Pith. sign in

REVIEW 5 major objections 7 minor 40 references

Continual Speech Learning with Fused Speech Features

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A learnable Gated-Fusion Layer over Whisper's frozen encoder lets a single speech model learn six tasks in sequence—keyword spotting, speaker identification, emotion recognition, intent classification, slot filling, and ASR—while…

desk verdict A useful empirical proof-of-concept for continual speech learning with a frozen Whisper encoder, but the headline result rests on oracle task identity and single-run numbers, so it needs variance estimates and honest scoping before publication. read the letter →

arxiv 2506.01496 v2 pith:WF4QS5IL submitted 2025-06-02 cs.CL cs.SDeess.AS

Guitao Wang , Jinming Zhao , Hao Yang , Guilin Qi , Tongtong Wu , Gholamreza Haffari This is my paper
classification cs.CLcs.SDeess.AS
keywords continuallearningspeechprocessinggatedfusionWhispertask-incrementalcatastrophicforgettinggenerativedecodingspokenlanguageunderstanding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes continual speech learning, a setting in which one speech model is trained sequentially on six distinct tasks—keyword spotting, speaker identification, emotion recognition, intent classification, slot filling, and automatic speech recognition. The authors argue that a learnable Gated-Fusion Layer placed over Whisper's frozen encoder lets the model dynamically combine features from different encoder layers for the current task, avoiding the overwriting that naive fine-tuning causes. They claim this design beats standard continual learning baselines such as experience replay and DERPP in both accuracy and average forgetting, with the best mean rank across the six tasks. If correct, the result is a practical recipe for extending speech models to new tasks without full retraining.

What carries the argument

The central object is the Gated-Fusion Layer (GFL). It introduces a parameter matrix $W \in \mathbb{R}^{N \times M}$, where $N$ is the number of tasks and $M$ the number of encoder layers; for task $k$, the one-hot vector $t_k$ selects a temperature-controlled softmax over layers, and the final representation is $h_k = t_k \cdot \mathrm{SoftMax}(W, t) \cdot H_{\mathrm{norm}}$, with $H_{\mathrm{norm}}$ the concatenation of LayerNorm'd hidden states from every encoder layer. The gate's job is to act as a differentiable router that lets the decoder attend to content, speaker, and prosodic features stored in different layers depending on the task, while the frozen encoder guarantees that no parameter drift occurs.

What would settle it

Run GFLD on the six tasks while withholding the task identity at inference so the model must infer the task from the audio itself; if accuracy and forgetting are unchanged, task-conditioned gating is not the cause of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that task-conditioned fusion of multi-layer encoder features is what makes continual learning work for heterogeneous speech tasks. A Gated-Fusion Layer takes the normalized hidden states of all Whisper encoder layers and, given a one-hot task vector, computes a task-specific weighted average of those layers using a softmax with a low temperature. Because the encoder stays frozen, previously learned features are not overwritten, and the decoder only needs to adapt its generative behavior. The double-stage variant GFLD first trains the gate jointly with the decoder, then freezes the gate and trains only the decoder, which the authors find yields the most stable performance and the best mean rank (2.50) across the six tasks.

Load-bearing premise

The load-bearing premise is that the model is always told which task it is performing at inference through a one-hot task vector, and that Whisper's frozen encoder already contains the features every task needs; if either fails, the gating mechanism cannot be expected to work.

Editorial extensions

If this is right

  • A single Whisper model can be extended to new speech tasks without full retraining, as long as the new task's audio features are representable by the frozen encoder.
  • GFLD retains stable performance across different task orders, so deployment does not require a fixed curriculum.
  • Task-specific output prompts plus constrained decoding yield large gains on classification-style tasks, making generative decoding a viable interface for discriminative speech tasks.
  • Pre-training the gate on only 10% of the data preserves nearly 95% of full-data performance, so the approach is data-efficient.
  • Negative forgetting scores on some tasks indicate that later tasks can improve earlier ones through positive transfer, not merely avoid interference.

