Pith. sign in

REVIEW 3 major objections 5 minor 30 references

A dynamic convolution layer that reads microphone locations lets one speech-enhancement model work on any array geometry, including arrays never seen in training.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 14:42 UTC pith:5BUOSHOA

load-bearing objection Interesting and cleanly designed geometry-conditioned dynamic convolution for array-invariant SE, but the key attribution to geometry is untested and the reported gains are not consistent across the paper's own tables. the 3 major comments →

arxiv 2607.18658 v1 pith:5BUOSHOA submitted 2026-07-21 eess.AS cs.SD

Towards Array-Invariant Speech Enhancement via Geometry-Aware Dynamic Convolution

Zhenglong Liu , Wangyou Zhang , Chenda Li , Yanmin Qian This is my paper
classification eess.AS cs.SD
keywords array-invariant speech enhancementdynamic convolutionmicrophone coordinatesgeometry-aware processingpermutation equivariancemulti-channel speech enhancementtransformerreal-recorded dataset
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 paper aims to show that the explicit geometry of a microphone array—something array-agnostic methods typically discard—can be fed into a standard fixed-array speech-enhancement model through a dynamic convolution. It proposes Geo-DConv, a layer that reads microphone coordinates and generates the convolution weights on the fly, turning models like SpatialNet and TF-GridNet into systems that accept any number or layout of microphones. On real recordings, the adapted models match or beat prior array-agnostic models and generalize to an unseen six-microphone geometry without retraining. If the claim holds, it offers a practical route to a single enhancement system that can be deployed across devices with different arrays and trained jointly on data with heterogeneous geometries.

Core claim

The paper's central claim is that a geometry-aware dynamic convolution layer can convert fixed-array speech-enhancement models into array-invariant systems. Microphone coordinates are Fourier-encoded and passed through a small transformer (TACT) that predicts a transformation matrix. This matrix linearly combines a set of basis convolution kernels to produce a geometry-specific kernel for each input array. Because the coordinate transformer is permutation-equivariant, the output is stable under any ordering of the microphones. The authors show that two state-of-the-art fixed-array models, fitted with this adapter, can be trained on random sub-arrays of one to four microphones and still perfo

What carries the argument

Geo-DConv (Geometry-Aware Dynamic Convolution): the first-layer convolution has a fixed set of basis kernels; its actual weights are generated per-array by linearly combining the basis kernels with coefficients output by a Topology-Aware Coordinate Transformer (TACT). TACT takes the microphone coordinates, applies Fourier positional encoding, and runs a multi-head self-attention transformer to produce the coefficient matrix in a permutation-equivariant way, which guarantees that the convolution output is independent of the arbitrary ordering of input channels.

Load-bearing premise

The paper attributes the performance improvement to explicitly leveraging array geometry, but it never runs a control with random or constant coordinates, so the claimed causal role of geometry is not established by the experiments.

What would settle it

Train the same Geo-DConv-equipped models on the same data but feed them a fixed, array-independent coordinate matrix (or random coordinates) instead of the true microphone positions. If SI-SDR and PESQ on the same test arrays do not drop, then the geometry signal is not the source of the improvement, and the paper's central claim would be falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fixed-array speech-enhancement architectures can be reused without redesign to handle variable microphone counts and geometries.
  • Training data from different devices and array layouts can be pooled into one dataset, since the model no longer needs a fixed input channel configuration.
  • Permutation invariance is obtained automatically, removing the need for special input-order preprocessing or averaging over channel permutations.
  • Training on real-recorded multi-channel data becomes practical for array-invariant systems, potentially closing the simulation-to-real gap.
  • The approach extends to arrays with more microphones than seen in training, as demonstrated by generalization from four to six microphones.

