TGIF: Talker Group-Informed Familiarization of Target Speaker Extraction

Minje Kim; Tsun-An Hsieh

REVIEW 2 major objections 5 minor 38 references

TGIF: Talker Group-Informed Familiarization of Target Speaker Extraction

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TGIF shows that a lightweight student model, fine-tuned on a small talker group via knowledge distillation, can match or exceed a large generalist teacher in target speaker extraction, especially under heavy interference.

desk verdict Novel task framing, but experiments confound talker-group adaptation with acoustic-domain shift, so the central claim is not yet supported. read the letter →

arxiv 2507.14044 v1 pith:WDSLZNQK submitted 2025-07-18 eess.AS

Tsun-An Hsieh , Minje Kim This is my paper · ORCID

classification eess.AS

keywords targetspeakerextractiontalkergroupspecializationknowledgedistillationtest-timeadaptationpersonalizedspeechenhancementmodelcompressionSI-SDR

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 proposes TGIF (talker group-informed familiarization), a new setting for target speaker extraction where the system specializes to a small fixed set of talkers, such as a family, rather than trying to generalize to all speakers. Because clean speech from the group is unavailable after deployment, the paper uses knowledge distillation: a large generalist teacher produces pseudo-clean target estimates, and a smaller student model is fine-tuned on those estimates using group-specific mixtures. On the proposed evaluation protocol, the adapted student specialists improve overall SI-SDR by 1.26 dB (hidden size 128) and 1.08 dB (hidden size 256) over their generalist baselines, and they surpass the teacher in four- and five-speaker conditions. The aim is to show that specialization to a talker group is a practical model-compression route that preserves quality where it matters most.

What carries the argument

The mechanism is a two-stage knowledge distillation pipeline built on the TGIF definition. In stage one, both a large teacher and a small student are pretrained as generalists on a generic multi-talker corpus; the teacher is SpEx+ and the students are TD-SpeakerBeam variants with hidden sizes 128 and 256. In stage two, for each talker group $S(l)$, the frozen teacher processes group-specific mixtures and produces pseudo-clean target estimates $\hat{s}_{k^*}$, which replace the unavailable clean speech in the KD loss $L_{\text{MTL}} := -\text{SI-SDR}(\tilde{s}_{k^*} \| \hat{s}_{k^*}) + \gamma L_{\text{CE}}$. The student, initialized from its generalist checkpoint, is fine-tuned per group, adapting to the group's voices, household noises, and room acoustics. The combinatorial reduction from $\binom{|G|}{K}$ possible speaker mixtures to $\binom{|S(l)|}{K}$ is what makes the small student effective.

What would settle it

Look for a talker group where the teacher's average SI-SDR is low, then adapt a student with teacher pseudo-targets and separately with oracle clean targets; if the KD student's output is no better than the generalist baseline and tracks the teacher's errors rather than the oracle's, the assumption that pseudo-targets are a reliable proxy is falsified for that regime.

Watch

Extended reading notes

Core claim

This paper claims that target speaker extraction can be made group-aware without any clean reference signal at adaptation time, by treating a large pretrained teacher's outputs as proxy targets for a small student. The central empirical discovery is that a TD-SpeakerBeam student with reduced hidden size, fine-tuned per talker group using the SpEx+ teacher's estimates, not only closes much of the gap to the teacher overall but exceeds the teacher's SI-SDR in the four-speaker and five-speaker mixtures, where the group constraint most reduces ambiguity. The gains grow with task difficulty: improvement over the generalist rises from about 0.12 dB for one speaker to more than 3.4 dB for five speakers, and in very low input SDR regions the KD student outperforms the teacher even though it was trained on the teacher's own outputs. Oracle experiments, where ground-truth clean speech replaces teacher outputs, show the ceiling is higher still, with the larger oracle student coming within 0.24 dB of the teacher overall.

Load-bearing premise

The method assumes that the large model's guess at the target voice is a good stand-in for the true clean recording of that voice on the family's own data; if the guess is bad, the small model learns the bad guess and has no clean recording to correct it.

