Echo-Aware Modulation for Compact-Latent Frequency-Time Modeling in Lightweight Acoustic Echo Cancellation

Bj\"orn W. Schuller; Cairong Zou; Kai Xie; Qingyun Wang; Ruiyu Liang; Ye Ni

REVIEW 3 major objections 5 minor 29 references

Echo-Aware Modulation for Compact-Latent Frequency-Time Modeling in Lightweight Acoustic Echo Cancellation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a cheap echo-aware modulation block can restore most of the speech quality lost when an acoustic echo cancellation model compresses audio into a compact Bark-domain bottleneck, at a fraction of the compute cost of freq

desk verdict A solid engineering contribution with a real ablation gap: the EAM gains could be generic capacity, and there are no error bars. read the letter →

arxiv 2608.03650 v1 pith:WMEFNOOK submitted 2026-08-04 eess.AS

Ye Ni , Ruiyu Liang , Qingyun Wang , Kai Xie , Cairong Zou , Björn W. Schuller This is my paper · ORCID

classification eess.AS

keywords acousticechocancellationlightweightneuralnetworkBark-domainmodelingfrequency-timeecho-awaremodulationdual-branchencodercompactlatentbottleneckhybridAEC

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 tries to establish that the performance penalty of compact Bark-domain acoustic echo cancellation comes mainly from lost microphone–reference interaction cues, and that a lightweight modulation block can put those cues back. It proposes MSA-EchoLite, a hybrid system where a classical adaptive filter provides an echo reference and a small neural network suppresses residual echo in a compressed Bark-domain latent space. The key addition, the echo-aware modulation (EAM) block, injects the difference and product of microphone and reference features into the bottleneck. The authors report consistent gains across all compact configurations, with the largest improvement at the most compressed setting, and show that the Bark-domain model with EAM approaches the quality of a frequency-domain model needing nearly twice the FLOPs. This matters because it suggests low-resource devices can get near-full-band echo cancellation quality without the usual compute cost.

What carries the argument

The echo-aware modulation (EAM) block is the load-bearing component: Eq. (2) builds an interaction tensor from the fused, microphone, and reference representations plus their difference and element-wise product; Eq. (3) extracts joint time–frequency, temporal-only, and frequency-only cues with three depthwise convolutional branches; Eq. (4) re-weights those cues with a channel gate and a local gate and adds them back to the fused representation through a learnable scaling factor. The difference term flags where microphone and reference do not match, which is where near-end speech lives, while the product term flags where they correlate, which is where echo lives. The gated residual allows th

What would settle it

Retrain the Bark baseline and the Bark+EAM model at the C32H64 configuration under, say, five random seeds using the paper's own protocol. If the PESQ, ESTOI, or SDR difference between the two models across seeds overlaps zero or changes sign, then the claim that EAM consistently improves compact-latent modeling is not supported.

Watch

Extended reading notes

Core claim

The central claim is that information needed to separate speech from echo—namely, where the microphone and echo-reference representations disagree and where they agree—is precisely what Bark-domain compression and encoder downsampling discard, and that a modulation module can cheaply re-inject it. EAM concatenates the fused, microphone, and reference latent features with their pointwise difference S−R and product S⊙R, extracts joint, temporal-only, and frequency-only patterns with three depthwise convolutional branches, applies channel and local gates, and adds the result back through a learnable scale. At the C32H64 configuration, the paper reports that this improves PESQ-WB from 2.02 to 2.

Load-bearing premise

The load-bearing premise is that the reported single-run metric differences, such as PESQ 2.02 versus 2.10, are stable across random initializations and training runs; the paper reports no seeds, variances, or significance tests, and the gains are small enough that run-to-run spread could blur them.