Where Pith is reading between the lines

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

  • Without an ablation that replaces microphone coordinates with constant, random, or identity vectors while keeping all other components unchanged, the reported gains could stem from the added capacity of the dynamic convolution rather than from geometry. A simple control experiment would settle this attribution.
  • The same adapter mechanism might transfer to other multi-channel tasks where permutation and geometry matter, such as speaker separation, source localization, or beamforming, provided their front-end convolutions are similarly replaced.
  • Because TACT operates on the coordinate set alone, it could be trained with simulated noiseless geometry and then imported into a downstream enhancer, reducing the need for a large real-recorded multi-array corpus.
  • The paper does not address whether the dynamic kernel generation generalizes to radically different array topologies (e.g., line arrays vs. circular arrays) when trained on only one topology; this is a testable limitation.

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 / 5 minor

Summary. The paper proposes Geo-DConv, a module that generates dynamic convolution weights from microphone coordinates. The coordinates are Fourier-encoded, processed by a transformer (TACT) to produce a per-array coefficient matrix, and used to linearly combine learned basis kernels into a per-input-channel convolution kernel. This is designed to let fixed-array SE models (SpatialNet, TF-GridNet) accept an arbitrary number of microphones while remaining permutation-invariant. Experiments on RealMAN compare against array-agnostic baselines (FaSNet-TAC, USES2-comp) and include a zero-shot CHiME-4 evaluation. The central claim is that explicit geometry converts fixed-array models into array-invariant systems with consistent performance improvements.

Significance. If validated, the contribution is practically useful: a lightweight adapter that imports strong fixed-array SE models into variable-geometry deployment, with a clean permutation-equivariance argument and a real-recorded cross-dataset test. The paper is transparent about using real-recorded data and controls the fixed-vs-random array training setting for the base models. However, the causal role of geometry is never isolated, and the headline performance claim is not uniformly supported by the reported numbers. The missing ablation is load-bearing for the title, abstract, and conclusion.

major comments (3)
  1. [§2.2–2.4, Eq. (2)–(3); §4] The central claim that gains come from 'explicitly leveraging array geometry' is not controlled. No experiment replaces the coordinate input G in Eq. (3) with constant, random, or index-only inputs while keeping the TACT and basis-kernel machinery intact. The §4.1 fixed-vs-random control is applied to SpatialNet/TF-GridNet, not to the Geo-DConv variants. Therefore the reported improvements over FaSNet-TAC and USES2-comp in Tables 1–2 could stem from added dynamic capacity or from training on variable channel counts rather than from geometry. This is load-bearing because the title, abstract, and conclusion attribute the gains to explicit geometry.
  2. [§4.2, Tables 1–2] The abstract and §5 say 'consistent performance improvements' over array-invariant approaches, but the paper's own numbers contradict this. In Table 1 (Geometry-Invariant), TF-GridNet-Geo-DConv has SI-SDR 3.90, below USES2-comp's 4.17. In Table 2 (5-mic), SpatialNet-Geo-DConv (4.65) and TF-GridNet-Geo-DConv (4.54) both trail USES2-comp (4.91). The proposed methods improve PESQ/OVRL, but the strong wording is not supported. Please qualify the claim by metric and condition, or provide an analysis of the SI-SDR gap.
  3. [§4.2, Table 2] Cross-dataset CHiME-4 generalization is a strength, but only DNSMOS OVRL is reported (2.55/2.64/2.73 versus 1.42 unprocessed). The sentence 'can generalize well... confirms general geometry-aware spatial patterns' is under-supported without standard SE metrics (SI-SDR, PESQ, STOI) or an array-agnostic baseline evaluated under the same zero-shot protocol. Please add these numbers or temper the conclusion.
minor comments (5)
  1. [§4.2] Typo: 'previous FaSNet-TAC previous FaSNet-TAC' appears twice in the same sentence. Table 2 header also writes 'ChiME-4' instead of 'CHiME-4'.
  2. [§3.2] The SI-SDR citation [18] points to a 2025 study of SI-SDR with noisy references; the original SI-SDR definition should be cited (Le Roux et al., ICASSP 2019) unless a different variant is used.
  3. [§2.2–2.3] Coordinate preprocessing is underspecified. The text mentions spherical coordinates, but not the convention (azimuth/elevation/radius), normalization range, or whether the coordinates are centered at the array centroid. This affects reproducibility and the PE frequency scaling.
  4. [§3.2] No training recipe is given (optimizer, learning rate, epochs, batch size, seed). Also no code or model checkpoints are provided. Please add reproducibility details.
  5. [§4] No variance or significance testing is reported. Many key differences are small (e.g., 0.05–0.1 in PESQ/OVRL); error bars or significance tests for the main comparisons would strengthen the claims.

