,

    Bandhav Veluri, Malek Itani, Justin Chan, Takuya Yoshioka, and Shyamnath Gollakota, “Semantic hearing: Programming acoustic scenes with binaural hearables,” inProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023, pp. 1–15

    2023

  21. [29]

    Echo-aware adaptation of sound event localization and de- tection in unknown environments,

    Masahiro Yasuda, Yasunori Ohishi, and Shoichiro Saito, “Echo-aware adaptation of sound event localization and de- tection in unknown environments,” in2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 226–230

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.