A Comparative Study on Positional Encoding for Time-frequency Domain Dual-path Transformer-based Source Separation Models

Kohei Saijo; Tetsuji Ogawa

REVIEW 3 major objections 4 minor 2 cited by

A Comparative Study on Positional Encoding for Time-frequency Domain Dual-path Transformer-based Source Separation Models

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read In dual-path Transformer separation models, removing positional encoding gives better length extrapolation than explicit PEs, especially with convolutional layers, while explicit PEs win on matched or shorter inputs.

desk verdict NoPE length-extrapolation advantage is real, but the conv-specific claim is overreached and variance is missing. read the letter →

arxiv 2504.19605 v2 pith:DVKVMKS5 submitted 2025-04-28 eess.AS cs.SD

Kohei Saijo , Tetsuji Ogawa This is my paper · ORCID

classification eess.AScs.SD

keywords positionalencodinglengthextrapolationdual-pathtransformertime-frequencydomainsourceseparationconvolutionalimplicitpositionsampling-rategeneralizationTF-Locoformer

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 asks how positional encoding (PE) should be chosen in Transformer-based time-frequency dual-path source separation models, using TF-Locoformer as the test bed. It compares absolute encoding (APE), a relative encoding (KERPLE), rotary encoding (RoPE), and no encoding (NoPE) on speech separation and music source separation. The central claim is a trade-off: explicit PEs improve performance when inputs are no longer than the training segments, but NoPE generalizes better to longer inputs and to unseen sampling rates, and this extrapolation advantage is strongest when the model contains convolutional layers. The authors argue that convolutional layers already carry implicit position information, so explicit PE can be omitted without a matched-length penalty. This matters because length extrapolation is what lets a model trained at one sampling rate handle signals at another, a common practical need.

What carries the argument

The carrying mechanism is the TF-Locoformer block, which alternates frequency-axis and time-axis modeling, each consisting of multi-head self-attention surrounded by ConvSwiGLU feed-forward layers. The treatment variable is how position enters those attention layers: additive sinusoidal embeddings (APE), KERPLE's learnable kernelized relative distances added to attention logits, rotary rotations of queries and keys (RoPE), or no explicit position signal (NoPE). The explanatory load is carried by the convolutional layers: because zero-padding lets convolutions encode position locally, the model with NoPE retains positional information when the kernel size K is greater than one, and this implicit encoding generalizes to unseen lengths better than any explicit scheme tested.

What would settle it

Retrain the key comparisons (e.g., RoPE vs NoPE on WHAMR! at 4-second training, and the MUSDB 3-second models) several times with different random seeds and report the distribution of separation metrics; if the NoPE extrapolation advantage over RoPE falls within the seed-to-seed spread, or if NoPE's matched-length parity disappears with repeated runs, the claimed PE trade-off is not established.

Watch

Extended reading notes

Core claim

The paper's central discovery is that in TF-domain dual-path Transformers, the best positional encoding depends on whether the input length matches training. With segments at or below training length, explicit PEs—particularly RoPE on matched data—deliver better separation scores; beyond training length, NoPE outperforms explicit PEs, and the gap widens as inference segments grow, for example in music source separation where NoPE keeps improving past the 8-second point at which RoPE saturates. The authors attribute this to position information encoded implicitly by zero-padded convolutions in the model: with kernel size greater than one, NoPE matches explicit PEs at the training length, while with kernel size one (K=1) the explicit PEs are necessary. The same trend transfers from the 8 kHz to the 16 kHz test condition, indicating that NoPE's extrapolation advantage extends to the frequency dimension and thus to sampling-rate generalization.

Load-bearing premise

The conclusions rest on single training runs per configuration, so small performance differences—some around 0.1 dB—are treated as real signals rather than run-to-run noise.

Editorial extensions

If this is right

For deployment where inputs may exceed training length or sampling rate may change, NoPE is the safer choice in convolution-containing dual-path Transformers.
For matched-length or shorter inputs, explicit PEs such as RoPE remain preferable, so the choice is a length-dependent design decision.
The implicit positional encoding from convolutions is sufficient to replace explicit PEs in this architecture class, allowing the 16% training-time saving NoPE showed in the WHAMR! setup.
NoPE's length extrapolation transfers from speech separation to music source separation, suggesting the result is not dataset-specific.
Sampling-rate generalization (e.g., 8 kHz training to 16 kHz test) is improved by NoPE relative to RoPE, supporting low-rate training for high-rate deployment.