Editorial extensions

If this is right

A device that serves a fixed set of users can deploy a much smaller TSE model and still retain, or even improve, extraction quality in difficult multi-speaker conditions.
Because adaptation needs no clean speech from the owner group, the method is usable after deployment with unlabeled recordings collected by the device.
The benefit of familiarization grows as the mixture gets harder: the reported gains over the generalist student rise from about 0.1 dB for a single speaker to more than 3.4 dB with five speakers.
Specialists can beat their teacher on group data, so for on-device use the expensive generalist teacher does not need to run at inference time.
The same recipe should transfer to other TSE architectures and other group-specific domains, since the loss only requires a teacher estimate and an enrollment signal.

Reading between the lines

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

The teacher's output is the ceiling the student can approach; the 3.28 dB gap between SKD and SKD-Oracle at five speakers suggests that a better teacher, or a confidence-weighted distillation loss, would directly buy more specialist quality.
If the group membership changes, the adaptation may need periodic re-familiarization; a testable extension is to include a speaker-identification gate that detects an out-of-group talker and triggers an update.
The combinatorial view suggests TGIF could be combined with mixture-of-experts routing: each family or group owns a small specialist, and a router selects it by a short enrollment utterance.
Because the adaptation data are unlabeled mixtures, privacy is a selling point, but the teacher must be run once on device or in a trusted environment to generate pseudo-targets; distilling without storing raw mixtures would be a 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

Colleague,

TGIF proposes a useful extension of personalized TSE from one speaker to a small talker group, using a teacher to distill pseudo-clean targets into a lightweight student at test time. That task framing is new and practically motivated. The writing is clear, the code is open, and the oracle-KD variant is a sensible upper bound. I agree with the reader's conditional take on the merits, but the stress-test note lands: the experiments confound talker-group identity with acoustic domain.

What the paper does well: the idea of specializing a small TSE model to a fixed group of users, e.g., a family, is a real gap between single-speaker personalization and a fully generalist TSE. The two-stage KD pipeline is directly inherited from prior PSE work, but applying it to multi-talker groups and showing consistent gains of 1.1–1.3 dB overall (larger in four/five-speaker conditions) is a legitimate positive result. The oracle variant shows the teacher's quality ceiling and helps frame the KD loss.

Where it falls short: the main claim is that TGIF improves TSE by familiarizing the model with the group's talkers. But the generalists are pretrained on AudioSet/Freesound noise and OpenSLR28 RIRs, while adaptation and evaluation use DEMAND household noise and OpenSLR26 small-room RIRs. So the KD student is fine-tuned on both the talker group and a new acoustic domain, and the baseline generalists are evaluated out-of-domain without adaptation. The observed gains could come entirely from acoustic adaptation. A control that adapts under the same acoustic conditions with a held-out speaker group, or that swaps noise/RIR while keeping the talker group, is needed to support the 'talker group-informed' label. The teacher is also never adapted, so the claim that SKD surpasses T in four/five-speaker conditions compares an adapted student to an unadapted teacher. On top of that, there are no error bars or significance tests, and the paper mentions a validation set to prevent overfitting but does not describe how it is used when early stopping is not applied during adaptation.

The single-speaker results show no improvement (actually a small loss for h=256), which is consistent with the idea that the benefits are concentrated in harder mixtures—but that is also where acoustic adaptation matters most.

Bottom line: the concept is worth pursuing, and the paper deserves referee time. I would send it to peer review with a request for a controlled experimental design that isolates talker-group adaptation from acoustic-domain adaptation, plus significance testing. As is, I would not cite its quantitative claims, but I would cite the task formulation if I worked in this area.

