DualStream Contextual Fusion Network: Efficient Target Speaker Extraction by Leveraging Mixture and Enrollment Interactions

Chang Sun; Jianping An; Ke Xue; Rongfei Fan; Shanping Yu

REVIEW 4 major objections 6 minor 36 references

DualStream Contextual Fusion Network: Efficient Target Speaker Extraction by Leveraging Mixture and Enrollment Interactions

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read DCF-Net claims that explicitly modeling interactions between a contextualized enrollment and the mixture, via a DualStream Fusion Block, lifts target speaker extraction to 21.6 dB SI-SDRi and cuts target confusion to 0.4%.

desk verdict Plausible incremental TSE architecture, but the internal 21.6 vs 21.9 dB inconsistency undermines the SOTA claim until fixed. read the letter →

arxiv 2502.08191 v1 pith:KGAFNJNF submitted 2025-02-12 cs.SD eess.AS

Ke Xue , Rongfei Fan , Shanping Yu , Chang Sun , Jianping An This is my paper · ORCID

classification cs.SDeess.AS

keywords targetspeakerextractiontime-frequencydomainDualStreamFusionBlockMGImechanismsqueeze-and-excitationSI-SDRiconfusioncocktailparty

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that target speaker extraction should not rely solely on fixed speaker embeddings from an enrollment, because doing so discards contextual information and the interplay between the enrollment and the mixed signal. It introduces DCF-Net, a time-frequency domain network whose DualStream Fusion Block (DSFB) processes the mixture and a contextualized enrollment in two symmetric streams and lets them interact through the MGI mechanism and channel recalibration. On WSJ0-2Mix the model reports 21.6 dB SI-SDRi, above the previous best of 21.4 dB, with a target confusion rate of 0.4%. The paper takes this as evidence that explicitly modeling mixture-enrollment interactions improves extraction quality and robustness in noise and reverberation.

What carries the argument

The load-bearing component is the DualStream Fusion Block (DSFB). It takes the encoder's mixture representation $Y$ and contextualized enrollment $\bar{E}$, normalizes each with RMS-Norm, splits each into two halves via the MGI mechanism, and mixes them by element-wise multiplication ($\hat{Y} = Y_1 \circ \bar{E}_2$ and $\hat{E} = Y_2 \circ \bar{E}_1$), so each stream carries information from the other. Squeeze-and-Excitation blocks then recalibrate channel importance, and residual connections preserve the original features. The DSFB's job is to produce interaction-enriched features that guide the downstream dual-path improved transformer mask estimator.

What would settle it

Train DCF-Net and its DSFB-free baseline several times with different random seeds under identical data, optimizer, and epoch settings; if the 0.2 dB SI-SDRi gap on WSJ0-2Mix falls within the seed-to-seed variance, the claimed advantage of the fusion block is not established.

Watch

Extended reading notes

Core claim

The central claim is that jointly contextualizing the enrollment and the mixture, then fusing them across both spatial and channel dimensions, gives the extraction network a richer and more consistent representation than a plain speaker embedding. The authors claim this is what lets DCF-Net outperform prior target speaker extraction methods on WSJ0-2Mix, WHAM!, and WHAMR!, while cutting target confusion to 0.4%.

Load-bearing premise

The paper assumes the performance gain over CIENet-mDPTNet comes from the DSFB modules and not from unstated training differences or run-to-run variance, since no repeated runs or error bars are reported.

Editorial extensions

If this is right

On the WSJ0-2Mix benchmark, DCF-Net reports 21.6 dB SI-SDRi and 21.7 dB SDRi, beating the previous state of the art CIENet-mDPTNet (21.4 dB SI-SDRi) with 3.9M parameters.
On WHAM!, DCF-Net reaches 16.8 dB SI-SDRi, and on WHAMR! 15.8 dB, indicating the fusion design holds up under additive noise and reverberation.
The target confusion rate drops from 1% (CIENet) to 0.4% (DCF-Net) on WSJ0-2Mix, which the authors attribute to better use of enrollment information.
Ablations show that increasing the number of DSFB blocks from 0 to 8 monotonically improves SI-SDRi, with diminishing returns; the paper chooses 2 blocks as a parameter-efficiency trade-off.
Replacing the improved transformer with an RNN or base transformer in the extraction network lowers SI-SDRi to 20.8 dB or 21.3 dB, so the choice of extraction backbone interacts with the fusion design.

