Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

An Investigation of Test-time Adaptation for Audio Classification under Background Noise

T0 review · 4 major / 6 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 This paper claims that a modified version of the CoNMix test-time adaptation method achieves the highest audio classification accuracy under background noise, outperforming TTT and TENT on AudioMNIST and SpeechCommands.

desk verdict A readable but methodologically invalid comparison: the 'modified CoNMix' is selected on test-set errors and compared under unequal budgets, so the headline claim does not stand. read the letter →

arxiv 2507.15523 v1 pith:HOZ2C6Q4 submitted 2025-07-21 cs.LG cs.SDeess.AS

Weichuang Shao , Iman Yi Liao , Tomas Henrique Bode Maul , Tissa Chandesa This is my paper
classification cs.LGcs.SDeess.AS
keywords test-timeadaptationaudioclassificationbackgroundnoisedomainshiftCoNMixTENTTTTMel-spectrogram
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

Under background-noise domain shift, this paper asks whether test-time adaptation (TTA) methods designed for images can rescue audio classifiers that only saw clean speech. The authors convert audio to Mel-spectrograms and apply three TTA methods — TTT, TENT, and CoNMix — to AudioMNIST and SpeechCommands V1 with added noise at 3 dB and 10 dB SNR. Their central claim is that a modified CoNMix, which adjusts the pseudo-label loss per dataset, delivers the lowest error rates, down to 5.31% under 10 dB exercise-bike noise and 12.75% under 3 dB running-tap noise on AudioMNIST. They also report that TENT and TTT can hurt performance on SpeechCommands, a negative adaptation effect they believe has not been reported for audio classification before. If the claim holds, it points to CoNMix as the most reliable of the three TTA approaches for noisy audio.

What carries the argument

The load-bearing mechanism is the CoNMix teacher-student test-time adaptation framework, which combines three losses — nuclear-norm maximisation, pseudo-label cross-entropy, and consistency between strong and weak augmentations — to adapt a Vision Transformer on the entire unlabelled test set before making predictions. For audio, the paper converts raw waveforms to Mel-spectrograms so these image-oriented methods apply, replacing image augmentations with time shifts. The proposed modification changes only the pseudo-label component: it removes that loss for AudioMNIST and swaps softmax cross-entropy for negative log-likelihood loss on SpeechCommands, because the original pseudo-label loss caused negative adaptation on that dataset.

What would settle it

Run TENT and TTT with the same Vision Transformer backbone as CoNMix, the full test set, and the same number of adaptation epochs on the same noise conditions; if either method then matches or beats modified CoNMix's error rates on AudioMNIST and SpeechCommands, the claim that CoNMix is the most reliable TTA method would be refuted.

Watch

Extended reading notes

Core claim

The paper's core discovery is that CoNMix, when modified so that its pseudo-label loss is removed for AudioMNIST and replaced with a negative log-likelihood loss for SpeechCommands, consistently improves accuracy under background noise across both datasets and all tested noise conditions. In contrast, TENT and TTT often fail: on SpeechCommands their adapted error rates exceed the unadapted model's error rates, and TTT performs poorly on AudioMNIST. The paper attributes CoNMix's stability to its use of the full test set over many epochs and its Vision Transformer backbone, and it claims this is the first study to apply TTA to audio classification under domain shift.

Load-bearing premise

The comparison assumes CoNMix's advantage comes from its adaptation strategy rather than from its larger Vision Transformer backbone and its access to the whole test set for many epochs, neither of which the other two methods are given.

Editorial extensions

If this is right

  • Audio classifiers deployed under real-world background noise can be protected without retraining or labels by using a full-test-set adaptation method like modified CoNMix.
  • TENT and TTT, though effective in vision, should not be assumed safe for audio: on SpeechCommands they can raise error rates above the unadapted baseline.
  • Pseudo-label quality, not just entropy minimisation, is the fragile component of CoNMix when the domain shift is acoustic, since adjusting that loss is what removes the negative effect.
  • Converting audio to Mel-spectrograms is a viable bridge that lets image-oriented TTA methods be evaluated on audio tasks, which may speed up transfer of future vision TTA ideas.