Referee Report

2 major / 5 minor

Summary. This paper introduces TGIF, a two-stage knowledge-distillation framework for target speaker extraction (TSE) specialized to a small talker group. A high-capacity SpEx+ teacher and smaller TD-SpeakerBeam students (hidden sizes 128 and 256) are first pretrained as generalists; at deployment, the teacher's outputs on group-specific mixtures are used as pseudo-clean targets to fine-tune the student, avoiding the need for clean speech of the group. Experiments on DNS-derived data with VCTK talkers, DEMAND noise, and OpenSLR26 RIRs report that the KD specialists improve over the student generalists by about 1.1–1.3 dB overall, with larger gains in four- and five-speaker conditions, and that an oracle variant with clean targets approaches the teacher's performance.

Significance. If the reported gains are attributable to talker-group familiarization, TGIF is a practically useful extension of personalized speech enhancement to small groups, and the paper's release of an open-source implementation is a strength. The inclusion of an oracle-KD upper bound and the per-input-SDR analysis in Fig. 2 are also valuable. However, the current experimental design does not isolate the talker-group factor from acoustic domain adaptation, so the central attribution remains unproven; the absence of variance reporting further tempers the strength of the quantitative claims.

major comments (2)

[§3.1.1–3.1.2, Table 1] The experimental design conflates talker-group familiarization with unsupervised domain adaptation to the test acoustic conditions. Generalist pretraining (§3.1.1) uses AudioSet/Freesound noise, OpenSLR28 RIRs, and an SNR range of [-5, 25] dB, whereas test-time adaptation (§3.1.2) uses DEMAND household noise, OpenSLR26 small-room RIRs, a fixed room per group, and an SNR range of [-15, 15] dB. Since T and S are evaluated without adaptation while SKD and SKD-Oracle are fine-tuned on the test-domain mixtures through Eq. (5), the gains in Table 1 could reflect adaptation to the noise/reverberation distribution rather than familiarization with the group's talkers. The claim in §4.2 that SKD surpasses T in four- and five-speaker conditions is likewise a comparison between an adapted student and an unadapted teacher. A control experiment that adapts the student under the same test acoustics with a held-out talker group is needed to attribute the improvement to the talker group itself.
[Table 1, §4.1] The headline results are reported as single mean SI-SDR values without variance, confidence intervals, or significance tests. The overall gains are 1.26 dB (h=128) and 1.08 dB (h=256), while the single-speaker row shows -0.15 dB for h=256, so it is unclear whether the average improvements are statistically reliable or within run-to-run variation. Reporting standard deviations across the 20 groups, per-mixture confidence intervals, and a paired significance test would substantially strengthen the central claim.

minor comments (5)

[Abstract/Introduction] There are minor typos: 'speech font-end applications' should be 'front-end', and 'SpEx+ is a also' should be 'also a'.
[Eq. (5), §3.3] The value of γ during KD adaptation is not stated explicitly. Section 2.3 says the student uses only the reconstruction loss L_SI-SDR, but Eq. (5) includes the L_CE term; the adaptation loss configuration should be made explicit.
[Fig. 2] The color scale of the heatmap and the precise definition of input SDR used for binning are not given in the caption; please add these details.
[Table 1] The SI-SDRi metric is used throughout the table and text but never defined; please define it explicitly (presumably the improvement over the input SI-SDR).
[§3.1.2] The paper assumes clean enrollment segments for group members, but in a realistic deployed family scenario the enrollment could be noisy or reverberant; a brief discussion of this assumption would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KD pseudo-targets bound training, but final scores use clean references.