Reading between the lines

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

If the DSFB's interaction design is the true driver, a natural test is whether a simpler cross-attention between mixture and enrollment streams achieves the same gain at lower cost; the paper does not run this comparison.
Because the MGI mechanism is borrowed from single-image reflection separation, its success here suggests that 'component synergy' operations may transfer broadly across modalities whenever two related signals need mutual disambiguation.
The reported 0.4% confusion rate, if reproducible, would make the model substantially safer for hearing-assist and automatic meeting transcription, where grabbing the wrong speaker is costly.
A controlled study varying enrollment length and noise type would clarify whether the robustness claims on WHAM!/WHAMR! are due to the fusion block or to the interaction block inherited from CIENet.

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

Plainly: this is a solid incremental TSE paper with one serious internal inconsistency that needs fixing before I'd trust the headline. The DualStream Fusion Block is a genuine new combination: taking the MGI mechanism from image reflection separation, pairing it with SE channel attention, RMS-Norm, and a dual-path transformer, all in the T-F domain. That's a reasonable design move, and the ablation in Fig. 5 shows DSFB contributes meaningfully over the concatenation baseline. The target-confusion analysis is a nice practical addition, and reporting on WHAM!/WHAMR! is good practice.

The soft spots are real. The abstract and Table 1 say 21.6 dB SI-SDRi, while Section 5.1 says the model 'demonstrates 21.9dB and 22.1dB' on SI-SDRi and SDRi. No explanation is given, and Fig. 5's ablation at O=2 matches 21.61 dB, so the 21.9 figure looks like an error or a different run. That's a load-bearing problem because the claimed SOTA margin over CIENet-mDPTNet is only 0.2 dB; at 21.9 it's 0.5 dB. Without code, checkpoints, or error bars, a reader can't tell which number is real. Also, all comparisons to prior work use published numbers, and their own no-DSFB baseline (O=0) lands at 21.39 dB, essentially equal to CIENet-mDPTNet's 21.4 dB. That suggests the comparison is highly sensitive to implementation details. The ablation only varies the number of DSFB blocks; it doesn't isolate MGI or SE. Minor: Table 3 header says 'DFC-Net' (typo), and Fig. 5 labels O=0 and O=1 both as 'CC', which is confusing.

None of this says the architecture is conceptually wrong. The design is coherent and the improvements are plausible. The paper just hasn't nailed down its own central quantitative claim, and the reproducibility basis is too thin for a SOTA claim.

I'd want to see this after a major revision: code release, corrected and consistent numbers, error bars, and an ablation that turns off MGI and SE separately. As it stands, a serious referee would be justified in sending it back for those fixes. It's not a desk-reject: the idea is worth engaging with. I just wouldn't cite the 21.6 dB number until the inconsistency is resolved.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes DCF-Net, a time-frequency domain target speaker extraction model whose main contribution is the DualStream Fusion Block (DSFB), which processes the mixture and enrollment jointly through an interaction block, an MGI-style multiplicative interaction mechanism, and squeeze-and-excitation channel recalibration. The model is evaluated on WSJ0-2Mix, WHAM!, and WHAMR!, with a reported peak SI-SDRi of 21.6 dB on WSJ0-2Mix and a target confusion rate of 0.4%. The paper also includes ablations over the number of DSFB blocks and over the choice of extraction backbone.

Significance. If the reported numbers are reproducible, DCF-Net would represent a modest but genuine improvement over the strong CIENet-mDPTNet baseline on a standard benchmark, and the target-confusion analysis is a practically relevant addition. The paper's strengths are its use of standard corpora, comparison with a recent strong baseline, and ablations that vary the DSFB count and the extraction backbone. However, the central claim is an empirical SOTA claim resting on a 0.2 dB margin over one baseline, and the manuscript currently does not provide enough statistical or reproducibility evidence to establish that margin.

major comments (4)