Reading between the lines

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

  • Editorial inference: a fairer comparison that gives TENT and TTT the same Vision Transformer backbone and multi-epoch full-test-set access could substantially shrink or eliminate CoNMix's measured advantage, since the paper does not isolate adaptation strategy from architecture and adaptation budget.
  • Editorial inference: the negative adaptation of TENT on SpeechCommands may reflect batch-normalisation statistics being dominated by the noisy test distribution, a mechanism the paper suggests but does not test directly.
  • Editorial inference: the 'first study' claim is sensitive to scope; TTA has been applied to speech recognition and enhancement, so the novelty is best read as first for audio classification specifically, not for audio processing generally.
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.

Referee Report

4 major / 6 minor

Summary. The paper investigates test-time adaptation (TTA) methods—TTT, TENT, and CoNMix—for audio classification under background noise. The authors convert audio to mel-spectrograms, add real-world noise (doing dishes, exercise bike, running tap) at 3 dB and 10 dB SNR to test sets of AudioMNIST (AM) and SpeechCommands V1 (SC) and its variants, and report classification error rates with and without adaptation. They propose a modified CoNMix that omits the pseudo-label loss for AM and replaces softmax+CE with log-softmax+NLL for SC, and report that this modified CoNMix achieves the lowest error rates, calling it the most reliable TTA method. They also claim this is the first study of TTA for audio classification under background noise.

Significance. If the empirical claims were valid, this would be a useful comparative study of TTA methods for a new modality (audio classification under background noise), with practical relevance for deployment in noisy environments. The paper is transparent about its ablations and includes a code link, and it explicitly acknowledges several confounds in §5.3. However, the central numerical claim is undermined by a protocol that selects the CoNMix variant after observing test-set errors (Appendix A.3), by uncontrolled differences in architecture and adaptation budget, and by the absence of any repeated-run statistics. As presented, the results do not provide a valid test of the claimed TTA advantage; the positive aspects are the clear problem framing, the multi-noise/multi-SNR evaluation, and the honest disclosure of the post-hoc selection.

major comments (4)
  1. [Appendix A.3 and §3.5] The choice of CoNMix variant is made after inspecting test-set error rates. Appendix A.3 states that 'in a post-mortem manner, we retrospectively manipulated the scaling factor λ2' and that for AM the pseudo-label loss was removed entirely, while §3.5 describes a different replacement (log-softmax + NLL) for SC. These per-dataset decisions were made using the same test sets whose error rates are then reported as the method's performance in Tables 2 and 4. This makes the headline claim that modified CoNMix is 'the most reliable TTA method' circular: the reported numbers are test-set-tuned, not a valid evaluation of a fixed TTA algorithm. A correct protocol would fix the loss variant and λ2 a priori (e.g., using a validation split, as exists for SC) or report all variants without declaring a single winner.
  2. [Table 2] The row labeled 'SC EB-3' appears twice with different values: '83.46 86.2 21.64 14.13' and '34.27 41.43 24.98 14.79'. The second row must be a copy-paste error, almost certainly intended to be 'SCR EB-3', given the pattern for DD and RT. This error directly affects the data presented for one of the two main datasets and needs to be corrected; more generally, this call for careful checking of all entries in Tables 2–4.
  3. [§5.3] The comparison is confounded by architecture and adaptation budget. CoNMix uses a Vision Transformer and adapts for many epochs on the entire test set, while TENT and TTT use ResNet backbones and one epoch. The paper acknowledges these differences in §5.3 but the follow-up experiment in Figure 4 varies only the epoch count for TENT and TTT; it does not control for architecture (ResNet vs. ViT) or for adaptation protocol (online mini-batch vs. full test set). Consequently, the superior error rates of the modified CoNMix could be due to the larger backbone, the full-set multi-epoch budget, or the test-set-tuned loss configuration, rather than to the CoNMix TTA strategy itself. Without a controlled comparison—for example, running all methods on the same backbone with the same adaptation budget and the same a-priori loss configuration—the central claim is not established.
  4. [Tables 2, 3, and 4] No error bars, confidence intervals, or repeated runs are reported for any of the experimental results. All numbers are single deterministic runs, so small differences (e.g., 5.31% vs. 5.37% error on AM EB-10) cannot be distinguished from run-to-run variation. The claims that one method is 'most stable' or 'most reliable' require at least three or more random seeds per configuration, especially given the stochasticity of TTA updates and the randomness in noise clip selection (§4.2).
minor comments (6)
  1. [§3.4.3, Eq. (11)] The two terms in Eq. (11) are identical: both are λ·CE(ỹ_ij, ŷ_ij) and (1−λ)·CE(ỹ_ij, ŷ_ij), which simplifies to CE(ỹ_ij, ŷ_ij). The equation presumably should distinguish the two contributions (e.g., teacher-student or two different synthetic branches); as written, it is nonsensical and should be corrected or removed.
  2. [§3.5] The text refers to 'SC, SCR, and SCN' before the SCN dataset has been introduced; these abbreviations should be defined at first use in §3.5 or the section should refer to Section 4.1.
  3. [Table 4 description] The table caption mentions 'under background noise conditions EB, DD, and TR' but the rows use 'RT' for running tap. The caption should use 'RT' consistently.
  4. [§1 and §6] The claim of being 'the first study to leverage TTA techniques for audio classification under domain shift' is too strong, since the related-work section discusses several TTA works in audio (ASR, speech enhancement, pathological speech detection). The novelty claim should be narrowed to the specific setup of TTA for audio classification with real-world background noise, or supported by a more systematic literature search.
  5. [Figures 1–7] The figure captions are often telegraphic (e.g., 'Figure 1: Demonstrates the Tent Adaptation and Norm Adaptation error rates for SC, SCR, and SCN, under 10 dB SNR level EB background noise.') and the plots are not described in the text; adding axis labels, legend explanations, and a short textual reference would improve comprehension.
  6. [§4.4] The sentence 'a minimum of 26.72% increase in error rate' seems to refer to AM across all methods except CoNMix; clarify that this is the minimum increase among the other methods on AM, and check the arithmetic for consistency with Table 2.

