Pith. sign in

REVIEW 3 major objections 7 minor 29 references

Combining global statistics with segment-level attention yields more accurate automatic MOS prediction across music, speech, and sound.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A dual-branch pooling method, DRASP, combines global statistics with segment-level attention and improves MOS prediction correlation with human ratings.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely useful pooling module for MOS prediction with consistent gains, but the undisclosed segment size and missing error bars keep the headline numbers from being fully reproducible. the 3 major comments →

arxiv 2508.21407 v1 pith:UI66TLTX submitted 2025-08-29 cs.SD cs.AI

DRASP: A Dual-Resolution Attentive Statistics Pooling Framework for Automatic MOS Prediction

Cheng-Yeh Yang , Kuan-Tang Huang , Chien-Chun Wang , Hung-Shin Lee , Hsin-Min Wang , Berlin Chen This is my paper
classification cs.SD cs.AI
keywords mean opinion score predictionattentive statistics poolingsegmental attentiondual-resolution poolingaudio quality assessmenttext-to-music evaluationSpearman rank correlationpooling layer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Automatic mean-opinion-score (MOS) prediction must turn variable-length audio into one vector, and the pooling layer that does this largely determines whether the model hears the whole clip or only its loudest moments. This paper argues that the two standard pooling philosophies—global statistics pooling, which averages everything, and attentive statistics pooling, which weights individual frames—are complementary but individually incomplete. It introduces DRASP, a pooling layer with two branches: one computes global mean and standard deviation over the full utterance, the other splits audio into fixed-size segments and applies attention to those segments instead of to noisy frames. Combined by a learnable fusion, the two views produce a single representation that the paper shows yields higher system-level rank correlation with human ratings across music, speech, and general audio, on two backbones and two datasets. If the results hold, DRASP is a drop-in replacement for the pooling layer in existing MOS predictors that improves ranking of generative audio systems without retraining the whole architecture.

Core claim

The paper's central claim is that a pooling module should represent an audio clip at two resolutions at once. The global branch summarizes the entire utterance with an unweighted mean and standard deviation, capturing overall temporal dynamics. The segmental branch partitions the frame sequence into S = T/n non-overlapping blocks, averages the frames inside each block to form a segment embedding, and then learns attention weights over those segments to compute a weighted mean and standard deviation. A trainable linear fusion, initialized to trust global statistics, combines the two vectors. On MusicEval and the AES-Natural subset, this dual view outperforms average pooling, statistics poolin

What carries the argument

DRASP is a dual-branch pooling layer. The key components are: (1) a global statistics pooling branch that computes mean and standard deviation over all frames; (2) a segmental attentive statistics branch that groups frames into non-overlapping segments of length n, averages each segment, scores segments with a tanh attention layer, softmax-normalizes, and computes a weighted mean and standard deviation; and (3) a linear fusion p = α[μ; σ] + β[μ̃; σ̃], with α and β initialized to 1 and 0 so training starts from global statistics and then adds local emphasis. The segment-level averaging is the load-bearing mechanism: it suppresses frame noise and shortens the attention sequence from T frames t

Load-bearing premise

That a fixed, evenly spaced partition of the audio into n-frame blocks keeps the perceptually important details inside a block where attention can find them; if an artifact straddles a boundary or sits inside a large block, the segment-level branch cannot isolate it and degrades toward global averaging.