[Abstract, Table 1, §5.1] The headline SI-SDRi is reported inconsistently: the abstract and Table 1 state 21.6 dB, while §5.1 states that the model demonstrates 21.9 dB SI-SDRi and 22.1 dB SDRi. Figure 5 at O=2 shows 21.61/21.71 dB, which matches Table 1 but not §5.1. Since the claimed gain over CIENet-mDPTNet is either 0.2 dB or 0.5 dB depending on which number is used, the manuscript must state which result is real and explain the discrepancy, ideally with training logs or released code.
[§5.1, Table 1, §5.4] The SOTA comparison is not supported by variance information. No error bars, repeated-seed results, significance tests, or same-configuration reruns of baselines are provided; the baselines are quoted from prior publications, and the in-house no-DSFB baseline labeled CC in Figure 5 gives 21.39 dB, essentially equal to the reported CIENet-mDPTNet value of 21.4 dB. Given typical run-to-run variance in this setting, a 0.2 dB margin cannot be distinguished from training noise. Please provide at least three runs with mean and standard deviation for DCF-Net, and ideally for the baseline under identical conditions.
[§3.2, Eqs. (2)-(4)] The central architectural innovation is the adaptation of the MGI mechanism from single-image reflection separation to T-F speech features, but the paper provides no evidence that this transfer is beneficial beyond the overall DSFB. The ablation in Figure 5 only varies the number of DSFB blocks; it does not isolate the MGI mechanism, the SE block, the residual connections, or the dual-stream design. A component ablation, for example replacing MGI with plain concatenation or attention and removing the SE block, is needed to support the claim that these specific modules rather than extra capacity or the interaction block are responsible for the reported gain.
[§5.3, Figure 4] The target confusion rate of 0.4% is presented without a precise definition of the rate's denominator, the number of test utterances, or confidence intervals. Since the rate is a small-count statistic, the manuscript should specify exactly how 'target confusion' is counted and how many test samples the percentage is based on, and ideally compare against more than one system.

minor comments (6)

[Abstract] The phrase 'calledtarget confusion problem' in the abstract appears to be a spacing error and should read 'called target confusion problem'.
[Table 1] The column header 'Domian' is a typo and should be 'Domain'.
[Table 3] The header 'diferent' is a typo, and the model name 'DFC-Net' in Table 3 is inconsistent with 'DCF-Net' used elsewhere.
[§3.2] The function X(·) that separates an input into two equal-channel outputs is never defined; please clarify whether it is a learned split operation or a specific layer.
[Eq. (5)] The SI-SDR definition has an apparent typo in the text below the equation ('˜s, ˜s are normalized'), and the relationship between the notation in Eq. (6) and the standard SI-SDR formulation should be made clearer.
[§5.2] The text contains minor wording errors such as 'inspite of this' and lacks details on the WHAM! and WHAMR! evaluation configurations; please specify the versions and SNR conditions used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DCF-Net's reported results are empirical training outcomes on held-out benchmarks, with no fitted quantity renamed as a prediction and no load-bearing self-citation.

full rationale

The paper's central claims are empirical, not derivational: DCF-Net is trained with SI-SDR loss and evaluated on WSJ0-2Mix, WHAM!, and WHAMR!. There is no chain in which an output is defined in terms of an input or in which a fitted parameter predicts a correlate of itself by construction. The building blocks (DRC, RMS-Norm, the MGI mechanism, the SE block, and the dual-path improved transformer) are all imported from external prior work, and the authors do not cite their own prior work as evidence for any premise. The choice O=2 for the number of DSFB blocks is an ablation choice made on the reported benchmark, which is a methodological concern about test-set selection rather than a circular reduction. The internal inconsistency between 21.6 dB in the abstract/Table 1 and 21.9 dB in Section 5.1 is a reporting or reproducibility problem, not a circularity problem. Because no load-bearing step reduces to its own input by definition or by self-citation, the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, so the ledger records the hyperparameters that were set by hand and the domain assumptions about transferring modules from other fields. No new physical entities are postulated.

free parameters (5)

Number of DSFB blocks O = 2
Chosen from the ablation in Sec. 5.4; directly affects final SI-SDRi and parameter count.
Transformer hidden dimension = 128
Set in Sec. 4.2 without ablation; influences extraction network capacity.
Channel dimension C = 256
Set in Sec. 4.2; affects all feature maps.
Learning rate = 0.0005 with 5-epoch warmup
Set in Sec. 4.2; no sensitivity analysis.
Number of dual-path transformer blocks N = 6
Set in Sec. 4.2; no sensitivity analysis.

assumptions (4)

standard math STFT and inverse STFT with dynamic range compression reconstruct the target speech from the estimated mask.
Assumed in Sec. 3.1 and Sec. 3.4; standard invertible transforms.
ad hoc to paper The MGI mechanism from single-image reflection separation (ref [29]) is beneficial when applied to time-frequency feature maps of speech.
Introduced in Sec. 3.2 with no adaptation argument; no ablation isolates its contribution.
domain assumption The dual-path improved transformer of ref [7] is a suitable extraction network for TSE.
Used in Sec. 3.3; borrowed from speech separation, not re-derived for extraction.
domain assumption SI-SDR loss is an appropriate training objective; minimizing it improves extraction quality.
Used as loss in Sec. 3.4; standard in the field.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DualStream Contextual Fusion Network: Efficient Target Speaker Extraction by Leveraging Mixture and Enrollment Interactions." pith.science (2026). https://pith.science/paper/KGAFNJNF