Circularity Check

0 steps flagged

No significant circularity: the zero-shot CHiME-4 evaluation and the permutation-stability proof are independent of any fitted parameter.

full rationale

This is an empirical architecture paper rather than a derivation. The core array-invariance claim rests on two independent supports. First, the permutation-stability argument is a direct proof from stated assumptions: Fourier positional encoding is pointwise, MHSA is permutation-equivariant, and the resulting dynamic convolution satisfies P X ⊛ P W_dyn = X ⊛ W_dyn. This is not a fitted result. Second, the generalization claim is tested on a genuine out-of-distribution benchmark: models trained on RealMAN with at most 4 microphones are evaluated on the 6-microphone CHiME-4 test set without fine-tuning. The improvement from 1.42 to 2.64/2.73 OVRL is a measured external result, not a quantity that was fit to CHiME-4. The self-references, principally USES2-comp [5] and [11], supply baseline methods and background, not the load-bearing derivation; they are published works with their own evaluations and do not by themselves justify the architecture's performance. The main weakness is that the causal attribution to explicit array geometry is not tested by an ablation that replaces coordinates with constant or random inputs; however, that is a missing control for a mechanistic claim, not a circularity in the sense of an equation reducing to its own input or a fitted parameter being renamed as a prediction. No step in the paper makes the output equivalent to the input by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

All free variables are architectural hyperparameters; no physical constants are fitted. The lowest-support ingredient is the assumption that a small learned basis can represent geometry-dependent filters, together with the unverified coordinate normalization. No new physical entities are introduced.

free parameters (7)
  • basis kernel dimension b = 8
    Hand-set in Sec. 3.2; bounds the rank of the geometry-conditioned weight space; no ablation reported.
  • output channel size O = 16
    Hand-set fixed output dimension passed to the downstream fixed-array model.
  • Fourier PE frequency bands L = 6
    Hand-set coordinate encoding resolution; no sensitivity analysis.
  • TACT hidden dimension d_hidden = 64
    Hand-set transformer width; no ablation.
  • TACT transformer layers/heads = 2 layers / 4 heads
    Hand-set depth and heads; no ablation.
  • coordinate system = spherical
    Choice of spherical over Cartesian coordinates is a modeling decision; no comparison reported.
  • coordinate scale/normalization
    Fourier PE uses sin/cos of raw coordinates; the scale of coordinates before encoding is not specified, and the results are sensitive to it.
axioms (6)
  • standard math Multi-head self-attention over a sequence of coordinate tokens is permutation-equivariant with respect to token order.
    Used in Sec. 2.3 to assert that permuting G permutes M; assumes no extra positional encoding beyond the coordinates themselves.
  • standard math Convolution is invariant under joint permutation of input channels and corresponding weights: (PX) ⊛ (PW_dyn) = X ⊛ W_dyn.
    Used in Sec. 2.3 to establish that the output feature map does not depend on microphone ordering.
  • ad hoc to paper The b-dimensional learned basis kernels can span the effective array-specific filter needed for any geometry.
    Eq. 2 restricts W_dyn to a linear combination of b fixed kernels; no expressiveness or capacity argument is offered beyond final performance.
  • domain assumption Microphone coordinates are known exactly and share the same coordinate conventions across training and evaluation arrays.
    Geometry input is only as good as the annotations; coordinate noise or convention mismatch between RealMAN and CHiME-4 is not analyzed.
  • domain assumption RealMAN's direct-path-filtered speech is a valid clean target for enhancement evaluation.
    The target is obtained by filtering source speech through an estimated direct-path propagation filter rather than measured anechoically, as stated in Sec. 3.1.
  • domain assumption Training on sub-arrays of at most 4 microphones from one 32-mic array transfers to 5/6-microphone arrays with different geometry.
    Table 2's cross-topology and CHiME-4 generalization results depend on this extrapolation.