Reading between the lines

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

A direct extension is to test NoPE at sampling rates far beyond training (e.g., 8 kHz to 32 or 44.1 kHz) to see whether the frequency-dimension extrapolation holds beyond the 16 kHz case reported.
The K=1 experiments imply an architecture-level rule: the more convolutional the Transformer, the less need for explicit PE; a systematic sweep of kernel size versus PE type could quantify when each becomes redundant.
The length-dependent trade-off suggests a hybrid design—explicit PE for short-range in-distribution accuracy plus convolution-provided position for long-range extrapolation—could beat either extreme.
Because only single-seed results are reported, the precise ordering among PE methods on matched-length inputs should be treated as provisional until replicated.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: this is a competent, genuinely useful empirical comparison of positional encodings in TF-domain dual-path separation. The headline result — NoPE matches or beats explicit PEs on matched lengths and generalizes better to longer inputs — is real and consistently observed across WHAMR! and MUSDB. The paper is worth reading and deserving of a serious referee, but the abstract slightly overstates the convolution-specific mechanism.

The new piece is the systematic comparison: previous work only looked at APE vs RPE in enhancement, and NoPE was mostly a side observation. Here they control the architecture and vary PE, including KERPLE and RoPE, and show the extrapolation benefit holds across two datasets and several training lengths. The 16% training speedup for NoPE over RoPE is a useful practical detail. The MUSDB curves make the saturation-contrast point well: RoPE saturates around 8s while NoPE keeps improving.

Where it gets soft: no variance or significance testing anywhere. Differences of 0.1–0.2 dB are treated as meaningful, and with a single run per configuration that's not defensible. Also, their short-sequence claim (explicit PEs better when evaluating shorter than training) is contradicted by their own matched-length results: on WHAMR!, NoPE is comparable or better on 8k min for all matched sets (A4 vs A1-A3, B4 vs B1-B3). The C1 vs C2 comparison is a single difference of 0.1–0.3 dB. They push this conclusion harder than the data support.

The bigger issue: the (ii) claim — 'particularly pronounced when the model contains convolutional layers' — rests on the K=1 ablation (Table II). That ablation removes local receptive field and capacity, not just implicit position encoding. The uniform ~4 dB drop confirms capacity loss. And Table II actually shows NoPE's extrapolation advantage over RoPE on 16k is bigger without conv (B4 vs B3 in Table I: ~0.4 dB for min, 0.5 for max; D4 vs D3 in Table II: 4.6 and 5.3 dB), which is the opposite of 'particularly pronounced.' So the conv-specific mechanism is unverified; what is verified is that NoPE generalizes well in this architecture.

All that said, the main practical takeaway — use NoPE in TF-domain dual-path with convolution when you expect longer inputs or unseen sampling rates — is reasonably supported. The paper would benefit from repeated runs with error bars, toning down the short-sequence and conv-specific language, and a fairer ablation (e.g., keep K=1 but increase capacity, or compare with a model that has conv but no padding). No code is provided, which limits reproducibility but does not falsify the result.

Recommendation: send to peer review. It's a useful empirical study with a clear question, and the flaws are fixable in revision. I would bring it to reading group to discuss extrapolation methodology.

Referee Report

3 major / 4 minor

Summary. The paper compares four positional encoding strategies (APE, KERPLE, RoPE, and NoPE) in the TF-Locoformer TF-domain dual-path source separation architecture. Experiments on WHAMR! and MUSDB18-HQ evaluate both matched-length performance and extrapolation to longer input segments and to higher sampling rates. The authors report that explicit PEs perform better on inputs at or shorter than the training length, whereas NoPE yields superior length extrapolation, and they attribute this to convolutional layers implicitly encoding position. A K=1 ablation is presented to support the convolutional mechanism, and additional results show NoPE scaling to a medium model and faster training.

Significance. If the findings are robust, the paper provides practical design guidance for TF-domain dual-path separators: use explicit PEs for in-distribution lengths and NoPE when inputs may exceed training length or when sampling-rate generalization is needed. The study is the first systematic PE comparison for this architecture class, and the consistent NoPE advantage on MUSDB extrapolation (Figures 1–2) and on WHAMR! 16 kHz (Table I, B4 vs B3) is a useful empirical result. However, the evidence is weakened by single training runs without variance reporting and by a confounded K=1 ablation, so the central mechanism claim is not yet established.

major comments (3)