@misc{pith2026250208191,
  author       = {Pith},
  title        = {Pith review of: DualStream Contextual Fusion Network: Efficient Target Speaker Extraction by Leveraging Mixture and Enrollment Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGAFNJNF}},
  note         = {Machine review of arXiv:2502.08191}
}

read the original abstract

Target speaker extraction focuses on extracting a target speech signal from an environment with multiple speakers by leveraging an enrollment. Existing methods predominantly rely on speaker embeddings obtained from the enrollment, potentially disregarding the contextual information and the internal interactions between the mixture and enrollment. In this paper, we propose a novel DualStream Contextual Fusion Network (DCF-Net) in the time-frequency (T-F) domain. Specifically, DualStream Fusion Block (DSFB) is introduced to obtain contextual information and capture the interactions between contextualized enrollment and mixture representation across both spatial and channel dimensions, and then rich and consistent representations are utilized to guide the extraction network for better extraction. Experimental results demonstrate that DCF-Net outperforms state-of-the-art (SOTA) methods, achieving a scale-invariant signal-to-distortion ratio improvement (SI-SDRi) of 21.6 dB on the benchmark dataset, and exhibits its robustness and effectiveness in both noise and reverberation scenarios. In addition, the wrong extraction results of our model, called target confusion problem, reduce to 0.4%, which highlights the potential of DCF-Net for practical applications.

Figures

Figures reproduced from arXiv: 2502.08191 by the authors.

**Figure 2.** The overall architecture of our proposed network. There are four main components, different capital letters [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Multi-range 2D convolution. 3.2 DualStream Fusion Block (DSFB) DSFB consists of two fully symmetric channel flows, receiving the outputs Y and E¯ of the encoder respectively, followed by RMS layer normalization (RMS-Norm) [28]. Compared to LayerNorm, RMS-Norm effectively stabilizes the magnitude of layer activations, which ensures invariance to the rescaling of weights and datasets, thereby promoting more robust and… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The target confusion problem (TCP) in CIENet [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Performance comparison of varying numbers of DSFB (O) on the WSJ0-2Mix dataset, including parameter(M) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

[1]

Some experiments on the recognition of speech, with one and with two ears,

E. C. Cherry, “Some experiments on the recognition of speech, with one and with two ears,” J. Acoust. Soc. Am. , vol. 25, no. 5, pp. 975–979, 1953

1953
[2]

The cocktail party phenomenon revisited: The importance of working memory capacity,

A. R. Conway, N. Cowan, and M. F. Bunting, “The cocktail party phenomenon revisited: The importance of working memory capacity,” Psychon. Bull. Rev., vol. 8, pp. 331–335, 2001

work page 2001
[3]

An event-related potential study of selective auditory attention in children and adults,

D. Coch, L. D. Sanders, and H. J. Neville, “An event-related potential study of selective auditory attention in children and adults,” J. Cogn. Neurosci., vol. 17, no. 4, pp. 605–622, 2005

work page 2005
[4]

Selective cortical representation of attended speaker in multi-talker speech perception,

N. Mesgarani and E. F. Chang, “Selective cortical representation of attended speaker in multi-talker speech perception,” Nature, vol. 485, no. 7397, pp. 233–236, 2012

work page 2012
[5]

Supervised speech separation based on deep learning: An overview,

D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,”IEEE/ACM Trans. Audio Speech Lang. Process., vol. 26, no. 10, pp. 1702–1726, 2018

2018
[6]

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 in Proc. IEEE ICASSP, 2020, pp. 46–50

work page 2020
[7]

Dual-path transformer network: Direct context-aware modeling for end-to-end monaural speech separation,

J. Chen, Q. Mao, and D. Liu, “Dual-path transformer network: Direct context-aware modeling for end-to-end monaural speech separation,” arXiv preprint arXiv:2007.13975, 2020

arXiv 2007
[8]

Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation,

Z.-Q. Wang, C. Samuele, and C. Shukjae, “Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation,” in in Proc. IEEE ICASSP, 2023, pp. 1–5

work page 2023

Show all 36 references

[9]

Crossnet: Leveraging global, cross-band, narrow-band, and positional encoding for single-and multi-channel speaker separation,