Editorial extensions

If this is right

Bark-domain models with EAM reach 99.1% of the PESQ of a frequency-domain model that uses nearly twice the FLOPs, and surpass it in SDR.
EAM improves every reported compact-latent configuration in the paper, with the largest benefit appearing at the most compressed setting (C32H64).
The full MSA-EchoLite system keeps 0.2 M parameters and about 102 M FLOPs per second while reporting stronger echo suppression than existing lightweight baselines on real blind test sets.
Bark-domain modeling is shown to be more sensitive to channel-dimension reduction than to hidden-dimension reduction, and EAM partially compensates for that sensitivity.

Reading between the lines

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

Editorial inference: Because the degradation mechanism is compression itself, the same difference-and-product interaction injection should transfer to other compact auditory scales such as ERB; this is a testable prediction the paper does not make.
Editorial inference: The EAM block is generic enough to be dropped into any dual-branch echo or noise suppression model that has a target branch and a correlated reference branch, so its usefulness may extend beyond Bark-domain AEC.
Editorial inference: All reported gains are single-run point estimates, so the open question is not whether the mechanism helps in one training run but whether the magnitude of the benefit is stable across random initializations; a multi-seed repetition study is the natural next step.

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

Best read as a careful engineering paper on lightweight Bark-domain AEC. The genuinely new piece is EAM, a bottleneck module that combines difference and product of microphone/reference latents with multi-receptive-field depthwise convs and gating; that combination is not in the prior work they cite. The systematic Bark-vs-frequency comparison under a shared encoder is also useful: it shows Bark dominates the low-complexity Pareto frontier and is more sensitive to channel compression than hidden-size reduction.

What the paper does well: all baselines are retrained under the same protocol and loss, the ablation spans several channel/hidden/Bark-band settings, and +EAM consistently improves every reported metric in every configuration. The real-blind-test comparison against EchoFree, DTLN-AEC, DeepVQE-S, and their own MSA-DPCRN is fair and they disclose reuse of their prior encoder and loss.

Two soft spots, one more fundamental than the other. First, the ablation is not capacity-matched: +EAM adds 33-76% parameters and 26-65% FLOPs, and there is no control that increases baseline capacity similarly without the echo-aware S-R and S*R terms. So the gains could be generic capacity rather than the echo-aware interaction the paper motivates. The data-processing inequality argument in Sec. II-C is generic and doesn't single out the echo-aware structure. This weakens the central novelty claim but does not invalidate the module's practical usefulness. Second, all numbers are single-run point estimates; the headline gains are small (PESQ +0.08, ESTOI +1.01 points), so a few seeds with variance or significance testing would be needed to establish that the effect is real. Minor: no code or artifacts, and CAGCRN, which they cite as a closely related lightweight baseline, is not included in the experiments.

This paper deserves a serious referee. The architecture is plausible, the evaluation is honest in coverage, and the questions it raises about compact-latent FT modeling are useful to the subfield. I would ask for a capacity-matched control and repeated seeds before acceptance. A reader looking for a deployable low-FLOP AEC module gets value here; a reader who cares about mechanism will need the missing control.

Referee Report

3 major / 5 minor

Summary. The paper proposes MSA-EchoLite, a lightweight hybrid acoustic echo cancellation (AEC) system that combines a PBFDAF linear filter with a Bark-domain neural suppressor. The central contribution is an echo-aware modulation (EAM) module inserted in the frequency-time LSTM bottleneck; Eqs. (2)–(4) define an interaction representation built from the fused, microphone, and echo-related latents plus their difference S−R and element-wise product S⊙R, processed through multi-branch depthwise convolutions and gated residual scaling. The paper compares Bark-domain and frequency-domain variants across bottleneck channels C and hidden sizes H, varies the number of Bark bands, and evaluates on the AEC Challenge blind test sets. The authors claim that +EAM consistently improves all compact-latent configurations with modest additional FLOPs, that Bark-domain modeling offers a better performance-complexity trade-off, and that Bark models are more sensitive to channel compression than frequency models.