Reading between the lines

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

  • Because the task identity is injected as a one-hot vector, the paper's evaluation is task-incremental; whether the same gate could infer the task from the audio alone is an open question the paper does not address.
  • The method's ceiling is set by Whisper's frozen encoder: any task whose decisive features are absent from Whisper's pretrained representations (for example, an unseen language or an acoustic attribute not covered by its training) would likely fall outside what the gate can select.
  • The same gating recipe—selecting from frozen layer-wise features under an oracle task identity—could transfer to other frozen foundation models, not just speech encoders, as a generic continual-learning module.
  • The reported negative average forgetting suggests that positive backward transfer is real; a targeted study could test whether this comes from the frozen anchor, the gate, or the generative decoder re-adaptation.
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

5 major / 7 minor

Summary. The paper proposes a continual speech learning setup (CSL) in which Whisper-base is sequentially adapted to six speech tasks (KS, SID, ER, IC, SF, ASR) by freezing the encoder and adding a Gated-Fusion Layer (GFL) that combines normalized hidden states from all encoder layers using a task one-hot vector. Two training strategies are introduced: GFLS jointly trains the GFL and decoder, while GFLD initializes a new model with the GFL parameters learned by GFLS, freezes the GFL, and trains only the decoder. Experiments compare GFLS and GFLD against MTL, FT, Replay, LwF, and DERPP using accuracy, WER, STF1, average forgetting, and mean rank. The paper claims that GFLD achieves the best mean rank and maintains stable performance under different task orders.

Significance. If the claims were fully supported, the paper would be a useful contribution to speech continual learning: it unifies six heterogeneous tasks under a generative decoding format, uses a parameter-efficient frozen-encoder approach, and provides a public code link. The comparison with standard continual-learning baselines and the task-order analysis are valuable. The evaluation uses held-out test sets, and the design is motivated by prior layer analyses rather than fitted to the results, so I do not see a circularity problem. However, the central empirical claim currently rests on the GFLD two-stage protocol, the oracle task-identity assumption, and single-run results, so the significance is conditional on resolving those issues.

major comments (5)
  1. [§3.2 (GFLD definition)] GFLD as described in Section 3.2 is initialized with the GFL parameters trained during the GFLS run, which means the fusion weights have already been learned over the full task sequence before the decoder-only continual phase begins. This gives GFLD access to information about future tasks that the baselines (FT, Replay, LwF, DERPP) and GFLS do not receive, so the comparison in Table 1 is not a like-for-like continual-learning comparison. Please clarify what data the GFLS pre-training stage uses, and either match the baselines by giving them an equivalent pre-training stage or remove this pre-training from the evaluated protocol. If GFLD is intended as a two-pass post-hoc consolidation method, the claims should be scoped accordingly.
  2. [Eq. (3) and §5.5] The GFL computation in Eq. (3) requires the one-hot task vector t_k at inference, and the paper only evaluates the task-incremental setting where the task label is given. The claim in Section 5.5 that GFLD is suitable for 'open-world environments with unpredictable task scheduling' is not tested, because in such an environment the task would have to be inferred from the audio. Please either add a task-agnostic evaluation (e.g., inferring the task via the decoder or a classifier) or explicitly scope all adaptability claims to the task-incremental setting with known task labels throughout the abstract, introduction, and conclusion.
  3. [Table 1 and §4.3] The headline comparison in Table 1 is based on a single run per method. The mean-rank gap between GFLD (2.50) and Replay (2.83) is small, and per-task results go in opposite directions (Replay is better on KS and SF; GFLD is better on SID and ER). Without multiple seeds, error bars, or significance tests, the abstract's claim that the approach 'improves accuracy significantly' is not supported. Please report standard deviations over at least three to five seeds and apply a significance test (e.g., paired bootstrap or Wilcoxon) to the per-task metrics and mean ranks.
  4. [§4.4 and Table 1 vs Table 3] The exact task order used for the six-task results in Table 1 is never stated, although Section 5.5 shows that performance is order-dependent for a three-task subset (STDEV up to 8.82 for GFLS on SID). Since continual-learning results are sensitive to order, the main numbers are not reproducible without this information, and the claim of order-robustness is only demonstrated for KS/SID/ER, not for the full six-task set. Please report the task order for Table 1 and, ideally, repeat the main comparison under at least two or three orders.
  5. [§4.4 and Table 2] The prompt and constrained-decoding configuration applied to each task and each method in Table 1 is not specified. Table 2 shows that these choices have large effects: KS accuracy is 97.11 with no prompt and 89.78 with the 'xxx is ...' prompt plus constrained decoding, while IC accuracy varies from 75.67 to 99.31. Because Section 5.3 states that prompts must be 'carefully tailored' per task, the main evaluation should either use the best configuration per task for all methods or report the exact prompt/decoding setup for every cell; otherwise the comparisons in Table 1 are difficult to reproduce and the fairness of the baselines is unclear.
minor comments (7)
  1. [Abstract and Title] The abstract uses 'continuous speech learning' while the title and the rest of the paper use 'continual speech learning'; please choose one term and use it consistently.
  2. [Section 2] There are typographical issues: 'Given a sequence of task{T1,T2,...}' should be 'tasks', and the definition of eT^test_k is garbled; please rewrite for clarity.
  3. [Eq. (3)] The dimensions of H_norm and the axis along which SoftMax is applied are not defined; please state that H_norm is M x d and that SoftMax operates over the M layers.
  4. [Section 5.2] The 'few-shot pre-training' experiment is not defined clearly; please clarify what 10% of the data means, how the subset is sampled, and what the axes of Figure 2 represent.
  5. [Table 3 and Section 5.5] There are typos: 'GSLS' and 'GSLD' should be 'GFLS' and 'GFLD', and 'GSL D' should be 'GFLD'.
  6. [Section 5.4] The paper uses both 'constraint decoding' and 'constrained decoding'; please use one term consistently.
  7. [Table 1] For MTL, the AF entries are '-'; please add a footnote explaining that AF is not defined when there is no sequential training.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the empirical claims are supported by held-out evaluation, and the GFL weights are learned rather than fitted to the reported predictions.