What would settle it

Inject a single short, clearly audible distortion into an otherwise clean clip at a known time offset, then compare DRASP's predicted score to the clean original. If the predicted ranking fails to drop when the distortion crosses a segment boundary but drops when the same distortion sits inside one segment, the fixed uniform segmentation is doing the work, not the attention.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Replacing the pooling layer in an existing MOS predictor with DRASP improves system-level SRCC on both MusicEval and the AES-Natural subset, across the CLAP-based and AudioBox-Aesthetics backbones.
  • DRASP beats every single-granularity baseline on MusicEval, including frame-level attentive statistics pooling, which the paper finds can underperform plain statistics pooling.
  • Segment size n is a controllable trade-off: DRASP stays above average pooling for all tested sizes, but too-small segments pick up localized noise and too-large segments blur perceptual events.
  • Because attention runs over S segments rather than T frames, DRASP is computationally cheaper than multi-head and multi-resolution attentive pooling while achieving higher scores in these experiments.
  • The DRASP-augmented CLAP model fine-tuned on about 35 hours of data outperforms the much larger pre-trained AudioBox-Aesthetics baseline, suggesting the module is data-efficient.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the fixed uniform segmentation is the most likely failure point; an adaptive segmenter aligned to phonemes or notes would let attention isolate events that currently straddle boundaries.
  • Editorial inference: because the headline SRCC gain is computed after per-clip predictions are averaged to the system level, part of the gain may be variance reduction across clips; clip-level Spearman is not reported, so per-clip calibration remains untested.
  • Editorial inference: the same dual-branch design should transfer to video or image quality assessment, wherever a sequence of frame or patch embeddings exists, though the paper does not test that.
  • Editorial inference: injecting a brief artificial artifact at controlled positions and inspecting the segment attention weights would directly test the claimed noise-suppression mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes DRASP, a dual-branch pooling mechanism for automatic MOS prediction. One branch computes global statistics (mean and standard deviation) over all frame-level embeddings; the other partitions the sequence into fixed-length segments, averages frames within each segment, applies a learned segment-level attention, and computes attentive statistics. The two branches are fused with trainable scalar weights. DRASP is integrated into two MOS backbones (a CLAP-based model and AudioBox-Aesthetics) and evaluated on MusicEval and a subset of AES-Natural. The reported results show consistent improvements over average, statistics, and attentive pooling baselines, including a 10.39% relative improvement in system-level SRCC over average pooling on MusicEval.

Significance. If the empirical results are robust, DRASP is a simple and modular pooling change that improves MOS prediction across music, speech, and general audio, and across two different upstream architectures. The core idea of applying attention over acoustic segments rather than individual frames is well motivated by the observed instability of frame-level attentive statistics pooling, and the authors provide open-source code. The claims are falsifiable and the experimental scope is broader than many pooling-only papers. The main limitations are that the key segment-size hyperparameter is not reported for the main tables, the AES-Natural evaluation uses an unavailability-driven subset without split details, and no statistical uncertainty or significance testing is provided. These issues currently prevent the "consistent generalization" claim from being fully established.

major comments (3)
  1. [§III-B, Eq. (1) and Fig. 4] The main results in Tables I and II do not state the segment size n used to form S=T/n segments in Eq. (1). Fig. 4 shows that system-level SRCC on MusicEval varies by roughly 0.05 depending on n (e.g., Musical Impression from about 0.90 to 0.95). Since the claimed consistent outperformance across datasets depends on this hyperparameter, the paper must report the n used in Tables I and II, describe how it was selected, and provide a sensitivity analysis on AES-Natural. Without this, the cross-dataset generalization claim is not established; the result could be a tuned hyperparameter rather than a robust mechanism.
  2. [§IV-A, §V-D] The AES-Natural experiments use an unavailability-driven subset of 2,776 clips (950 speech, 932 music, 894 general). The paper does not report how this subset was split into train/validation/test, whether the split is stratified by domain, or how the subset differs from the full benchmark. The data-efficiency claim ('only 35 hours') is also incomplete because the amount of in-domain training data used for the AudioBox-Aesthetics baseline is not stated. These omissions make the generalization and data-efficiency results difficult to verify. Please provide the split, per-domain test sizes, data amounts, and ideally evaluate on the full AES-Natural test set or justify the subset.
  3. [Tables I and II] No variance or significance information is reported. The headline comparisons are single-run system-level correlations over 31 systems (MusicEval) and four dimensions (AES-Natural). Differences such as average SRCC 0.924 vs. 0.897 (Table I) or 0.909 vs. 0.871 (Table II) could be within run-to-run or system-sampling noise. Report multiple seeds (mean ± std) and at least one paired test (e.g., bootstrap over systems) for the main comparisons. This is necessary to support the word 'consistently' in the abstract and to justify the use of 'significantly' in Section V-C.
minor comments (7)
  1. [Eq. (1)] S=T/n assumes n divides T. Clarify how variable-length audio is handled: padding, truncation, or a modified final segment.
  2. [§IV-A] Provide a reproducible link or exact list for the AES-Natural subset, including the original source IDs and any filtering criteria, so the subset can be reconstructed.
  3. [Table I] The baseline 'Segmental Attentive Statistics Pooling' is not described in Section II or the experimental setup. Add a brief description and citation.
  4. [Fig. 4] When n=1, the segmental branch reduces to frame-level attentive statistics pooling. Stating this explicitly would help the reader interpret the n-axis.
  5. [§V-C] Replace 'significantly outperforms' with 'consistently outperforms' unless a statistical test is added, since no significance test is reported.
  6. [Abstract] The 10.39% relative improvement is presented without specifying that it is on MusicEval system-level SRCC. Clarify the scope of this number.
  7. [§V-A] The text says 'underlined in Table I', but the rendered table does not show underlines. Check the formatting or rephrase.