Significance. If the reported effects are reproducible, the EAM module is a simple and computationally cheap way to improve compact-latent AEC modeling, and the systematic Bark-versus-frequency comparison within a single framework is a useful contribution to lightweight AEC. The paper's strengths include retraining all baselines under a common data and loss protocol, evaluating on real blind test sets, and ablating across multiple Bark-band/channel/hidden-size configurations. The central claims, however, rest entirely on single-run point estimates without variance or significance testing, and the EAM ablation is not capacity-matched; consequently, the specific benefit of the echo-aware interaction over generic added capacity is not yet established.

major comments (3)

[§III-B, Tables I–III] All reported numbers are single-run point estimates with no variance, confidence intervals, or significance tests. The headline EAM gains are small — e.g., at C32H64, PESQ-WB 2.02→2.10, ESTOI +1.01 percentage points, SDR +0.34 dB; at B48 in Table II, +EAM improves PESQ by only 0.02. The central claim that '+EAM consistently improves all configurations' is load-bearing and depends on these differences being larger than run-to-run noise. Please report at least 3–5 independent training runs per configuration, give means and standard deviations, and apply a paired test or equivalent across the Table I and Table II configurations; the Table III blind-test results should also be accompanied by variance information where feasible.
[§III-D, Eq. (2), Table I] The EAM ablation is not capacity-matched. At C32H64, +EAM increases parameters from 0.15M to 0.20M (+33%) and FLOPs from 80.83 to 101.90 M/s (+26.1%); at C64H64 the parameter increase is +76% (0.25M→0.44M) and FLOPs +64.7%. Since the observed improvements are small, they could be caused by the added parameters and compute rather than by the specific S−R and S⊙R terms in Eq. (2). The data-processing-inequality motivation in §II-C only argues for more expressive bottleneck processing; it does not single out echo-aware interactions. Please add a matched-capacity control in which the baseline is given comparable extra parameters/computation without the echo-aware terms (e.g., wider C, an extra DWC/PWC block, or an EAM variant unable to use S−R and S⊙R), and show that the echo-aware structure is what drives the gains.
[§III-C, Fig. 2, Table I] The Bark-versus-frequency analysis and the compression-sensitivity conclusions in Fig. 2 are also derived from single runs and lack uncertainty estimates. For example, the claim that Bark is more sensitive when C is reduced from 64 to 32 (normalized drop 2.47%→5.50%) uses one point per configuration, and the same run-to-run noise concern applies. In addition, the Freq and Bark models in Table I are not parameter-matched (e.g., C64H64: Bark 0.25M, +EAM 0.44M, Freq 0.31M), so the comparison is not holding capacity fixed across domains. Please provide uncertainty bars on the Pareto and normalized-drop analyses and clarify which components are held fixed when comparing Bark and frequency-domain modeling.

minor comments (5)

[Fig. 1] The figure contains typos: 'Covn2d' should likely be 'Conv2d'. The abbreviations RFP and SFP should be introduced before or near the figure caption to improve readability.
[§II-C] The data processing inequality is true but used loosely as a motivation for EAM: EAM operates on the compressed latents and cannot increase I(Z;Y) beyond what the encoder preserves. Consider rephrasing to say that EAM is a more expressive post-encoder transformation that better exploits the information remaining in the bottleneck, rather than one that recovers information lost to compression.
[Eq. (5)–(6)] The reconstruction in Eq. (6) applies a real Bark gain to the complex microphone spectrum. Please clarify the exact multiplication (e.g., whether the gain is applied to the magnitude and the microphone phase is preserved) to avoid ambiguity.
[Table II] The B48 setting shows a very small EAM benefit (PESQ 1.90→1.92). A brief discussion of why the gain is smaller at the lowest Bark-band count would strengthen the ablation narrative.
[General] No code availability statement is included. A public implementation or a detailed reproducibility note would be valuable in a systems-oriented paper of this type.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EAM is an architecturally defined module evaluated on held-out metrics, not a quantity fitted from the targets it predicts.