pith-pipeline@v1.3.0-alltime-deepseek · 7956 in / 15082 out tokens · 121804 ms · 2026-08-01T14:42:51.668779+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards Array-Invariant Speech Enhancement via Geometry-Aware Dynamic Convolution." pith.science (2026). https://pith.science/paper/5BUOSHOA

@misc{pith2026260718658,
  author       = {Pith},
  title        = {Pith review of: Towards Array-Invariant Speech Enhancement via Geometry-Aware Dynamic Convolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5BUOSHOA}},
  note         = {Machine review of arXiv:2607.18658}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-channel speech enhancement (SE) systems exhibit superior performance over single-channel methods but are constrained to fixed microphone array configurations. This restricts their real-world deployment across devices with diverse array geometries. While recent array-agnostic SE methods address variable microphone numbers and permutations, they largely fail to exploit explicit array geometry priors when available, missing a crucial cue for optimal spatial filtering. A Geometry-Aware Dynamic Convolution (Geo-DConv) framework is proposed, which explicitly leverages microphone coordinates to transform standard fixed-array SE models into robust array-invariant systems. Experiments are conducted on the recent real-recorded RealMAN multi-channel speech dataset. Results demonstrate that the proposed architecture enables two widely used fixed-array models to adapt to array-invariant settings, with consistent performance improvements across diverse array topologies.

Figures

Figures reproduced from arXiv: 2607.18658 by Chenda Li, Wangyou Zhang, Yanmin Qian, Zhenglong Liu.

Figure 1
Figure 1. Figure 1: Architecture of the proposed Geo-DConv. The left part illustrates the Geometry-Aware Dynamic Kernel Gener￾ation based on the Topology-Aware Coordinate Transformer (TACT). The right part shows the Dynamic Convolution pro￾cess applied to input acoustic features. we avoid the common issue of mismatch between simulated data and real-world environments in array-based speech en￾hancement. 2. Methods In common fi… 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

30 extracted references · 4 linked inside Pith

  1. [1]

    The ge- ometric variations across different arrays make it difficult to merge existing datasets into a unified, large-scale training cor- pus

    Introduction While multi-channel speech enhancement (SE) theoretically of- fers a higher performance upper bound than single-channel approaches, its real-world applicability is often hindered by its reliance on fixed microphone array geometries. The ge- ometric variations across different arrays make it difficult to merge existing datasets into a unified,...

    Pith/arXiv arXiv 2026

  2. [2]

    Input features are stacked along the channel dimension, and inter-channel infor- mation is fused within the layer to form the output [10, 11]

    Methods In common fixed-array SE, cross-channel modeling typically begins at the very first convolutional layer. Input features are stacked along the channel dimension, and inter-channel infor- mation is fused within the layer to form the output [10, 11]. However, a fundamental limitation of conventional convolu- tional layer is its strict requirement for...

  3. [3]

    Experiments 3.1. Datasets Simulated microphone-array datasets often suffer from se- vere real-world domain mismatch, which leads to perfor- mance degradation when models trained on such data are de- ployed in real acoustic scenarios. To improve model gener- alization, the Real-recorded and Annotated Microphone Array Speech&Noise (RealMAN) dataset [14] is ...

  4. [4]

    Impact of Fixed vs

    Results and Analysis 4.1. Impact of Fixed vs. Random Array Training Two data-feeding strategies are used to evaluate the impact of array geometry configuration during training: utilizing a ran- dom 4-mic array versus a fixed-geometry 4-mic array. During the testing phase, all models are evaluated on the same fixed- geometry 4-mic array (microphone indices...