[Section IV-D, Table II] The K=1 ablation does not isolate implicit positional encoding from model capacity: setting the ConvSwiGLU kernel size to 1 removes the local receptive field of every block, and the uniform ~4 dB drop on 8k min relative to Table I is consistent with underfitting rather than loss of position information. Therefore the claim that the NoPE extrapolation advantage is 'particularly pronounced when the model contains convolutional layers' (abstract and conclusion) is not supported by this experiment. In fact, Table II shows the opposite trend: without convolution, NoPE beats RoPE on 16k min/max by 4.6/5.3 dB (11.2/12.0 vs 6.6/6.7), a much larger extrapolation advantage than in the conv case (B4 vs B3: 17.3/18.1 vs 16.9/17.6). A clean test would require an architecture that removes or randomizes positional cues without reducing receptive-field capacity, or a matched-capacity control.
[Section IV-D, Table I and Section IV-E, Figures 1–2] All conclusions are based on single training runs, and differences as small as 0.1 dB are treated as decisive. For instance, the claim that explicit PEs are better on shorter inputs rests on C1 vs C2 at 8k min (17.4 vs 17.3 dB), which is within typical run-to-run variability, and the matched-length advantage of RoPE over NoPE is similarly small (B3 vs B4: 17.8 vs 17.8 on 8k min). Without multiple seeds, confidence intervals, or significance tests, findings (i) and parts of finding (ii) are not statistically grounded. I request at least three seeds for the key comparisons (A3/A4, B3/B4, C1/C2, and the MUSDB RoPE/NoPE pairs) or a stated variance across runs.
[Section III-C and Section IV-C] The RoPE base frequency (theta) is not reported, although this hyperparameter is known to strongly affect length extrapolation. Since TF-Locoformer's original RoPE configuration may not be optimized for the longer-sequence and higher-sampling-rate conditions tested here, the comparison could understate RoPE's extrapolation capability. Please report the theta value used, and ideally include a sensitivity check (e.g., the default and one larger or smaller theta).

minor comments (4)

[Abstract and Conclusion] The phrase 'particularly pronounced when the model contains convolutional layers' is repeatedly stated, but, as noted in Major Comment 1, the reported evidence in Table II points the other way; please soften or revise this claim.
[Section IV-D, Table I caption] The caption contains a typo: '‘ MIN ‘ VERSION IS USED FOR TRAINING' should read 'The min version is used for training'; similar spacing issues appear in other table captions.
[Section IV-D, Table III] It is unclear whether TF-Locoformer-NoPE (M) was trained with the same budget and early-stopping criterion as the reproduced TF-Locoformer (S), and whether the medium-size NoPE result is from a single run; a brief note would improve comparability.
[Section IV-E, Figures 1–2] The figures would benefit from error bars or at least point markers for each condition; currently the reader cannot assess whether the differences between curves (e.g., RoPE vs NoPE at T'=6 in Figure 2) are meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the PE comparison is empirical and no conclusion reduces to its inputs.

full rationale

This paper is an empirical comparison study, not a derivation. The central results (APE vs. KERPLE vs. RoPE vs. NoPE on WHAMR! and MUSDB) come from training and evaluating model variants, and no performance number is fitted and then renamed as a prediction. The only self-citations are to TF-Locoformer as the base architecture and to the authors' prior sampling-frequency-invariance work; these are used as background and architecture choice, not as the evidence for the PE ranking. The K=1 ablation in Table II could confound removal of convolution with reduction of model capacity, but that is a validity or experimental-design concern rather than circularity: NoPE's behavior is still measured, not constructed. No equation or claim in the paper is equivalent to its own input by definition, and no load-bearing step reduces to a self-citation chain. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical comparison and introduces no new theoretical entities. Its conclusions rest on a set of domain assumptions: STFT bin-resolution invariance across sampling rates, the transfer of convolution position-encoding evidence from vision to audio, the K=1 ablation as a proxy for removing convolution, and the representativeness of single runs. Learnable KERPLE parameters r1 and r2, and the unstated RoPE base, are the PE-specific values that could affect the results.

free parameters (3)

KERPLE r1 = learned during training, not reported
Equation (9) defines KERPLE bias as -r1 log(1 + r2 |i-j|); r1 and r2 are learnable and influence the relative-position prior.
KERPLE r2 = learned during training, not reported
Same as r1; controls the nonlinear distance scaling.
RoPE base frequency (theta) = not reported (typically 10000)
Section III-C describes RoPE without specifying theta; this hyperparameter controls the rotation frequency schedule and is known to affect length extrapolation.