K. V . Ahmadi and D. Wang, “Crossnet: Leveraging global, cross-band, narrow-band, and positional encoding for single-and multi-channel speaker separation,” arXiv preprint arXiv:2403.03411, 2024

2024 arXiv
[10]

Neural target speech extraction: An overview,

Z. Katerina, D. Marc, and O. Tsubasa, “Neural target speech extraction: An overview,” IEEE Signal Process. Mag., vol. 40, no. 3, pp. 8–29, 2023

2023
[11]

V oicefilter: Targeted voice separation by speaker-conditioned spectro- gram masking,

Q. Wang, H. Muckenhirn, and K. Wilson, “V oicefilter: Targeted voice separation by speaker-conditioned spectro- gram masking,” arXiv preprint arXiv:1810.04826, 2018

2018 arXiv
[12]

Time-domain speaker extraction network,

C. Xu, W. Rao, E. Chng, and H. Li, “Time-domain speaker extraction network,” in2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2019, pp. 327–334

2019
[13]

Spex: Multi-scale time domain speaker extraction network,

——, “Spex: Multi-scale time domain speaker extraction network,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 28, pp. 1370–1384, 2020

2020
[14]

Spex+: A complete time domain speaker extraction network,

M. Ge, C. Xu, L. Wang, E. Chng, J. Dang, and H. Li, “Spex+: A complete time domain speaker extraction network,” arXiv preprint arXiv:2005.04686, 2020

2005 arXiv
[15]

Adaptive-spex: Local and global perceptual modeling with speaker adaptation for target speaker extraction,

X. Xu, D. Yan, and D. Li, “Adaptive-spex: Local and global perceptual modeling with speaker adaptation for target speaker extraction,” in 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC) . IEEE, 2023, pp. 342–347

2023
[16]

Multi-stage speaker extraction with utterance and frame-level reference signals,

M. Ge, C. Xu, L. Wang, E. S. Chng, J. Dang, and H. Li, “Multi-stage speaker extraction with utterance and frame-level reference signals,” in in Proc. IEEE ICASSP, 2021, pp. 6109–6113

2021
[17]

Neural speaker extraction with speaker-speech cross-attention network

W. Wang, C. Xu, M. Ge, and H. Li, “Neural speaker extraction with speaker-speech cross-attention network.” inin Proc. Interspeech, 2021, pp. 3535–3539

2021
[18]

Robust speaker extraction network based on iterative refined adaptation,

C. Deng, S. Ma, Y . Zhang, Y . Sha, H. Zhang, H. Song, and X. Li, “Robust speaker extraction network based on iterative refined adaptation,” arXiv preprint arXiv:2011.02102, 2020

2011 arXiv
[19]

Target speaker extraction with ultra-short reference speech by ve-ve framework,

L. Yang, W. Liu, L. Tan, Y . Jaemo, and M. Han-Gil, “Target speaker extraction with ultra-short reference speech by ve-ve framework,” in in Proc. IEEE ICASSP, 2023, pp. 1–5

2023
[20]

X-sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,

K. Liu, Z. Du, X. Wan, and H. Zhou, “X-sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,” in in Proc. IEEE ICASSP, 2023, pp. 1–5

2023
[21]

Target speech extraction with pre-trained self-supervised learning models,

J. Peng, D. Marc, and O. Tsubasa, “Target speech extraction with pre-trained self-supervised learning models,” in in Proc. IEEE ICASSP, 2024, pp. 10 421–10 425

2024
[22]

Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,

K. Žmol, M. Delcroix, K. Kinoshita, T. Ochiai, T. Nakatani, and L. J. Burget, “Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,” IEEE J. Sel. Top. Signal Process., vol. 13, no. 4, pp. 800–814, 2019

2019
[23]

X-tf-gridnet: A time–frequency domain target speaker extraction network with adaptive speaker embedding fusion,

F. Hao, X. Li, and C. Zheng, “X-tf-gridnet: A time–frequency domain target speaker extraction network with adaptive speaker embedding fusion,” Information Fusion, vol. 112, 2024. 9 A PREPRINT - F EBRUARY 13, 2025

2024
[24]

Single-channel speech extraction using speaker inventory and attention network,

X. e. a. Xiao, “Single-channel speech extraction using speaker inventory and attention network,” in in Proc. IEEE ICASSP, 2019, pp. 86–90

2019
[25]

Sef-net: Speaker embedding free target speaker extraction network,

B. Zeng, H. Suo, Y . Wan, and M. Li, “Sef-net: Speaker embedding free target speaker extraction network,”Proc. Interspeech, pp. 3452–3456, 2023