full rationale

The paper's central claim is that the proposed EAM module improves compact-latent Bark-domain AEC performance. This is an empirical architectural claim. EAM is defined by Eqs. (2)-(4) as a fixed function of the fused, microphone, and echo-related latent representations (X, S, R), and its output is a residual modulation. The module parameters are learned end-to-end on reconstruction losses (Eq. 7), and all reported gains are on held-out test sets. No fitted constant is renamed as a prediction, and no quantity used for training is also reported as the predicted outcome. The data-processing-inequality statement in Section II-C is explicitly motivational ('To mitigate this, we embed EAM...') and does not derive the module's specific form; it only motivates adding capacity in the bottleneck. The self-citation to MSA-DPCRN [5] for the asymmetric dual-branch encoder is disclosed as prior work and is not load-bearing for the EAM-specific novelty: the encoder design is a starting point, while EAM is a new component evaluated against a non-EAM baseline under the same encoder. The loss function follows prior work [23], but this is standard practice and does not make the result circular. The paper does report only single-run results and lacks a capacity-matched control for the added EAM parameters; however, these are experimental-validity concerns, not circularity. They concern whether the improvement is due to the echo-aware terms or to added capacity, not whether the claimed prediction reduces by construction to the paper's inputs. Therefore the correct circularity finding is none, score 0.

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

The central claim is an empirical architecture claim, not a derivation, so the ledger lists hand-chosen design and training quantities plus the domain assumptions needed for the reported gains to transfer. No new physical entities are introduced.

free parameters (4)

Learnable scaling factor alpha in Eq. (4) = Learned; final value not reported
Controls the gating strength of the EAM output and is trained end-to-end on the AEC loss.
Loss coefficients in Eq. (7) = 0.3, 0.45, 0.2, 0.05, 0.3
Hand-selected weights for STFT, PMSQE, magnitude, time, complex, and anti-wrapping phase losses; they shape which distortions the model optimizes.
Bark band count B = 100 for final config; 48 and 64 ablated
Feature compression dimension chosen by hand; central to the compact-latent claim and varied in Table II.
Bottleneck channels C and hidden size H = C=32, H=64 in final config
Selected based on the Pareto analysis in Fig. 2(a); the sensitivity of Bark-domain models to these choices is a core empirical finding.

assumptions (4)

domain assumption Bark-scale power spectra and the transposed Bark filterbank (Eqs. 5-6) are sufficient representations for estimating the speech extraction mask.
The compact-latent design assumes Bark-domain processing preserves the spectral structure needed for echo suppression; used throughout Section II-D and III-C.
domain assumption The PBFDAF output d_hat_L(n) provides a useful echo-related reference for the dual-branch encoder.
The hybrid AEC paradigm relies on the linear echo estimate aligning with the far-end reference; introduced in Sections II-A and II-B.
domain assumption Synthetic training mixtures from the AEC Challenge and DNS Challenge generalize to real AEC Challenge blind test sets.
Training data described in Section III-A are synthesized from public challenge datasets, but the claimed real-world performance assumes these mixtures are representative of the real blind test conditions.
standard math Data processing inequality (Section II-C) applies to the compressed representation and motivates why compression limits task information.
This is used as motivation for EAM, not as a derivation of the architecture; the inequality itself is an accepted background result.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Echo-Aware Modulation for Compact-Latent Frequency-Time Modeling in Lightweight Acoustic Echo Cancellation." pith.science (2026). https://pith.science/paper/WMEFNOOK

@misc{pith2026260803650,
  author       = {Pith},
  title        = {Pith review of: Echo-Aware Modulation for Compact-Latent Frequency-Time Modeling in Lightweight Acoustic Echo Cancellation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WMEFNOOK}},
  note         = {Machine review of arXiv:2608.03650}
}