assumptions (5)

domain assumption STFT with fixed window and hop sizes yields constant TF-bin resolution regardless of sampling rate, so extrapolation in frequency bins corresponds to generalization to unseen sampling rates.
Section II-B; this equivalence motivates the 8kHz vs 16kHz frequency-extrapolation evaluation. If this assumption fails, the interpretation of frequency-axis length extrapolation as sampling-rate invariance breaks down.
domain assumption Convolutional layers in TF-Locoformer encode positional information implicitly through zero-padding, as demonstrated for vision CNNs.
Section III-D cites [16]; this is the premise for expecting NoPE to work in a model with convolutions. If ConvSwiGLU does not provide useful position cues, NoPE's matched-length performance would be surprising.
ad hoc to paper Setting the convolution kernel size to K=1 removes convolutional position encoding while preserving the rest of the architecture.
Section IV-D, Table II; the paper treats K=1 as the 'no convolution' condition, but this also reduces model capacity and changes the receptive field, a confound that is not flagged.
ad hoc to paper RoPE is used with a default base frequency (theta) not stated in the paper; this hyperparameter can influence length extrapolation.
Section III-C describes RoPE without specifying theta; the paper's RoPE results may be sensitive to this choice.
domain assumption Observed performance differences from single runs are representative and not dominated by training variance.
No repeated runs or significance tests are reported in Section IV-D or IV-E; the conclusions treat 0.1-0.4 dB gaps as meaningful.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Comparative Study on Positional Encoding for Time-frequency Domain Dual-path Transformer-based Source Separation Models." pith.science (2026). https://pith.science/paper/DVKVMKS5

@misc{pith2026250419605,
  author       = {Pith},
  title        = {Pith review of: A Comparative Study on Positional Encoding for Time-frequency Domain Dual-path Transformer-based Source Separation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVKVMKS5}},
  note         = {Machine review of arXiv:2504.19605}
}

read the original abstract

In this study, we investigate the impact of positional encoding (PE) on source separation performance and the generalization ability to long sequences (length extrapolation) in Transformer-based time-frequency (TF) domain dual-path models. The length extrapolation capability in TF-domain dual-path models is a crucial factor, as it affects not only their performance on long-duration inputs but also their generalizability to signals with unseen sampling rates. While PE is known to significantly impact length extrapolation, there has been limited research that explores the choice of PEs for TF-domain dual-path models from this perspective. To address this gap, we compare various PE methods using a recent state-of-the-art model, TF-Locoformer, as the base architecture. Our analysis yields the following key findings: (i) When handling sequences that are the same length as or shorter than those seen during training, models with PEs achieve better performance. (ii) However, models without PE exhibit superior length extrapolation. This trend is particularly pronounced when the model contains convolutional layers.

Figures

Figures reproduced from arXiv: 2504.19605 by the authors.

**Figure 2.** uSDR scores of models trained on 3-, 6-, or 8-second chunks. uSDRs [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Is MixIT Really Unsuitable for Correlated Sources? Exploring MixIT for Unsupervised Pre-training in Music Source Separation
eess.AS 2025-05 conditional novelty 6.0 of 10

MixIT pre-training on unlabeled Free Music Archive audio improves MUSDB18 separation after fine-tuning, with uSDR gains of about 0.3 to 0.5 dB.
FasTUSS: Faster Task-Aware Unified Source Separation
cs.SD 2025-07 conditional novelty 5.0 of 10

FasTUSS cuts TUSS's computational cost by up to 81 percent with minor SNR drops, and introduces a causal variant compatible with KVCache.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages · cited by 2 Pith papers

[1]

Deep clustering: Discrimi- native embeddings for segmentation and separation,

J. R. Hershey, Z. Chen, J. Le Roux et al., “Deep clustering: Discrimi- native embeddings for segmentation and separation,” in Proc. ICASSP, 2016

work page 2016
[2]

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

D. Yu, M. Kolbæk, Z. H. Tan et al., “Permutation invariant training of deep models for speaker-independent multi-talker speech separation,” in Proc. ICASSP, 2017

work page 2017
[3]

TasNet: Time-domain audio separation net- work for real-time, single-channel speech separation,

Y . Luo and N. Mesgarani, “TasNet: Time-domain audio separation net- work for real-time, single-channel speech separation,” in Proc. ICASSP, 2018

work page 2018
[4]

Conv-TasNet: Surpassing ideal time-frequency magnitude mask- ing for speech separation,

——, “Conv-TasNet: Surpassing ideal time-frequency magnitude mask- ing for speech separation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 27, no. 8, pp. 1256–1266, 2019

work page 2019
[5]