Circularity Check

1 steps flagged · score 6.0 of 10

CoNMix's reported advantage is partly a test-set-tuned number: the pseudo-label loss was removed or replaced per dataset after inspecting the noisy test sets, so the headline accuracy is not the accuracy of a pre-specified TTA method.

  1. fitted input called prediction [Appendix A.3; Section 3.5; Tables 2 and 4]
    "It is worth noting that for CoNMix, the pseudo-label loss function (see Eq.9 and Eq.18) contributed in a limited way to test-time adaptation performance. To tackle the issue in a post-mortem manner, we retrospectively manipulated the scaling factor λ2 of the pseudo-label loss function. Specifically, for the results shown in Table 2 section 5, we removed the pseudo-label loss entirely for the experiment on AM under 3.0 dB and 10.0 dB SNR levels in Eq.7."

    The paper's central claim is that the proposed modified CoNMix achieves the highest accuracy (5.31% EB-10, 12.75% RT-3 on AM). But the modified configuration is not fixed before testing: Appendix A.3 says the scaling factor λ2 was 'retrospectively manipulated' and that the pseudo-label loss was removed entirely for AM, while Section 3.5 states that for SC the pseudo-label loss was replaced with NLL after an ablation 'identified that the inclusion of pseudo label loss (Eq.9) was responsible for the negative adaptation effect.' Both choices were made by inspecting the same noisy test sets whose error rates are then reported as the method's performance.

full rationale

This is an empirical comparison paper, not a formal derivation, so most classic circular-derivation patterns are absent: there is no self-citation chain, no uniqueness theorem, and no renamed known result. The one genuine circularity is in the evaluation: the 'modified CoNMix' whose accuracy is reported was chosen post hoc from test-set performance. Appendix A.3 explicitly describes retrospectively manipulating λ2 and removing the pseudo-label loss for AM on the exact experiments reported in Table 2; Section 3.5 describes switching to NLL for SC after an ablation on the same test data. The reported 5.31% and 12.75% error rates therefore measure a test-set-tuned configuration, not the performance of a fixed TTA method, so the claim that modified CoNMix is the most reliable TTA approach is partly forced by construction. The paper also acknowledges in Section 5.3 that CoNMix uses a larger ViT backbone, the entire test set, and many epochs, while TENT/TTT use ResNet and one epoch; this is a serious confound for the comparison but is not itself circularity, so it is noted rather than scored as a separate circular step. No load-bearing self-citations were found. Score 6 reflects that the headline result is partially circular (test-set selection of the method), while the underlying components (TTT, TENT, CoNMix, the ablation losses) remain external and independently defined. If the authors had fixed the modified CoNMix variant a priori or chosen it on a validation set, the circularity would not arise.