read the original abstract

Existing lightweight acoustic echo cancellation (AEC) systems often combine linear AEC with Bark-domain DNN-based suppression to lower the computational footprint. In such systems, downsampling layers further compress the input features into a compact bottleneck representation, but this compression weakens frequency-time modeling capacity and degrades performance. To mitigate this limitation, we propose MSA-EchoLite, a lightweight Bark-domain AEC framework with an asymmetric dual-branch encoder and an echo-aware frequency-time modulation (EAM) module. The EAM module enriches the compressed bottleneck representation by modeling discrepancy and correlation cues between the dual-branch microphone and echo-related latent features. Experimental results show that the Bark-domain variant of MSA-EchoLite offers a better performance-complexity trade-off than its frequency-domain counterpart but is more sensitive to feature compression. With only 26.1% additional FLOPs over its non-EAM Bark-domain variant, its EAM-enhanced version achieves 99.1% of the PESQ of the frequency-domain counterpart, which requires nearly twice the FLOPs, and even surpasses it in SDR. Overall, MSA-EchoLite outperforms state-of-the-art lightweight AEC models while using only 0.2 M parameters and 100 M FLOPs/s.

Figures

Figures reproduced from arXiv: 2608.03650 by the authors.

**Figure 2.** Comparison of Bark- and frequency-domain modeling under com [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Spectrogram comparison on a double-talk test sample. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

[1]

Multi-task sub- band network for deep residual echo suppression,

J. Sun, D. Luo, Z. Li, J. Li, Y . Ju, and Y . Li, “Multi-task sub- band network for deep residual echo suppression,” in ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–2

work page 2023
[2]

DeepVQE: Real time deep voice quality enhancement for joint acoustic echo cancellation, noise suppression and dereverberation,

N. C. Ristea, E. Indenbom, A. Saabas, T. P ¨arnamaa, J. Guzhvin, and R. Cutler, “DeepVQE: Real time deep voice quality enhancement for joint acoustic echo cancellation, noise suppression and dereverberation,” in Proc. Interspeech 2023 , 2023, pp. 3819–3823

work page 2023
[3]

A progressive neural network for acoustic echo cancellation,

Z. Chen, X. Xia, S. Sun, Z. Wang, C. Chen, G. Xie, P. Zhang, and Y . Xiao, “A progressive neural network for acoustic echo cancellation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–2

work page 2023
[4]

Interfer- ence aware training target for dnn based joint acoustic echo cancellation and noise suppression

V . Khanagha, D. Koutsaidis, K. Kalgaonkar, and S. Srinivasan, “Interfer- ence aware training target for dnn based joint acoustic echo cancellation and noise suppression.” in Interspeech, 2024

work page 2024
[5]

MSA-DPCRN: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation,

Y . Ni, C. Pang, C. Huang, and C. Zou, “MSA-DPCRN: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation,” in Proc. Interspeech 2024, 2024, pp. 162–166

work page 2024
[6]

Two-stage language model framework for acoustic echo cancellation,

K. Xie, H. Li, N. Hou, H. Liu, and J. Chen, “Two-stage language model framework for acoustic echo cancellation,” in ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 14 562–14 566

work page 2026
[7]

Multi-task deep residual echo suppression with echo-aware loss,

S. Zhang, Z. Wang, J. Sun, Y . Fu, B. Tian, Q. Fu, and L. Xie, “Multi-task deep residual echo suppression with echo-aware loss,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9127–9131

work page 2022
[8]

Deep adaptive aec: Hybrid of deep learning and adaptive acoustic echo cancellation,

H. Zhang, S. Kandadai, H. Rao, M. Kim, T. Pruthi, and T. Kristjansson, “Deep adaptive aec: Hybrid of deep learning and adaptive acoustic echo cancellation,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 756–760

work page 2022

Show all 29 references

[9]

Low-complexity acoustic echo cancellation with neural kalman filtering,

D. Yang, F. Jiang, W. Wu, X. Fang, and M. Cao, “Low-complexity acoustic echo cancellation with neural kalman filtering,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

2023
[10]

A hybrid approach for low-complexity joint acoustic echo and noise reduction,

S. S. Shetu, N. K. Desiraju, J. M. M. Aponte, E. A. Habets, and E. Mabande, “A hybrid approach for low-complexity joint acoustic echo and noise reduction,” in 2024 18th International Workshop on Acoustic Signal Enhancement (IWAENC) . IEEE, 2024, pp. 349–353