TFPSNet: Time-frequency domain path scanning network for speech separation,

L. Yang, W. Liu, and W. Wang, “TFPSNet: Time-frequency domain path scanning network for speech separation,” in Proc. ICASSP, 2022

work page 2022
[6]

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

Z.-Q. Wang, S. Cornell, S. Choi et al. , “TF-GridNet: Integrating full- and sub-band modeling for speech separation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 3221–3236, 2023

work page 2023
[7]

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 Proc. ICASSP, 2020

2020
[8]

Attention is all you need in speech separation,

C. Subakan, M. Ravanelli, S. Cornell et al., “Attention is all you need in speech separation,” in Proc. ICASSP, 2021

work page 2021

Show all 36 references

[9]

Dual-path mamba: Short and long-term bidirectional selective structured state space models for speech separation,

X. Jiang, C. Han, and N. Mesgarani, “Dual-path mamba: Short and long-term bidirectional selective structured state space models for speech separation,” in Proc. ICASSP, 2025, pp. 1–5

2025
[10]

Toward universal speech enhancement for diverse input conditions,

W. Zhang, K. Saijo, Z.-Q. Wang et al. , “Toward universal speech enhancement for diverse input conditions,” in Proc. ASRU, 2023

2023
[11]

Task-aware unified source separation,

K. Saijo, J. Ebbers, F. G. Germain et al. , “Task-aware unified source separation,” in Proc. ICASSP, 2025, pp. 1–5

2025
[12]

Tf-locoformer: Transformer with local modeling by convolution for speech separation and enhance- ment,

K. Saijo, G. Wichern, F. G. Germain et al., “Tf-locoformer: Transformer with local modeling by convolution for speech separation and enhance- ment,” in Proc. IWAENC, 2024, pp. 205–209

2024
[13]

RoFormer: Enhanced transformer with rotary position embedding,

J. Su, Y . Lu, S. Pan et al., “RoFormer: Enhanced transformer with rotary position embedding,” arXiv preprint arXiv:2104.09864 , 2021

2021 arXiv
[14]

Train short, test long: Attention with linear biases enables input length extrapolation,

O. Press, N. A. Smith, and M. Lewis, “Train short, test long: Attention with linear biases enables input length extrapolation,” arXiv preprint arXiv:2108.12409, 2021

2021 arXiv
[15]

Kerple: Kernelized relative positional embedding for length extrapolation,

T.-C. Chi, T.-H. Fan, P. J. Ramadge et al., “Kerple: Kernelized relative positional embedding for length extrapolation,” Proc. NeurIPS, vol. 35, pp. 8386–8399, 2022

2022
[16]

How much position in- formation do convolutional neural networks encode?

M. A. Islam, S. Jia, and N. D. Bruce, “How much position in- formation do convolutional neural networks encode?” arXiv preprint arXiv:2001.08248, 2020

2001 arXiv
[17]

Conditional positional encodings for vision transformers,

X. Chu, Z. Tian, B. Zhang et al., “Conditional positional encodings for vision transformers,” arXiv preprint arXiv:2102.10882 , 2021

2021 arXiv
[18]

Conformer: Convolution-augmented transformer for speech recognition,

A. Gulati, J. Qin, C.-C. Chiu et al., “Conformer: Convolution-augmented transformer for speech recognition,” in Proc. Interspeech , 2020, pp. 5036–5040

2020
[19]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar et al. , “Attention is all you need,” Proc. NeurIPS, 2017

2017
[20]

Root mean square layer normalization,

B. Zhang and R. Sennrich, “Root mean square layer normalization,” Proc. NeurIPS, 2019

2019
[21]

GLU variants improve Transformer,

N. Shazeer, “GLU variants improve Transformer,” arXiv preprint arXiv:2002.05202, 2020

2002 arXiv
[22]

An exploration of length general- ization in transformer-based speech enhancement,

Q. Zhang, H. Zhu, X. Qian et al. , “An exploration of length general- ization in transformer-based speech enhancement,” in Proc. Interspeech, 2024, pp. 1725–1729

2024
[23]

Llama 2: Open foundation and fine-tuned chat models,

H. Touvron, L. Martin, K. Stone et al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288 , 2023

2023 arXiv
[24]

The impact of positional encoding on length generalization in transformers,

A. Kazemnejad, I. Padhi, K. Natesan Ramamurthy et al. , “The impact of positional encoding on length generalization in transformers,” Proc. NeurIPS, vol. 36, pp. 24 892–24 928, 2023