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

No new physical or architectural entities are introduced. The central claim rests on standard audio features, an unvalidated transfer of image TTA methods, a fragile pseudo-labeling step, and post-hoc per-dataset loss selection.

free parameters (4)
  • CoNMix loss scaling factors λ1, λ2, λ3 = not reported
    Weights of the nuclear-norm, pseudo-label, and consistency losses in Eq.7 and Eq.19. The paper states they are real scaling factors but never gives values. λ2 is set to zero for AM and retained for SC after the post-hoc analysis.
  • TTT time-shift percentage = not reported
    Section 3.2 says audio is shifted left or right by 'a specified percentage' using PyTorch roll, but the percentage is not given. It directly defines the pretext task used for adaptation.
  • Number of test-time adaptation epochs for CoNMix = not reported
    Section 5.3 reports results at 1 and 10 epochs, but the headline CoNMix numbers in Table 2 do not state which epoch count was used. This is central because CoNMix's multi-epoch full-set adaptation is a stated reason for its advantage.
  • Gaussian noise severity λ = 0.005 in the displayed experiment, otherwise unspecified
    Eq.21 defines a Gaussian distribution shift for the TENT probe experiments. Only one value is shown in Figure 2, and the choice is not justified.
assumptions (4)
  • domain assumption Mel-spectrogram conversion preserves the class-relevant information after background noise is added.
    Section 4.2 converts audio to mel-spectrograms to reuse image-oriented TTA methods. This is standard, but the paper does not quantify how much information is lost at 3 dB SNR.
  • domain assumption Image-oriented TTA methods remain valid when the input modality is changed and the rotation pretext is replaced by time shift.
    Sections 3.2 and 4.2 make this transfer without a validation study or an analysis of whether time shift is as informative for audio as rotation is for images.
  • domain assumption Pseudo-labels generated by class centers and iterative refinement are reliable enough to drive test-time adaptation.
    Section 3.4.2 describes the pseudo-label process. The paper's own ablation shows that the pseudo-label loss causes negative adaptation on SpeechCommands, so this assumption is fragile and dataset-dependent.
  • ad hoc to paper Dataset-specific loss variants (drop pseudo-label for AM, use NLL for SC) are chosen under a valid TTA protocol that does not use test labels.
    Appendix A.3 says the modifications were made 'post-mortem' and 'retrospectively' using test-set error rates. This contradicts the no-label premise of test-time adaptation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Investigation of Test-time Adaptation for Audio Classification under Background Noise." pith.science (2026). https://pith.science/paper/HOZ2C6Q4

@misc{pith2026250715523,
  author       = {Pith},
  title        = {Pith review of: An Investigation of Test-time Adaptation for Audio Classification under Background Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOZ2C6Q4}},
  note         = {Machine review of arXiv:2507.15523}
}
read the original abstract

Domain shift is a prominent problem in Deep Learning, causing a model pre-trained on a source dataset to suffer significant performance degradation on test datasets. This research aims to address the issue of audio classification under domain shift caused by background noise using Test-Time Adaptation (TTA), a technique that adapts a pre-trained model during testing using only unlabelled test data before making predictions. We adopt two common TTA methods, TTT and TENT, and a state-of-the-art method CoNMix, and investigate their respective performance on two popular audio classification datasets, AudioMNIST (AM) and SpeechCommands V1 (SC), against different types of background noise and noise severity levels. The experimental results reveal that our proposed modified version of CoNMix produced the highest classification accuracy under domain shift (5.31% error rate under 10 dB exercise bike background noise and 12.75% error rate under 3 dB running tap background noise for AM) compared to TTT and TENT. The literature search provided no evidence of similar works, thereby motivating the work reported here as the first study to leverage TTA techniques for audio classification under domain shift.

Figures

Figures reproduced from arXiv: 2507.15523 by the authors.