Circularity Check

0 steps flagged

No significant circularity: DRASP's performance claims are empirical and rest on external benchmarks, not on fitting the target into the model definition.

full rationale

The paper's central claim is that DRASP pooling improves MOS prediction over existing pooling methods. This is supported by system-level SRCC/LCC/MSE comparisons on MusicEval and AES-Natural, using two different backbones, with the reported improvements computed directly from Table I. The method's equations (1)-(5) define a pooling transformation with trainable attention and fusion parameters; none of these parameters is fitted to the reported SRCC values and then renamed as a prediction. The segment size n is a free design choice, and Fig. 4 is a sensitivity curve rather than a hidden fit: the paper claims robustness across n, and the undisclosed n used for Tables I and II is a reproducibility weakness, not a circularity. The motivation that frame-level attentive pooling is unstable comes from baseline results in Table I; using that observation to design an architecture and then testing on the same dataset is a possible model-selection bias, but not a definitional reduction. Citations to [16] and [17] point to external benchmarks with human annotations and pretrained models independent of the present paper's fitted values; even if any author overlaps, the cited data are externally falsifiable. No uniqueness theorem, ansatz-smuggling via citation, or renaming of a known result is load-bearing. Overall, the empirical derivation chain is self-contained: the performance claim is separable from the paper's definitions and can be checked against the published data.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim rests on the sufficiency of fixed non-overlapping segmentation and on the quality of upstream embeddings and labels; the reported numbers also depend on the segment size n. No invented physical entities are introduced.

free parameters (1)
  • Segment size n = not specified in main text (swept over 1, 2, ..., 512 in Fig. 4)
    The number of frames averaged into a segment controls the granularity of the segmental attention branch. The main tables (I, II) do not state which n was used; Fig. 4 shows strong dependence of SRCC on n, so the reported numbers depend on this hand-chosen hyperparameter.
axioms (4)
  • domain assumption Pre-trained audio encoders (CLAP, AudioBox-Aesthetics) provide frame embeddings sufficient for MOS prediction.
    Section III-D and IV-B: the entire framework operates on embeddings from these encoders; if the embeddings lack quality-relevant information, DRASP cannot help.
  • domain assumption Fixed-size non-overlapping segmentation preserves perceptually salient local events.
    Equation (1), Section III-B: no adaptive boundaries are used; assumes uniform segments capture the relevant acoustic units.
  • domain assumption System-level aggregation (averaging clip predictions per generative system) is the right evaluation target.
    Section IV-C: the paper adopts the protocol from [29], but does not justify why system-level SRCC is the primary measure beyond practical relevance.
  • domain assumption The AES-Natural subset available at the time of experiments is representative of the full benchmark.
    Section IV-A: the paper uses only 2,776 clips because some original files were unavailable; the subset may not be representative and no coverage analysis is given.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DRASP: A Dual-Resolution Attentive Statistics Pooling Framework for Automatic MOS Prediction." pith.science (2026). https://pith.science/paper/UI66TLTX

@misc{pith2026250821407,
  author       = {Pith},
  title        = {Pith review of: DRASP: A Dual-Resolution Attentive Statistics Pooling Framework for Automatic MOS Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UI66TLTX}},
  note         = {Machine review of arXiv:2508.21407}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A pooling mechanism is essential for mean opinion score (MOS) prediction, facilitating the transformation of variable-length audio features into a concise fixed-size representation that effectively encodes speech quality. Existing pooling methods typically operate at a singular granularity, concentrating either on a comprehensive global perspective or a detailed frame-level analysis, which may overlook complementary perceptual insights. To address this limitation, we introduce the Dual-Resolution Attentive Statistics Pooling (DRASP) framework. DRASP integrates both coarse-grained, global statistical summaries and fine-grained, attentive analyses of perceptually significant segments. This dual-view architecture empowers our model to formulate a more thorough and robust representation, capturing both the overarching structural context and salient local details concurrently. Extensive experiments validate the effectiveness and strong generalization ability of the proposed framework. It consistently outperforms various baseline methods across diverse datasets (MusicEval and AES-Natural), MOS prediction backbones (including a CLAP-based model and AudioBox-Aesthetics), and different audio generation systems, achieving a relative improvement of 10.39% in system-level Spearman's rank correlation coefficient (SRCC) over the widely-used average pooling approach.