full rationale

The paper's central claim is that a group-specific student fine-tuned via knowledge distillation improves target speaker extraction relative to generalist baselines, and can even surpass the teacher in hard multi-speaker conditions. The KD loss in Eq. (5) uses the teacher's output as a proxy target, so on the adaptation mixtures the student is trained to imitate the teacher. However, all reported headline results in Table 1 and Fig. 2 are evaluated with SI-SDR computed against clean reference speech, not against the teacher's output. The student is therefore not being scored on its own training target; the evaluation measure is independent of the distillation signal. The claim that the student surpasses the teacher is an empirical outcome on held-out mixtures, not a consequence of the loss definition. The self-citations [23], [28] provide the knowledge-distillation framework as prior art, but the present contribution is a new application to talker-group TSE with its own experimental validation, so the citation is not load-bearing in a circular way. The reviewer's concern about confounding talker-group familiarization with acoustic-domain adaptation is an experimental-design issue, not a circularity: it questions whether the improvement is attributable to the group identity versus the change in noise/RIR conditions, but it does not show that any prediction reduces by construction to its inputs. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the outcome. The derivation chain is empirical and self-contained against clean references.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the domain protocol (fixed group, unlabeled adaptation data, available enrollments) and on the quality of the teacher's pseudo-targets. No new physical entities or mathematical objects are introduced.

free parameters (2)

Test-time SNR range = [-15, 15] dB
The adaptation and evaluation dataset uses a noisier SNR range than the pretraining range ([-5, 25] dB), a design choice that likely makes adaptation benefits more visible.
Hidden size of student models = 128 and 256
Two student capacities are chosen to show the method works for small models; these are architecture choices, not fitted values.

assumptions (3)

domain assumption At test time, the set of talkers is a fixed small group and unlabeled mixtures from this group are available for adaptation
The entire TGIF protocol assumes the system is deployed to a known family-like group and can collect group mixtures without clean targets (Sec. 2.4, 3.1.2).
domain assumption Clean enrollment segments are available for each group member
TSE requires an enrollment utterance to identify the target; the paper assumes per-member clean enrollments exist (Sec. 3.1.2).
domain assumption Teacher model outputs are a sufficiently accurate proxy for clean speech in the group-specific domain
The KD loss (Eq. 5) replaces clean targets with teacher estimates, so the student inherits any teacher bias; this is the central limitation acknowledged indirectly by the SKD vs SKD-Oracle gap.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TGIF: Talker Group-Informed Familiarization of Target Speaker Extraction." pith.science (2026). https://pith.science/paper/WDSLZNQK

@misc{pith2026250714044,
  author       = {Pith},
  title        = {Pith review of: TGIF: Talker Group-Informed Familiarization of Target Speaker Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDSLZNQK}},
  note         = {Machine review of arXiv:2507.14044}
}

read the original abstract

State-of-the-art target speaker extraction (TSE) systems are typically designed to generalize to any given mixing environment, necessitating a model with a large enough capacity as a generalist. Personalized speech enhancement could be a specialized solution that adapts to single-user scenarios, but it overlooks the practical need for customization in cases where only a small number of talkers are involved, e.g., TSE for a specific family. We address this gap with the proposed concept, talker group-informed familiarization (TGIF) of TSE, where the TSE system specializes in a particular group of users, which is challenging due to the inherent absence of a clean speech target. To this end, we employ a knowledge distillation approach, where a group-specific student model learns from the pseudo-clean targets generated by a large teacher model. This tailors the student model to effectively extract the target speaker from the particular talker group while maintaining computational efficiency. Experimental results demonstrate that our approach outperforms the baseline generic models by adapting to the unique speech characteristics of a given speaker group. Our newly proposed TGIF concept underscores the potential of developing specialized solutions for diverse and real-world applications, such as on-device TSE on a family-owned device.

Figures

Figures reproduced from arXiv: 2507.14044 by the authors.

**Figure 2.** SI-SDRi trends across input SDR. To evaluate the effectiveness of adaptation under both ideal and more realistic conditions where clean references may not be available, the adaptation process considers both oracle and teacher-forcing adaptation scenarios. In the oracle setup, the student specialists are fine-tuned using the ground truth clean target speech as the supervision signal. In the teacher-forcing setup, whi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

[1]

Univer- sal speech enhancement with score-based diffusion,

J. Serr `a, S. Pascual, J. Pons, R. O. Araz, and D. Scaini, “Univer- sal speech enhancement with score-based diffusion,” arXiv preprint arXiv:2206.03065, 2022

arXiv 2022
[2]