Figure 2
Figure 2. Includes the Tent Adaptation and Norm Adapta [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 1
Figure 1. Demonstrates the Tent Adaptation and Norm [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 3
Figure 3. Fig.3. It is clear that TENT performed poorly on SC (error [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Demonstrates the test-time adaptation error rate [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Depicts error rates when processing the uncor [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The accuracy results at 10 dB ((a), (b), (c), and (d)) and 3 dB ((g), (h), (i), and (j)) SNR level for SC. To [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: The accuracy results at 10 dB ((a), (b), (c), and (d)) and 3 dB ((g), (h), (i), and (j)) SNR level for AM. To [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. DHAuDS: A Dynamic and Heterogeneous Audio Benchmark for Test-Time Adaptation

    cs.SD 2025-11 conditional novelty 6.0 of 10

    DHAuDS is a new audio benchmark that corrupts four existing datasets with dynamically varying and diverse acoustic noise, and evaluates three classifiers under test-time adaptation.

Reference graph

Works this paper leans on

44 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Test-time adaptation for automatic pathological speech detection in noisy environ- ments

    Mahdi Amiri and Ina Kodrasi. Test-time adaptation for automatic pathological speech detection in noisy environ- ments. In2024 32nd European Signal Processing Conference (EUSIPCO), pages 86–90, 2024

    work page 2024

  2. [2]

    Self-supervised test-time adaptation on video data

    Fatemeh Azimi, Sebastian Palacio, Federico Raue, J¨ orn Hees, Luca Bertinetto, and Andreas Dengel. Self-supervised test-time adaptation on video data. InProceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 3439–3448, 2022

    work page 2022

  3. [3]

    Audiomnist: Exploring explainable artificial intelli- gence for audio analysis on a simple benchmark.Journal of the Franklin Institute, 2023

    S¨ oren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert M¨ uller, Sebastian Lapuschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelli- gence for audio analysis on a simple benchmark.Journal of the Franklin Institute, 2023

    work page 2023

  4. [4]

    Self-supervision & meta- learning for one-shot unsupervised cross-domain detection

    Francesco Cappio Borlino, Salvatore Polizzotto, Barbara Caputo, and Tatiana Tommasi. Self-supervision & meta- learning for one-shot unsupervised cross-domain detection. Computer Vision and Image Understanding, 223:103549, 2022

    work page 2022

  5. [5]

    Source-free temporal at- tentive domain adaptation for video action recognition

    Peipeng Chen and Andy J Ma. Source-free temporal at- tentive domain adaptation for video action recognition. In Proceedings of the 2022 International Conference on Multi- media Retrieval, pages 489–497, 2022

    work page 2022

  6. [6]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  7. [7]

    Adaptive methods for real-world do- main generalization

    Abhimanyu Dubey, Vignesh Ramanathan, Alex Pentland, and Dhruv Mahajan. Adaptive methods for real-world do- main generalization. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 14340–14349, 2021

    work page 2021

  8. [8]

    A brief review of domain adaptation

    Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: pro- ceedings from ICDATA 2020 and IKE 2020, pages 877–894, 2021

    work page 2020

Show all 44 references
  1. [9]

    Model- agnostic meta-learning for fast adaptation of deep net- works

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model- agnostic meta-learning for fast adaptation of deep net- works. InInternational conference on machine learning, pages 1126–1135. PMLR, 2017

    2017

  2. [10]

    MIT press, 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep learning. MIT press, 2016

    2016

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  4. [12]

    Meta-learning in neural networks: A survey

    Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelli- gence, 44(9):5149–5169, 2021

    2021

  5. [13]

    Visualizing adapted knowledge in domain transfer

    Yunzhong Hou and Liang Zheng. Visualizing adapted knowledge in domain transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13824–13833, 2021

    2021

  6. [14]

    Mel-spectrogram augmentation for sequence to sequence voice conversion

    Yeongtae Hwang, Hyemin Cho, Hongsun Yang, Dong-Ok Won, Insoo Oh, and Seong-Whan Lee. Mel-spectrogram augmentation for sequence to sequence voice conversion. arXiv preprint arXiv:2001.01401, 2020

    2001 arXiv

  7. [15]

    Sgem: Test-time adaptation for automatic speech recognition via sequential-level generalized entropy mini- mization.arXiv preprint arXiv:2306.01981, 2023

    Changhun Kim, Joonhyung Park, Hajin Shim, and Eunho Yang. Sgem: Test-time adaptation for automatic speech recognition via sequential-level generalized entropy mini- mization.arXiv preprint arXiv:2306.01981, 2023