2024
[11]

Attention-based dual stream interactive network for nonlinear residual echo suppression,

K. Xie, Z. Yang, J. Chen, and J. Li, “Attention-based dual stream interactive network for nonlinear residual echo suppression,” in 2024 32nd European Signal Processing Conference (EUSIPCO) , 2024, pp. 221–225

2024
[12]

EchoFree: Towards ultra lightweight and efficient neural acoustic echo cancellation,

X. Li, B. Kang, Z. Wang, Z. Zhang, M. Liu, Z. Fu, and L. Xie, “EchoFree: Towards ultra lightweight and efficient neural acoustic echo cancellation,” arXiv preprint arXiv:2508.06271 , 2025

2025 arXiv
[13]

Residual echo suppression using dual-stream interactive transformers with selective multi-scale en- coding,

K. Xie, Z. Yang, J. Chen, and M. Zhu, “Residual echo suppression using dual-stream interactive transformers with selective multi-scale en- coding,” IEEE Transactions on Audio, Speech and Language Processing , vol. 33, pp. 1103–1116, 2025

2025
[14]

Efficient high-performance bark-scale neural network for residual echo and noise suppression,

E. Seidel, P. Mowlaee, and T. Fingscheidt, “Efficient high-performance bark-scale neural network for residual echo and noise suppression,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 1386–1390

2024
[15]

CAGCRN: Real-time speech enhancement with a lightweight model for joint acoustic echo cancellation and noise suppression,

Y . Wang, Y . Liu, J. Liu, K. Niu, and Z. He, “CAGCRN: Real-time speech enhancement with a lightweight model for joint acoustic echo cancellation and noise suppression,” in Proc. Interspeech 2025 , 2025, pp. 768–772

2025
[16]

Conformer: Convolution-augmented transformer for speech recognition,

A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu et al. , “Conformer: Convolution-augmented transformer for speech recognition,” arXiv preprint arXiv:2005.08100 , 2020

2005 arXiv
[17]

Dual-path RNN: Efficient long sequence modeling for time-domain single-channel speech separation,

Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path RNN: Efficient long sequence modeling for time-domain single-channel speech separation,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2020, pp. 46–50

2020
[18]

DPCRN: Dual-path convolu- tion recurrent network for single channel speech enhancement,

X. Le, H. Chen, K. Chen, and J. Lu, “DPCRN: Dual-path convolu- tion recurrent network for single channel speech enhancement,” arXiv preprint arXiv:2107.05429, 2021

2021 arXiv
[19]

Frequency-domain and multirate adaptive filtering,

J. Shynk, “Frequency-domain and multirate adaptive filtering,” IEEE Signal Processing Magazine , vol. 9, no. 1, pp. 14–37, 1992

1992
[20]

T. M. Cover and J. A. Thomas, Elements of information theory (wiley se- ries in telecommunications and signal processing) . Wiley-interscience, 2006

2006
[21]

Real time speech enhancement in the waveform domain,

A. Defossez, G. Synnaeve, and Y . Adi, “Real time speech enhancement in the waveform domain,” arXiv preprint arXiv:2006.12847 , 2020

2006 arXiv
[22]

A deep learning loss function based on the perceptual evaluation of the speech quality,

J. M. Martin-Donas, A. M. Gomez, J. A. Gonzalez, and A. M. Peinado, “A deep learning loss function based on the perceptual evaluation of the speech quality,” IEEE Signal processing letters , vol. 25, no. 11, pp. 1680–1684, 2018

2018
[23]

MP-SENet: A speech enhancement model with parallel denoising of magnitude and phase spectra,

Y .-X. Lu, Y . Ai, and Z.-H. Ling, “MP-SENet: A speech enhancement model with parallel denoising of magnitude and phase spectra,” in Proc. Interspeech, 2023, pp. 3834–3838

2023
[24]