full rationale

The paper's central claim is empirical: a learnable Gated-Fusion Layer over Whisper's frozen encoder, task tags, and generative decoding yields better accuracy and less forgetting than continual-learning baselines. Nothing in the derivation reduces to its inputs by construction. Equation (3) computes h_k = t_k · SoftMax(W, t) · H_norm, where W is a learned parameter matrix and t_k is a one-hot task indicator; the task indicator is an oracle input that scopes the evaluation to task-incremental continual learning, but it does not make the reported test accuracies equal to the training signal. Table 1 evaluates on held-out SUPERB test sets, Table 3 varies task order, and Figure 2 varies pre-training scale, so the results are self-contained against external benchmarks. The self-citations — especially [17] for layer-wise specialization of audio encoders and [18] for continual-learning challenges — motivate the design but do not determine the empirical outcome; the cited results are externally falsifiable empirical studies rather than uniqueness theorems or ansatz carriers. No fitted parameter is renamed as a prediction, and no known result is repackaged as a derivation. The task-agnostic limitation is a scoping concern for the 'open-world' generalization claim, not a circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests mainly on the empirical setup. The only invented component is the GFL module, which is learned rather than postulated as an entity. Key hand-set choices are the softmax temperature, per-task epochs, replay buffer size, and per-task prompt templates. The main assumptions are that Whisper's layers encode complementary features, that task identity is known at inference, and that the frozen encoder is sufficient for all tasks.

free parameters (4)
  • Softmax temperature t = 0.0005
    Hand-set temperature in Eq. (3) controls the sharpness of the per-task gating; reported in Section 4.4 with no sensitivity analysis.
  • Max training epochs per task = 40 (SID), 60 (ER), 20 (others)
    Chosen because of limited data (Section 4.4); affects final performance.
  • Experience replay buffer size = 1000
    Set for Replay and DERPP baselines; affects forgetting and accuracy.
  • Prompt template and constrained decoding per task = e.g., 'xxx is ...' for IC/ER, None for KS
    Selected on the basis of Table 2; materially changes accuracy (e.g., IC from 75.67 to 99.31).
assumptions (3)
  • domain assumption Different layers of Whisper's encoder encode distinct speech features (content, speaker, prosody), and fusing them helps diverse tasks.
    Invoked in Section 3.2 based on prior work [17]; if false, the gated fusion premise is weakened.
  • domain assumption Task identity is known at inference and is provided via task-specific tags and the one-hot vector t_k.
    Eq. (3) requires t_k to select the correct gate row; this assumes a task-incremental continual learning setting.
  • domain assumption Whisper's frozen encoder features are sufficient for all six tasks when combined with a trainable decoder and GFL.
    Section 4.4 freezes the encoder; if the pretrained features do not cover a task, the method cannot adapt them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continual Speech Learning with Fused Speech Features." pith.science (2026). https://pith.science/paper/WF4QS5IL

@misc{pith2026250601496,
  author       = {Pith},
  title        = {Pith review of: Continual Speech Learning with Fused Speech Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WF4QS5IL}},
  note         = {Machine review of arXiv:2506.01496}
}
read the original abstract

Rapid growth in speech data demands adaptive models, as traditional static methods fail to keep pace with dynamic and diverse speech information. We introduce continuous speech learning, a new set-up targeting at bridging the adaptation gap in current speech models. We use the encoder-decoder Whisper model to standardize speech tasks into a generative format. We integrate a learnable gated-fusion layer on the top of the encoder to dynamically select task-specific features for downstream tasks. Our approach improves accuracy significantly over traditional methods in six speech processing tasks, demonstrating gains in adapting to new speech tasks without full retraining.