Miipher: A Robust Speech Restoration Model Integrating Self-Supervised Speech and Text Representations,

Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, Y . Zhang, W. Han, A. Bapna, and M. Bacchiani, “Miipher: A Robust Speech Restoration Model Integrating Self-Supervised Speech and Text Representations,” in Proc. of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , 2023

work page 2023
[3]

Genhancer: High-Fidelity Speech Enhancement via Generative Modeling on Discrete Codec Tokens,

H. Yang, J. Su, M. Kim, and Z. Jin, “Genhancer: High-Fidelity Speech Enhancement via Generative Modeling on Discrete Codec Tokens,” in Proc. Interspeech, 2024

work page 2024
[4]

Storm: A diffusion-based stochastic regeneration model for speech enhancement and dereverberation,

J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Storm: A diffusion-based stochastic regeneration model for speech enhancement and dereverberation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, p. 2724–2737, 2023

work page 2023
[5]

Schr ¨odinger bridge for generative speech enhancement,

J. A., R. Korostik, J. Balam, and B. Ginsburg, “Schr ¨odinger bridge for generative speech enhancement,” in Proc. Interspeech, 2024

work page 2024
[6]

Universal Score- based Speech Enhancement with High Content Preservation,

R. Scheibler, Y . Fujita, Y . Shirahata, and T. Komatsu, “Universal Score- based Speech Enhancement with High Content Preservation,” in Proc. Interspeech, 2024

work page 2024
[7]

Separate and diffuse: Using a pretrained diffusion model for better source separation,

S. Lutati, E. Nachmani, and L. Wolf, “Separate and diffuse: Using a pretrained diffusion model for better source separation,” in Proc. of the International Conference on Learning Representations (ICLR) , 2024

work page 2024
[8]

MossFormer2: Combining Transformer and RNN-Free Recurrent Network for Enhanced Time-Domain Monaural Speech Separation,

S. Zhao, Y . Ma, C. Ni, C. Zhang, H. Wang, T. H. Nguyen, K. Zhou, J. Yip, D. Ng, and B. Ma, “MossFormer2: Combining Transformer and RNN-Free Recurrent Network for Enhanced Time-Domain Monaural Speech Separation,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2024

work page 2024

Show all 38 references

[9]

Exploring self-attention mechanisms for speech separation,

C. Subakan, M. Ravanelli, S. Cornell, F. Grondin, and M. Bronzi, “Exploring self-attention mechanisms for speech separation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 2169–2180, 2023

2023
[10]

Wavesplit: End-to-end speech separation by speaker clustering,

N. Zeghidour and D. Grangier, “Wavesplit: End-to-end speech separation by speaker clustering,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 2840–2849, 2021

2021
[11]

Single Channel Target Speaker Extraction and Recognition with Speaker Beam,

M. Delcroix, K. ˇZmol´ıkov´a, K. Kinoshita, A. Ogawa, and T. Nakatani, “Single Channel Target Speaker Extraction and Recognition with Speaker Beam,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2018

2018
[12]

Improving speaker discrimination of target speech extraction with time-domain speakerbeam,

M. Delcroix, T. Ochiai, K. Zmolikova, K. Kinoshita, N. Tawara, T. Nakatani, and S. Araki, “Improving speaker discrimination of target speech extraction with time-domain speakerbeam,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICA...

2020
[13]

Spex: Multi-scale time domain speaker extraction network,

C. Xu, W. Rao, E. S. Chng, and H. Li, “Spex: Multi-scale time domain speaker extraction network,” Proc. of the IEEE Workshop on Machine Learning for Signal Processing (MLSP) , 2020

2020
[14]

Spex+: A complete time domain speaker extraction network,

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

2020
[15]

Increasing compactness of deep learning based speech enhancement models with parameter pruning and quantization techniques,

J.-Y . Wu, C. Yu, S.-W. Fu, C.-T. Liu, S.-Y . Chien, and Y . Tsao, “Increasing compactness of deep learning based speech enhancement models with parameter pruning and quantization techniques,” IEEE Signal Processing Letters, vol. 26, no. 12, pp. 1887–1891, 2019