ICASSP 2022 acoustic echo cancellation challenge,

R. Cutler, A. Saabas, T. Parnamaa, M. Purin, H. Gamper, S. Braun, K. Sorensen, and R. Aichner, “ICASSP 2022 acoustic echo cancellation challenge,” in ICASSP 2022 , 2022

2022
[25]

The INTER- SPEECH 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

C. K. Reddy, V . Gopal, R. Cutler, E. Beyrami, R. Cheng, H. Dubey, S. Matusevych, R. Aichner, A. Aazami, S. Braun et al. , “The INTER- SPEECH 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” in INTERSPEECH, 2020

2020
[26]

Acoustic echo cancellation with the dual-signal transformation lstm network,

N. L. Westhausen and B. T. Meyer, “Acoustic echo cancellation with the dual-signal transformation lstm network,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 7138–7142

2021
[27]

AECMOS: A speech quality assessment metric for echo impairment,

M. Purin, S. Sootla, M. Sponza, A. Saabas, and R. Cutler, “AECMOS: A speech quality assessment metric for echo impairment,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 901–905

2022
[28]

Per- ceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,

A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Per- ceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,” in 2001 IEEE international conference on acoustics, speech, and signal processing. Proce...

2001
[29]

An algorithm for intelligibility prediction of time–frequency weighted noisy speech,

C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An algorithm for intelligibility prediction of time–frequency weighted noisy speech,” IEEE Transactions on audio, speech, and language processing , vol. 19, no. 7, pp. 2125–2136, 2011

2011

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Multi-task sub- band network for deep residual echo suppression,

J. Sun, D. Luo, Z. Li, J. Li, Y . Ju, and Y . Li, “Multi-task sub- band network for deep residual echo suppression,” in ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–2

work page 2023

[2] [2]

DeepVQE: Real time deep voice quality enhancement for joint acoustic echo cancellation, noise suppression and dereverberation,

N. C. Ristea, E. Indenbom, A. Saabas, T. P ¨arnamaa, J. Guzhvin, and R. Cutler, “DeepVQE: Real time deep voice quality enhancement for joint acoustic echo cancellation, noise suppression and dereverberation,” in Proc. Interspeech 2023 , 2023, pp. 3819–3823

work page 2023

[3] [3]

A progressive neural network for acoustic echo cancellation,

Z. Chen, X. Xia, S. Sun, Z. Wang, C. Chen, G. Xie, P. Zhang, and Y . Xiao, “A progressive neural network for acoustic echo cancellation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–2

work page 2023

[4] [4]

Interfer- ence aware training target for dnn based joint acoustic echo cancellation and noise suppression

V . Khanagha, D. Koutsaidis, K. Kalgaonkar, and S. Srinivasan, “Interfer- ence aware training target for dnn based joint acoustic echo cancellation and noise suppression.” in Interspeech, 2024

work page 2024

[5] [5]

MSA-DPCRN: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation,

Y . Ni, C. Pang, C. Huang, and C. Zou, “MSA-DPCRN: A multi-scale asymmetric dual-path convolution recurrent network with attentional feature fusion for acoustic echo cancellation,” in Proc. Interspeech 2024, 2024, pp. 162–166

work page 2024

[6] [6]

Two-stage language model framework for acoustic echo cancellation,

K. Xie, H. Li, N. Hou, H. Liu, and J. Chen, “Two-stage language model framework for acoustic echo cancellation,” in ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 14 562–14 566

work page 2026

[7] [7]

Multi-task deep residual echo suppression with echo-aware loss,

S. Zhang, Z. Wang, J. Sun, Y . Fu, B. Tian, Q. Fu, and L. Xie, “Multi-task deep residual echo suppression with echo-aware loss,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9127–9131

work page 2022

[8] [8]

Deep adaptive aec: Hybrid of deep learning and adaptive acoustic echo cancellation,

H. Zhang, S. Kandadai, H. Rao, M. Kim, T. Pruthi, and T. Kristjansson, “Deep adaptive aec: Hybrid of deep learning and adaptive acoustic echo cancellation,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 756–760

work page 2022