    2023 arXiv

  8. [16]

    Variational on-the-fly personalization

    Jangho Kim, Jun-Tae Lee, Simyung Chang, and Nojun Kwak. Variational on-the-fly personalization. InInterna- tional Conference on Machine Learning, pages 11134–11147. PMLR, 2022

    2022

  9. [17]

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

    Sunwoo Kim and Minje Kim. Test-time adaptation toward personalized speech enhancement: Zero-shot learning with knowledge distillation. In2021 IEEE Workshop on Appli- cations of Signal Processing to Audio and Acoustics (WAS- PAA), pages 176–180. IEEE, 2021

    2021

  10. [18]

    Lattice-based unsupervised test-time adapta- tion of neural network acoustic models.arXiv preprint arXiv:1906.11521, 2019

    Ondrej Klejch, Joachim Fainberg, Peter Bell, and Steve Renals. Lattice-based unsupervised test-time adapta- tion of neural network acoustic models.arXiv preprint arXiv:1906.11521, 2019

    1906 arXiv

  11. [19]

    Domain adap- tation by mixture of alignments of second-or higher-order scatter tensors

    Piotr Koniusz, Yusuf Tas, and Fatih Porikli. Domain adap- tation by mixture of alignments of second-or higher-order scatter tensors. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4478–4487, 2017

    2017

  12. [20]

    Conmix for source-free single and multi-target domain adaptation

    Vikash Kumar, Rohit Lal, Himanshu Patil, and Anirban Chakraborty. Conmix for source-free single and multi-target domain adaptation. InProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision, pages 4178–4188, 2023

    2023

  13. [21]

    Model adaptation: Unsupervised domain adaptation without source data

    Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 9641–9650, 2020

    2020

  14. [22]

    Imbalanced source-free domain adaptation

    Xinhao Li, Jingjing Li, Lei Zhu, Guoqing Wang, and Zi Huang. Imbalanced source-free domain adaptation. In Proceedings of the 29th ACM international conference on multimedia, pages 3330–3339, 2021

    2021

  15. [23]

    A comprehensive sur- vey on test-time adaptation under distribution shifts.Inter- national Journal of Computer Vision, pages 1–34, 2024

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive sur- vey on test-time adaptation under distribution shifts.Inter- national Journal of Computer Vision, pages 1–34, 2024

    2024

  16. [24]

    Contin- ual test-time adaptation for end-to-end speech recognition on noisy speech.arXiv preprint arXiv:2406.11064, 2024

    Guan-Ting Lin, Wei-Ping Huang, and Hung-yi Lee. Contin- ual test-time adaptation for end-to-end speech recognition on noisy speech.arXiv preprint arXiv:2406.11064, 2024

    2024 arXiv

  17. [25]

    Listen, adapt, better wer: Source-free single-utterance test-time adaptation for automatic speech recognition.arXiv preprint arXiv:2203.14222, 2022

    Guan-Ting Lin, Shang-Wen Li, and Hung-yi Lee. Listen, adapt, better wer: Source-free single-utterance test-time adaptation for automatic speech recognition.arXiv preprint arXiv:2203.14222, 2022

    2022 arXiv

  18. [26]

    Unified deep supervised domain adapta- tion and generalization

    Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gi- anfranco Doretto. Unified deep supervised domain adapta- tion and generalization. InProceedings of the IEEE inter- national conference on computer vision, pages 5715–5725, 2017

    2017

  19. [27]

    Source free domain adaptation using an off-the-shelf clas- sifier

    Arun Reddy Nelakurthi, Ross Maciejewski, and Jingrui He. Source free domain adaptation using an off-the-shelf clas- sifier. In2018 IEEE International conference on big data (Big Data), pages 140–145. IEEE, 2018

    2018

  20. [28]

    Fast adaptation to super-resolution net- works via meta-learning

    Seobin Park, Jinsu Yoo, Donghyeon Cho, Jiwon Kim, and Tae Hyun Kim. Fast adaptation to super-resolution net- works via meta-learning. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pages 754–769. Springer, 2020

    2020

  21. [29]

    Semi-supervised domain adaptation via minimax entropy

    Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Dar- rell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. InProceedings of the IEEE/CVF in- ternational conference on computer vision, pages 8050–8058, 2019

    2019

  22. [30]

    Test- time prompt tuning for zero-shot generalization in vision- language models.Advances in Neural Information Process- ing Systems, 35:14274–14289, 2022

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test- time prompt tuning for zero-shot generalization in vision- language models.Advances in Neural Information Process- ing Systems, 35:14274–14289, 2022

    2022

  23. [31]

    Test-time training with self- supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–

  24. [32]

    A survey on deep trans- fer learning

    Chuanqi Tan, Fuchun Sun, Tao Kong, Wenchang Zhang, Chao Yang, and Chunfang Liu. A survey on deep trans- fer learning. InArtificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, P...