2023
[25]

Pvt v2: Improved baselines with pyramid vision transformer,

W. Wang, E. Xie, X. Li et al., “Pvt v2: Improved baselines with pyramid vision transformer,” Computational visual media, vol. 8, no. 3, pp. 415– 424, 2022

2022
[26]

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

M. Maciejewski, G. Wichern, E. McQuinn et al., “WHAMR!: Noisy and reverberant single-channel speech separation,” in Proc. ICASSP, 2020

2020
[27]

J. S. Garofolo et al. , CSR-I (WSJ0) Complete LDC93S6A , Linguistic Data Consortium, Philadelphia, 1993, web Download

1993
[28]

WHAM!: Extending speech separation to noisy environments,

G. Wichern, J. Antognini, M. Flynn et al., “WHAM!: Extending speech separation to noisy environments,” in Proc. Interspeech, 2019

2019
[29]

Music source separation with band-split rnn,

Y . Luo and J. Yu, “Music source separation with band-split rnn,” IEEE/ACM Trans. Audio, Speech, Lang. Process. , vol. 31, pp. 1893– 1901, 2023

1901
[30]

Music source separation with band-split rope transformer,

W.-T. Lu, J.-C. Wang, Q. Kong et al. , “Music source separation with band-split rope transformer,” in Proc. ICASSP, 2024, pp. 481–485

2024
[31]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. ICLR, 2018

2018
[32]

SDR — half-baked or well done?

J. Le Roux, S. Wisdom, H. Erdogan et al., “SDR — half-baked or well done?” in Proc. ICASSP, 2019

2019
[33]

Flashattention: Fast and memory- efficient exact attention with io-awareness,

T. Dao, D. Fu, S. Ermon et al. , “Flashattention: Fast and memory- efficient exact attention with io-awareness,” Proc. NeurIPS, vol. 35, pp. 16 344–16 359, 2022

2022
[34]

What’s all the fuss about free universal sound separation data?

S. Wisdom, H. Erdogan, D. P. Ellis et al., “What’s all the fuss about free universal sound separation data?” in Proc. ICASSP, 2021, pp. 186–190

2021
[35]

Music demixing challenge 2021,

Y . Mitsufuji, G. Fabbro, S. Uhlich et al. , “Music demixing challenge 2021,” Frontiers in Signal Processing , vol. 1, p. 808395, 2022

2021
[36]

MossFormer2: Combining transformer and RNN-free recurrent network for enhanced time-domain monaural speech separation,

S. Zhao, Y . Ma, C. Ni et al. , “MossFormer2: Combining transformer and RNN-free recurrent network for enhanced time-domain monaural speech separation,” in Proc. ICASSP, 2024

2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Deep clustering: Discrimi- native embeddings for segmentation and separation,

J. R. Hershey, Z. Chen, J. Le Roux et al., “Deep clustering: Discrimi- native embeddings for segmentation and separation,” in Proc. ICASSP, 2016

work page 2016

[2] [2]

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

D. Yu, M. Kolbæk, Z. H. Tan et al., “Permutation invariant training of deep models for speaker-independent multi-talker speech separation,” in Proc. ICASSP, 2017

work page 2017

[3] [3]

TasNet: Time-domain audio separation net- work for real-time, single-channel speech separation,

Y . Luo and N. Mesgarani, “TasNet: Time-domain audio separation net- work for real-time, single-channel speech separation,” in Proc. ICASSP, 2018

work page 2018

[4] [4]

Conv-TasNet: Surpassing ideal time-frequency magnitude mask- ing for speech separation,

——, “Conv-TasNet: Surpassing ideal time-frequency magnitude mask- ing for speech separation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 27, no. 8, pp. 1256–1266, 2019

work page 2019

[5] [5]

TFPSNet: Time-frequency domain path scanning network for speech separation,

L. Yang, W. Liu, and W. Wang, “TFPSNet: Time-frequency domain path scanning network for speech separation,” in Proc. ICASSP, 2022

work page 2022

[6] [6]

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

Z.-Q. Wang, S. Cornell, S. Choi et al. , “TF-GridNet: Integrating full- and sub-band modeling for speech separation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 3221–3236, 2023

work page 2023

[7] [7]

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 Proc. ICASSP, 2020

2020

[8] [8]

Attention is all you need in speech separation,

C. Subakan, M. Ravanelli, S. Cornell et al., “Attention is all you need in speech separation,” in Proc. ICASSP, 2021

work page 2021