  5. [5]

    Conclusions This paper reveals that while traditional fixed-array algorithms are inherently constrained by specific geometries, they can more effectively exploit spatial information. Building upon these in- sights, we propose the Geometry-Aware Dynamic Convolution module, which not only addresses the limitation of conventional convolutional layers for var...

  6. [6]

    U25A20409, and in part by SJTU Med-X (Medicine & Engineering) Translational Research Grant (YG2025LC09)

    Acknowledgments This work was supported in part by China NSFC project under Grants No. U25A20409, and in part by SJTU Med-X (Medicine & Engineering) Translational Research Grant (YG2025LC09)

  7. [7]

    All scientific content, ideas, analysis, and conclusions are original and fully authored by the researchers

    Generative AI Use Disclosure Generative AI tools were used solely for language polishing and grammatical improvement in the writing process. All scientific content, ideas, analysis, and conclusions are original and fully authored by the researchers. The authors take full responsibility for the final manuscript

  8. [8]

    SenSE: Semantic-aware high-fidelity universal speech enhancement,

    X. Li, H. Xie, Z. Wang, Z. Zhang, L. Xiao, and L. Xie, “SenSE: Semantic-aware high-fidelity universal speech enhancement,”

  9. [9]

    DNN-based geometry- invariant DOA estimation with microphone positional encoding and complexity gradual training,

    M.-S. Baek, J.-H. Chang, and I. Cohen, “DNN-based geometry- invariant DOA estimation with microphone positional encoding and complexity gradual training,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 2360–2376, 2025

    2025

  10. [10]

    AnyEnhance: A unified generative model with prompt-guidance and self-critic for voice enhancement,

    J. Zhang, J. Yang, Z. Fang, Y . Wang, Z. Zhang, Z. Wang, F. Fan, and Z. Wu, “AnyEnhance: A unified generative model with prompt-guidance and self-critic for voice enhancement,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 3085–3098, 2025

    2025

  11. [11]

    End-to- end microphone permutation and number invariant multi-channel speech separation,

    Y . Luo, Z. Chen, N. Mesgarani, and T. Yoshioka, “End-to- end microphone permutation and number invariant multi-channel speech separation,” inIEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2020, pp. 6394– 6398

    2020

  12. [12]

    TPARN: Triple-path attentive recurrent network for time-domain multichannel speech enhancement,

    A. Pandey, B. Xu, A. Kumar, J. Donley, P. Calamia, and D. Wang, “TPARN: Triple-path attentive recurrent network for time-domain multichannel speech enhancement,” inIEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6497–6501

    2022

  13. [13]

    Improving design of in- put condition invariant speech enhancement,

    W. Zhang, J.-w. Jung, and Y . Qian, “Improving design of in- put condition invariant speech enhancement,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10 696–10 700

    2024

  14. [14]

    AmbiDrop: Array-agnostic speech enhancement using ambisonics encoding and dropout- based learning,

    M. Tatarjitzky and B. Rafaely, “AmbiDrop: Array-agnostic speech enhancement using ambisonics encoding and dropout- based learning,” Sep. 2025. [Online]. Available: http://arxiv.org/ abs/2509.14855

    Pith/arXiv arXiv 2025

  15. [15]

    UniArray: Unified spectral-spatial modeling for array-geometry-agnostic speech separation,

    W. Chen, J. Zhang, J. Yang, E. S. Chng, and X. Zhong, “UniArray: Unified spectral-spatial modeling for array-geometry-agnostic speech separation,”IEEE Signal Processing Letters, vol. 32, pp. 2164–2168, 2025

    2025

  16. [16]

    A memory-based gravitational search al- gorithm for enhancing minimum variance distortionless response beamforming,

    S. Darzi, T. Sieh Kiong, M. Tariqul Islam, H. Rezai Soleyman- pour, and S. Kibria, “A memory-based gravitational search al- gorithm for enhancing minimum variance distortionless response beamforming,”Applied Soft Computing, vol. 47, pp. 103–118, 2016