    2018

  25. [33]

    Mel spectrogram-based audio forgery detection using cnn.Sig- nal, Image and Video Processing, 17(5):2211–2219, 2023

    Arda Ustubioglu, Beste Ustubioglu, and Guzin Ulutas. Mel spectrogram-based audio forgery detection using cnn.Sig- nal, Image and Video Processing, 17(5):2211–2219, 2023

    2023

  26. [34]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 5018–5027, 2017

    2017

  27. [35]

    Tent: Fully test-time adapta- tion by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adapta- tion by entropy minimization. InInternational Conference on Learning Representations, 2021

    2021

  28. [36]

    Test-time training on video streams.arXiv preprint arXiv:2307.05014, 2023

    Renhao Wang, Yu Sun, Yossi Gandelsman, Xinlei Chen, Alexei A Efros, and Xiaolong Wang. Test-time training on video streams.arXiv preprint arXiv:2307.05014, 2023

    2023 arXiv

  29. [37]

    Efficient test time adapter ensem- bling for low-resource language varieties.arXiv preprint arXiv:2109.04877, 2021

    Xinyi Wang, Yulia Tsvetkov, Sebastian Ruder, and Graham Neubig. Efficient test time adapter ensem- bling for low-resource language varieties.arXiv preprint arXiv:2109.04877, 2021

    2021 arXiv

  30. [38]

    Speech commands: A dataset for limited-vocabulary speech recognition.arXiv preprint arXiv:1804.03209, 2018

    Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition.arXiv preprint arXiv:1804.03209, 2018

    2018 arXiv

  31. [39]

    Group normalization, 2018

    Yuxin Wu and Kaiming He. Group normalization, 2018

    2018

  32. [40]

    Semi-supervised domain adaptation with subspace learning for visual recognition

    Ting Yao, Yingwei Pan, Chong-Wah Ngo, Houqiang Li, and Tao Mei. Semi-supervised domain adaptation with subspace learning for visual recognition. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2142–2150, 2015

    2015

  33. [41]

    Alleviating style sensitivity then adapting: Source-free domain adaptation for medical image segmenta- tion

    Yalan Ye, Ziqi Liu, Yangwuyong Zhang, Jingjing Li, and Hengtao Shen. Alleviating style sensitivity then adapting: Source-free domain adaptation for medical image segmenta- tion. InProceedings of the 30th ACM International Confer- ence on Multimedia, pages 1935–1944, 2022

    1935

  34. [42]

    An unsupervised domain adaptation method for compressed video quality enhancement

    Wang Zeyang. An unsupervised domain adaptation method for compressed video quality enhancement. In2022 19th In- ternational Computer Conference on Wavelet Active Media Technology and Information Processing (ICCWAMTIP), pages 1–5. IEEE, 2022

    2022

  35. [43]

    Source-style transferred mean teacher for source-data free object detection

    Dan Zhang, Mao Ye, Lin Xiong, Shuaifeng Li, and Xue Li. Source-style transferred mean teacher for source-data free object detection. InProceedings of the 3rd ACM Interna- tional Conference on Multimedia in Asia, pages 1–8, 2021

    2021

  36. [44]

    Transfer adaptation learning: A decade survey.IEEE Transactions on Neural Networks and Learning Systems, 2022

    Lei Zhang and Xinbo Gao. Transfer adaptation learning: A decade survey.IEEE Transactions on Neural Networks and Learning Systems, 2022. A Ablation Study for CoNMix A.1 Ablation Analysis for SC As shown in Fig.6a, 6b, 6c, 6i, 6h, and 6g, the original CoN- Mix pseudo-label loss ...

    2022

Pith tools

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