Figures

Figures reproduced from arXiv: 2506.01496 by the authors.

Figure 1
Figure 1. Framework of Continual Speech Learning with Fused Speech Features. For our continual learning setting, we make the following modifications: (1) Task-specific tags. To recognize and decode the current task, we introduce additional task-specific tags into Whisper’s vocabulary. For example, we use <|KS|> for the Keyword Spotting task. (2) Special tokens. We also introduce special tokens which are needed to label speake… view at source ↗
Figure 2
Figure 2. Performance of GFLD initialized by weights pre￾trained from different scale of data. 5.2. Analysis of Few-shot Pre-training As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance progression of continual learning meth￾ods across six tasks. Each curve represents the accuracy (for classification tasks), slot type F1 score (STF1, for SF), or word error rate (WER, for ASR) at different stages of training. enabling reliable deployment in open-world environments with unpredictable task scheduling. 5.6. Progression of Continual Learning Process [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

  1. [1]

    While CL has been widely studied in natural language processing (NLP) [4] and computer vision (CV) [5], its application to speech processing remains largely unexplored

    Introduction Continual learning (CL) [1–3] has emerged as a promising solu- tion to enable models to adapt to new data streams while retain- ing previously learned knowledge. While CL has been widely studied in natural language processing (NLP) [4] and computer vision (CV) [5], its application to speech processing remains largely unexplored. Traditional s...

    work page

  2. [2]

    Given a sequence of task{T 1,T 2,

    Continual Speech Learning Setup Continual learning sequentially trains a model on a series of tasks. Given a sequence of task{T 1,T 2, . . . ,TN }, whereNis the number of tasks. Each task has its own dataset, including training, validation, and test sets. At thek th step, the model is trained onT k to learn new knowledge and will be evaluated on eT test k...

    work page

  3. [3]

    Continual Speech Learning with Fused Speech Features

    Methodology 3.1. Model Architecture and Modification We adopt WHISPER-BASE 1 with encoder-decoder architec- ture as our backbone model, which is a multitask model primarily pre-trained on transcribing and translating tasks. Whisper executes tasks by using the corresponding prefixes (e.g.,“transcribe” and “translate”) as prompts to guide the model. 1https:...

    work page Pith review arXiv 2025

  4. [4]

    happy”, “sad

    Experiment 4.1. Dataset We conduct our experiments on six tasks with the corresponding datasets from SUPERB [20]. These tasks can be categorized into discriminative (e.g., IC) and generative tasks (e.g., ASR). To align the decoding strategies for these two types of tasks, we decode discriminative tasks in the same way of generative tasks. These tasks are:...

    work page

  5. [5]

    happy” directly, the model is required to generate a sentence like “Emotion is happy

    Results and Analysis 5.1. Analysis of Gated-Fusion Layer We compare results of various continual learning methods af- ter training on the last task in Table 1. Our proposed GFL framework demonstrates strong continual learning capabilities by dynamically selecting task-specific features while mitigat- ing catastrophic forgetting. Compared to baseline metho...

    work page

  6. [6]

    Our method significantly improves performance across multiple tasks, particularly in knowledge retention and mitigating catastrophic forgetting

    Conclusion We introduced a novel framework for speech continual learn- ing by integrating a Gated-Fusion Layer (GFL) into Whis- per, enabling dynamic fusion of speech features for improved adaptability. Our method significantly improves performance across multiple tasks, particularly in knowledge retention and mitigating catastrophic forgetting. Our resul...

    work page

  7. [7]

    U21A20488

    Acknowledgements This work is partially supported by National Nature Science Foundation of China under No. U21A20488. We thank the Big Data Computing Center of Southeast University for providing the facility support on the numerical calculations in this paper. The initial draft of this paper was written manually. Chat- GPT was later used to enhance clarit...

    work page

  8. [8]

    Contin- ual learning for large language models: A survey,

    T. Wu, L. Luo, Y . Li, S. Pan, T. Vu, and G. Haffari, “Contin- ual learning for large language models: A survey,”CoRR, vol. abs/2402.01364, 2024

    arXiv 2024

Show all 40 references
  1. [9]

    Dealing with unknowns in continual learning for end-to-end automatic speech recognition,

    M. Sustek, S. Sadhu, and H. Hermansky, “Dealing with unknowns in continual learning for end-to-end automatic speech recognition,” in23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, H. Ko and ...