2019
[16]

Incremental binarization on recurrent neural networks for single-channel source separation,

S. Kim, M. Maity, and M. Kim, “Incremental binarization on recurrent neural networks for single-channel source separation,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2019

2019
[17]

Towards model compression for deep learning based speech enhancement,

K. Tan and D. Wang, “Towards model compression for deep learning based speech enhancement,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 1785–1794, 2021

2021
[18]

Towards fully quantized neural networks for speech enhancement,

E. Cohen, H. V . Habi, and A. Netzer, “Towards fully quantized neural networks for speech enhancement,” Proc. Interspeech, 2023

2023
[19]

Dynamic slimmable network for speech separation,

M. Elminshawi, S. R. Chetupalli, and E. A. P. Habets, “Dynamic slimmable network for speech separation,” IEEE Signal Processing Letters, vol. 31, pp. 2205–2209, 2024

2024
[20]

Scalable speech enhancement with dynamic channel pruning,

R. Miccini, C. Laroche, T. Piechowiak, and L. Pezzarossa, “Scalable speech enhancement with dynamic channel pruning,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2025

2025
[21]

Zero-shot personalized speech enhancement through speaker-informed model selection,

A. Sivaraman and M. Kim, “Zero-shot personalized speech enhancement through speaker-informed model selection,” in Proc. of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2021

2021
[22]

Speech enhancement with zero-shot model selection,

R. E. Zezario, C.-S. Fuh, H.-M. Wang, and Y . Tsao, “Speech enhancement with zero-shot model selection,” in Proc. of the European Signal Processing Conference (EUSIPCO), 2021

2021
[23]

Zero-Shot Test- Time Adaptation Via Knowledge Distillation for Personalized Speech Denoising and Dereverberation,

S. Kim, M. Athi, G. Shi, M. Kim, and T. Kristjansson, “Zero-Shot Test- Time Adaptation Via Knowledge Distillation for Personalized Speech Denoising and Dereverberation,” Journal of the Acoustical Society of America, vol. 155, no. 2, pp. 1353–1367, 2024

2024
[24]

Distilling the knowledge in a neural network,

G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

2015 arXiv
[25]

Efficient Personalized Speech Enhancement Through Self-Supervised Learning,

A. Sivaraman and M. Kim, “Efficient Personalized Speech Enhancement Through Self-Supervised Learning,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1342–1356, 2022

2022
[26]

Personalized Speech Enhancement: New Models and Comprehensive Evaluation,

S. E. Eskimez, T. Yoshioka, H. Wang, X. Wang, Z. Chen, and X. Huang, “Personalized Speech Enhancement: New Models and Comprehensive Evaluation,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2022

2022
[27]

ICASSP 2021 Deep Noise Suppression Challenge,

C. K. A. Reddy, H. Dubey, V . Gopal, R. Cutler, S. Braun, H. Gamper, R. Aichner, and S. Srinivasan, “ICASSP 2021 Deep Noise Suppression Challenge,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2021

2021
[28]

Test-time adaptation toward personalized speech enhancement: Zero-shot learning with knowledge distillation,

S. Kim and M. Kim, “Test-time adaptation toward personalized speech enhancement: Zero-shot learning with knowledge distillation,” in Proc. of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2021

2021
[29]

SDR – half- baked or well done?