Figures

Figures reproduced from arXiv: 2508.21407 by Berlin Chen, Cheng-Yeh Yang, Chien-Chun Wang, Hsin-Min Wang, Hung-Shin Lee, Kuan-Tang Huang.

Figure 1
Figure 1. Figure 1: Overview of the proposed DRASP mechanism. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of two distinct MOS prediction backbones, each enhanced [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of various segment sizes n on system-level SRCC performance of our DRASP mechanism on MusicEval. TABLE II SYSTEM-LEVEL SRCC RESULTS ON THE AES-NATURAL DATASET FOR ASSESSING MODEL GENERALIZATION. Method PQ PC CE CU Avg. AudioBox-Aesthetics 0.877 0.935 0.839 0.832 0.871 CLAP-based + DRASP 0.881 0.938 0.829 0.877 0.881 AudioBox-Aesthetics + DRASP 0.900 0.936 0.890 0.911 0.909 and more linear cluster ar… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [1]

    MusicLDM: Enhancing novelty in text-to-music generation using beat-synchronous mixup strategies,

    K. Chen, Y . Wu, H. Liu, M. Nezhurina, T. Berg-Kirkpatrick, and S. Dubnov, “MusicLDM: Enhancing novelty in text-to-music generation using beat-synchronous mixup strategies,” in Proc. ICASSP, 2024

    work page 2024

  2. [2]

    Masked audio generation using a single non- autoregressive transformer,

    A. Ziv, I. Gat, G. L. Lan, T. Remez, F. Kreuk, A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “Masked audio generation using a single non- autoregressive transformer,” in Proc. ICLR, 2024

    work page 2024

  3. [3]

    V ALL-E 2: Neural codec language models are hu- man parity zero-shot text to speech synthesizers,

    S. Chen, S. Liu, L. Zhou, Y . Liu, X. Tan, J. Li, S. Zhao, Y . Qian, and F. Wei, “V ALL-E 2: Neural codec language models are hu- man parity zero-shot text to speech synthesizers,” in Arxiv preprint arXiv:2406.05370, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,” in Arxiv preprint arXiv:2410.06885 , 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Tango 2: Aligning diffusion-based text-to-audio generations through direct preference optimization,

    N. Majumder, C.-Y . Hung, D. Ghosal, W.-N. Hsu, R. Mihalcea, and S. Poria, “Tango 2: Aligning diffusion-based text-to-audio generations through direct preference optimization,” in Proc. ACM MM, 2024

    work page 2024

  6. [6]

    Fr ´echet audio distance: A metric for evaluating music enhancement algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr ´echet audio distance: A metric for evaluating music enhancement algorithms,” in Arxiv preprint arXiv:1812.08466 , 2019

    Pith/arXiv arXiv 2019

  7. [7]

    Deep neural network embeddings for text-independent speaker verification,

    D. Snyder, D. Garcia-Romero, D. Povey, and S. Khudanpur, “Deep neural network embeddings for text-independent speaker verification,” in Proc. Interspeech, 2017

    work page 2017

  8. [8]

    Attentive statistics pooling for deep speaker embedding,

    K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics pooling for deep speaker embedding,” in Proc. Interspeech, 2018

    work page 2018

  9. [9]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. NeurIPS, 2017

    work page 2017

  10. [10]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,” in Proc. Interspeech, 2020

    work page 2020

  11. [11]

    CAM++: A fast and efficient network for speaker verification using context-aware masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “CAM++: A fast and efficient network for speaker verification using context-aware masking,” in Arxiv preprint arXiv: 2303.00332 , 2023

    Pith/arXiv arXiv 2023

  12. [12]

    V oiceID loss: Speech enhancement for speaker verification,

    S. Shon, H. Tang, and J. Glass, “V oiceID loss: Speech enhancement for speaker verification,” in Proc. Interspeech, 2019

    work page 2019

  13. [13]

    Improved RawNet with feature map scaling for text-independent speaker verifica- tion using raw waveforms,