    2016

  17. [17]

    TF-GridNet: Integrating full- and sub-band modeling for speech separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watan- abe, “TF-GridNet: Integrating full- and sub-band modeling for speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 3221–3236, 2023

    2023

  18. [18]

    Beam-TasNet: Time-domain audio separation net- work meets frequency-domain beamformer,

    T. Ochiai, M. Delcroix, R. Ikeshita, K. Kinoshita, T. Nakatani, and S. Araki, “Beam-TasNet: Time-domain audio separation net- work meets frequency-domain beamformer,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 6384–6388

    2020

  19. [19]

    Closing the gap between time-domain multi-channel speech enhancement on real and simulation conditions,

    W. Zhang, J. Shi, C. Li, S. Watanabe, and Y . Qian, “Closing the gap between time-domain multi-channel speech enhancement on real and simulation conditions,” inIEEE Workshop on Applica- tions of Signal Processing to Audio and Acoustics (WASPAA), 2021, pp. 146–150

    2021

  20. [20]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” inEuropean Conference on Computer Vision (ECCV), 2020

    2020

  21. [21]

    Learning neural acoustic fields,

    A. Luo, Y . Du, M. Tarr, J. Tenenbaum, A. Torralba, and C. Gan, “Learning neural acoustic fields,” inAdvances in Neural Informa- tion Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Asso- ciates, Inc., 2022, pp. 3165–3177

    2022

  22. [22]

    RealMAN: A real-recorded and anno- tated microphone array dataset for dynamic speech enhancement and localization,

    B. Yang, C. Quan, Y . Wang, P. Wang, Y . Yang, Y . Fang, N. Shao, H. Bu, X. Xu, and X. Li, “RealMAN: A real-recorded and anno- tated microphone array dataset for dynamic speech enhancement and localization,” inAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Pa- quet, J. Tomczak, and C. Zhang, Eds., vol. ...

    2024

  23. [23]

    Music source separation with band-split RNN,

    Y . Luo and J. Yu, “Music source separation with band-split RNN,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 1893–1901, 2023

    1901

  24. [24]

    SpatialNet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,

    C. Quan and X. Li, “SpatialNet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1310–1323, 2024

    2024

  25. [26]

    A study of the scale invariant signal to distortion ratio in speech separation with noisy references,

    S. D. Jepsen, M. G. Christensen, and J. R. Jensen, “A study of the scale invariant signal to distortion ratio in speech separation with noisy references,”arXiv preprint arXiv:2508.14623, 2025. [Online]. Available: https://arxiv.org/abs/2508.14623

    Pith/arXiv arXiv 2025

  26. [27]

    Perceptual eval- uation of speech quality (PESQ)-a new method for speech qual- ity assessment of telephone networks and codecs,

    A. Rix, J. Beerends, M. Hollier, and A. Hekstra, “Perceptual eval- uation of speech quality (PESQ)-a new method for speech qual- ity assessment of telephone networks and codecs,” inIEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 2, 2001, pp. 749–752 vol.2

    2001

  27. [28]

    A short- time objective intelligibility measure for time-frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “A short- time objective intelligibility measure for time-frequency weighted noisy speech,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2010, pp. 4214–4217

    2010

  28. [29]

    DNSMOS P. 835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. Reddy, V . Gopal, and R. Cutler, “DNSMOS P. 835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” inIEEE international conference on acous- tics, speech and signal processing (ICASSP). IEEE, 2022, pp. 886–890

    2022

  29. [30]

    An analysis of environment, microphone and data simulation mismatches in robust speech recognition,

    E. Vincent, S. Watanabe, A. A. Nugraha, J. Barker, and R. Marxer, “An analysis of environment, microphone and data simulation mismatches in robust speech recognition,”Computer Speech & Language, vol. 46, pp. 535–557, 2017

    2017

  30. [2025]

    Available: https://arxiv.org/abs/2509.24708

    [Online]. Available: https://arxiv.org/abs/2509.24708

    Pith/arXiv arXiv