2023
[26]

Target speaker extraction by directly exploiting contextual information in the time-frequency domain,

X. Yang, C. Bao, J. Zhou, and X. Chen, “Target speaker extraction by directly exploiting contextual information in the time-frequency domain,” in in Proc. IEEE ICASSP, 2024, pp. 10 476–10 480

2024
[27]

On the importance of power compression and phase estimation in monaural speech dereverberation,

A. Li, C. Zheng, R. Peng, and X. Li, “On the importance of power compression and phase estimation in monaural speech dereverberation,” J. Acoust. Soc. Am. Express Lett. , vol. 1, no. 1, 2021

2021
[28]

Root mean square layer normalization,

B. Zhang and R. Sennrich, “Root mean square layer normalization,” Adv. Neural Inf. Process. Syst., vol. 32, 2019

2019
[29]

Single image reflection separation via component synergy,

Q. Hu and X. Guo, “Single image reflection separation via component synergy,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 13 138–13 147

2023
[30]

Squeeze-and-excitation networks,

J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 7132–7141

2018
[31]

Sdr–half-baked or well done?

L. Jonathan, W. Scott, E. Hakan, and H. John, “Sdr–half-baked or well done?” in in Proc. IEEE ICASSP, 2019, pp. 626–630

2019
[32]

Csr-i (wsj0) complete ldc93s6a,

G. John, P. David, G. Doug, and P. David, “Csr-i (wsj0) complete ldc93s6a,” Web Download. Philadelphia: Linguistic Data Consortium, vol. 83, 1993

1993
[33]

Wham!: Extending speech separation to noisy environments,

G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, and J. L. Roux, “Wham!: Extending speech separation to noisy environments,” arXiv preprint arXiv:1907.01160, 2019

1907 arXiv
[34]

Whamr!: Noisy and reverberant single-channel speech separation,

M. Maciejewski, G. Wichern, E. McQuinn, and J. Le Roux, “Whamr!: Noisy and reverberant single-channel speech separation,” in in Proc. IEEE ICASSP, 2020, pp. 696–700

2020
[35]

Target confusion in end-to-end speaker extraction: Analysis and approaches,

Z. Zhao, D. Yang, R. Gu, H. Zhang, and Y . Zou, “Target confusion in end-to-end speaker extraction: Analysis and approaches,” arXiv preprint arXiv:2204.01355, 2022

2022 arXiv
[36]

Attention is all you need,

A. Vaswani, “Attention is all you need,” Adv. Neural Inf. Process. Syst., 2017. 10

2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Some experiments on the recognition of speech, with one and with two ears,

E. C. Cherry, “Some experiments on the recognition of speech, with one and with two ears,” J. Acoust. Soc. Am. , vol. 25, no. 5, pp. 975–979, 1953

1953

[2] [2]

The cocktail party phenomenon revisited: The importance of working memory capacity,

A. R. Conway, N. Cowan, and M. F. Bunting, “The cocktail party phenomenon revisited: The importance of working memory capacity,” Psychon. Bull. Rev., vol. 8, pp. 331–335, 2001

work page 2001

[3] [3]

An event-related potential study of selective auditory attention in children and adults,

D. Coch, L. D. Sanders, and H. J. Neville, “An event-related potential study of selective auditory attention in children and adults,” J. Cogn. Neurosci., vol. 17, no. 4, pp. 605–622, 2005

work page 2005

[4] [4]

Selective cortical representation of attended speaker in multi-talker speech perception,

N. Mesgarani and E. F. Chang, “Selective cortical representation of attended speaker in multi-talker speech perception,” Nature, vol. 485, no. 7397, pp. 233–236, 2012

work page 2012

[5] [5]

Supervised speech separation based on deep learning: An overview,

D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,”IEEE/ACM Trans. Audio Speech Lang. Process., vol. 26, no. 10, pp. 1702–1726, 2018

2018

[6] [6]

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 in Proc. IEEE ICASSP, 2020, pp. 46–50

work page 2020

[7] [7]

Dual-path transformer network: Direct context-aware modeling for end-to-end monaural speech separation,

J. Chen, Q. Mao, and D. Liu, “Dual-path transformer network: Direct context-aware modeling for end-to-end monaural speech separation,” arXiv preprint arXiv:2007.13975, 2020

arXiv 2007

[8] [8]

Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation,

Z.-Q. Wang, C. Samuele, and C. Shukjae, “Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation,” in in Proc. IEEE ICASSP, 2023, pp. 1–5

work page 2023