    J.-w. Jung, S.-b. Kim, H.-j. Shim, J.-h. Kim, and H.-J. Yu, “Improved RawNet with feature map scaling for text-independent speaker verifica- tion using raw waveforms,” in Arxiv preprint arXiv: 2004.00526 , 2020

    Pith/arXiv arXiv 2004

  14. [14]

    ResSKNet-SSDP: Effective and light end-to-end architecture for speaker recognition,

    F. Deng, L. Deng, P. Jiang, G. Zhang, and Q. Yang, “ResSKNet-SSDP: Effective and light end-to-end architecture for speaker recognition,” Sensors, vol. 23, no. 3, p. 1203, 2023

    work page 2023

  15. [15]

    Deep segment attentive embedding for duration robust speaker verification,

    B. Liu, S. Nie, Y . Zhang, S. Liang, and W. Liu, “Deep segment attentive embedding for duration robust speaker verification,” in Proc. APSIPA, 2019

    work page 2019

  16. [16]

    MusicEval: A generative music dataset with expert ratings for automatic text-to-music evaluation,

    C. Liu, H. Wang, J. Zhao, S. Zhao, H. Bu, X. Xu, J. Zhou, H. Sun, and Y . Qin, “MusicEval: A generative music dataset with expert ratings for automatic text-to-music evaluation,” in Proc. ICASSP, 2025

    work page 2025

  17. [17]

    Meta audiobox aesthetics: Unified automatic quality assessment for speech, music, and sound,

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

    Pith/arXiv arXiv 2025

  18. [18]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” in Proc. ICASSP, 2023

    work page 2023

  19. [19]

    Exploring the encoding layer and loss function in end-to-end speaker and language recognition system,

    W. Cai, J. Chen, and M. Li, “Exploring the encoding layer and loss function in end-to-end speaker and language recognition system,” in Proc. Odyssey, 2018

    work page 2018

  20. [20]

    Self-attentive speaker embeddings for text-independent speaker verification,

    Y . Zhu, T. Ko, D. Snyder, B. Mak, and D. Povey, “Self-attentive speaker embeddings for text-independent speaker verification,” in Proc. Interspeech, 2018

    work page 2018

  21. [21]

    Self multi-head attention for speaker recognition,

    M. India, P. Safari, and J. Hernando, “Self multi-head attention for speaker recognition,” in Proc. Interspeech, 2019

    work page 2019

  22. [22]

    Vector-based attentive pooling for text-independent speaker verification,

    Y . Wu, C. Guo, H. Gao, X. Hou, and J. Xu, “Vector-based attentive pooling for text-independent speaker verification,” in Proc. Interspeech, 2020

    work page 2020

  23. [23]

    RSKNet-MTSP: Effective and portable deep architecture for speaker verification,

    Y . Wu, C. Guo, J. Zhao, X. Jin, and J. Xu, “RSKNet-MTSP: Effective and portable deep architecture for speaker verification,” Neurocomput- ing, vol. 511, pp. 259–272, 2022

    work page 2022

  24. [24]

    Multi-resolution multi-head attention in deep speaker embedding,

    Z. Wang, K. Yao, X. Li, and S. Fang, “Multi-resolution multi-head attention in deep speaker embedding,” in Proc. ICASSP, 2020

    work page 2020

  25. [25]

    NetVLAD: CNN architecture for weakly supervised place recognition,

    R. Arandjelovi ´c, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “NetVLAD: CNN architecture for weakly supervised place recognition,” in Proc. CVPR, 2016

    work page 2016

  26. [26]

    GhostVLAD for set- based face recognition,

    Y . Zhong, R. Arandjelovi ´c, and A. Zisserman, “GhostVLAD for set- based face recognition,” in Proc. ACCV, 2018

    work page 2018

  27. [27]

    Neural aggregation network for video face recognition,

    J. Yang, P. Ren, D. Zhang, D. Chen, F. Wen, H. Li, and G. Hua, “Neural aggregation network for video face recognition,” in Proc. CVPR, 2017

    work page 2017

  28. [28]

    Quality aware network for set to set recognition,

    Y . Liu, J. Yan, and W. Ouyang, “Quality aware network for set to set recognition,” in Proc. CVPR, 2017

    work page 2017

  29. [29]

    Generalization ability of MOS prediction networks,

    E. Cooper, W.-C. Huang, T. Toda, and J. Yamagishi, “Generalization ability of MOS prediction networks,” in Proc. ICASSP, 2022

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.