J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR – half- baked or well done?” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2019

2019
[30]

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

K. ˇZmol´ıkov´a, M. Delcroix, K. Kinoshita, T. Ochiai, T. Nakatani, L. Burget, and J. ˇCernock´y, “Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,”IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 4, pp. 800–814, 2019

2019
[31]

Librivox: Free public domain audiobooks,

J. Kearns, “Librivox: Free public domain audiobooks,” Reference Reviews, vol. 28, no. 1, pp. 7–8, 2014

2014
[32]

Audio Set: An Ontology and Human-Labeled Dataset for Audio Events,

J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio Set: An Ontology and Human-Labeled Dataset for Audio Events,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2017

2017
[33]

Freesound datasets: A platform for the creation of open audio datasets

E. Fonseca, J. Pons, X. Favory, F. Font, D. Bogdanov, A. Ferraro, S. Oramas, A. Porter, and X. Serra, “Freesound datasets: A platform for the creation of open audio datasets.” in Proc. of the International Society for Music Information Retrieval Conference (ISMIR) , 2017

2017
[34]

A study on data augmentation of reverberant speech for robust speech recognition,

T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2017

2017
[35]

Superseded-cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit,

C. Veaux, J. Yamagishi, K. MacDonald et al. , “Superseded-cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit,” URL http://datashare.is.ed.ac.uk/handle/10283/2651, 2016

2016
[36]

The Diverse Environments Multi-channel Acoustic Noise Database (DEMAND): A database of multichannel environmental noise recordings,

J. Thiemann, N. Ito, and E. Vincent, “The Diverse Environments Multi-channel Acoustic Noise Database (DEMAND): A database of multichannel environmental noise recordings,” Journal of the Acoustical Society of America , vol. 133, no. 5, pp. 3591–3591, 2013

2013
[37]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) , 2016, pp. 770–778

2016
[38]

Adam: A method for stochastic optimization,

D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. of the International Conference on Learning Representations (ICLR), 2015

2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Univer- sal speech enhancement with score-based diffusion,

J. Serr `a, S. Pascual, J. Pons, R. O. Araz, and D. Scaini, “Univer- sal speech enhancement with score-based diffusion,” arXiv preprint arXiv:2206.03065, 2022

arXiv 2022

[2] [2]

Miipher: A Robust Speech Restoration Model Integrating Self-Supervised Speech and Text Representations,

Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, Y . Zhang, W. Han, A. Bapna, and M. Bacchiani, “Miipher: A Robust Speech Restoration Model Integrating Self-Supervised Speech and Text Representations,” in Proc. of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , 2023

work page 2023

[3] [3]

Genhancer: High-Fidelity Speech Enhancement via Generative Modeling on Discrete Codec Tokens,

H. Yang, J. Su, M. Kim, and Z. Jin, “Genhancer: High-Fidelity Speech Enhancement via Generative Modeling on Discrete Codec Tokens,” in Proc. Interspeech, 2024

work page 2024

[4] [4]

Storm: A diffusion-based stochastic regeneration model for speech enhancement and dereverberation,

J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Storm: A diffusion-based stochastic regeneration model for speech enhancement and dereverberation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, p. 2724–2737, 2023

work page 2023

[5] [5]

Schr ¨odinger bridge for generative speech enhancement,

J. A., R. Korostik, J. Balam, and B. Ginsburg, “Schr ¨odinger bridge for generative speech enhancement,” in Proc. Interspeech, 2024

work page 2024

[6] [6]

Universal Score- based Speech Enhancement with High Content Preservation,

R. Scheibler, Y . Fujita, Y . Shirahata, and T. Komatsu, “Universal Score- based Speech Enhancement with High Content Preservation,” in Proc. Interspeech, 2024

work page 2024

[7] [7]

Separate and diffuse: Using a pretrained diffusion model for better source separation,

S. Lutati, E. Nachmani, and L. Wolf, “Separate and diffuse: Using a pretrained diffusion model for better source separation,” in Proc. of the International Conference on Learning Representations (ICLR) , 2024

work page 2024

[8] [8]

MossFormer2: Combining Transformer and RNN-Free Recurrent Network for Enhanced Time-Domain Monaural Speech Separation,

S. Zhao, Y . Ma, C. Ni, C. Zhang, H. Wang, T. H. Nguyen, K. Zhou, J. Yip, D. Ng, and B. Ma, “MossFormer2: Combining Transformer and RNN-Free Recurrent Network for Enhanced Time-Domain Monaural Speech Separation,” in Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) , 2024

work page 2024