    2022 doi

  2. [10]

    A comprehensive survey of continual learning: Theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 8, pp. 5362–5383,

  3. [11]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Des- jardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska- Barwinskaet al., “Overcoming catastrophic forgetting in neural networks,”Proceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  4. [12]

    Continual learning of natural language pro- cessing tasks: A survey,

    Z. Ke and B. Liu, “Continual learning of natural language pro- cessing tasks: A survey,”arXiv preprint arXiv:2211.12701, 2022

    2022 arXiv

  5. [13]

    Recent ad- vances of continual learning in computer vision: An overview,

    H. Qu, H. Rahmani, L. Xu, B. Williams, and J. Liu, “Recent ad- vances of continual learning in computer vision: An overview,” arXiv preprint arXiv:2109.11369, 2021

    2021 arXiv

  6. [14]

    Continual learning with pre-trained models: A survey,

    D. Zhou, H. Sun, J. Ning, H. Ye, and D. Zhan, “Continual learning with pre-trained models: A survey,”CoRR, vol. abs/2401.16386, 2024

    2024 arXiv

  7. [15]

    A comprehensive survey of continual learning: Theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”CoRR, vol. abs/2302.00487, 2023

    2023 arXiv

  8. [16]

    Experience replay for continual learning,

    D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, and G. Wayne, “Experience replay for continual learning,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  9. [17]

    Gradient episodic memory for continual learning,

    D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  10. [18]

    Efficient lifelong learning with A-GEM,

    A. Chaudhry, M. Ranzato, M. Rohrbach, and M. Elhoseiny, “Efficient lifelong learning with A-GEM,” in7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. [On- line]. Available: https://openreview.net/for...

    2019

  11. [19]

    A comprehensive sur- vey of continual learning: theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive sur- vey of continual learning: theory, method and application,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  12. [20]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,”IEEE transac- tions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935–2947, 2017

    2017

  13. [21]

    Piggyback: Adapting a single network to multiple tasks by learning to mask weights,

    A. Mallya, D. Davis, and S. Lazebnik, “Piggyback: Adapting a single network to multiple tasks by learning to mask weights,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 67–82

    2018

  14. [22]

    Overcoming catastrophic forgetting with hard attention to the task,

    J. Serra, D. Suris, M. Miron, and A. Karatzoglou, “Overcoming catastrophic forgetting with hard attention to the task,” inInterna- tional conference on machine learning. PMLR, 2018, pp. 4548– 4557

    2018

  15. [23]

    Learning to prompt for continual learning,

    Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2022, pp. 139–149

    2022

  16. [24]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  17. [25]

    Investigating pre-trained audio encoders in the low-resource condition,

    H. Yang, J. Zhao, G. Haffari, and E. Shareghi, “Investigating pre-trained audio encoders in the low-resource condition,”arXiv preprint arXiv:2305.17733, 2023

    2023 arXiv

  18. [26]

    Pretrained language model in continual learning: A comparative study,

    T. Wu, M. Caccia, Z. Li, Y . Li, G. Qi, and G. Haffari, “Pretrained language model in continual learning: A comparative study,” inThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29,

    2022

  19. [27]

    Semi- supervised spoken language understanding via self-supervised speech and language model pretraining,

    C.-I. Lai, Y .-S. Chuang, H.-Y . Lee, S.-W. Li, and J. Glass, “Semi- supervised spoken language understanding via self-supervised speech and language model pretraining,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,...

    2021

  20. [28]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  21. [29]

    Superb: Speech processing universal performance benchmark,

    S.-w. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakho- tia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Linet al., “Superb: Speech processing universal performance benchmark,” arXiv preprint arXiv:2105.01051, 2021

    2021 arXiv

  22. [30]

    Towards relation extraction from speech,

    T. Wu, G. Wang, J. Zhao, Z. Liu, G. Qi, Y . Li, and G. Haffari, “Towards relation extraction from speech,” inProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, 2022, pp. 10 7...

    2022 doi

  23. [31]

    Towards event extraction from speech with contextual clues,

    J. Kang, T. Wu, J. Zhao, G. Wang, G. Qi, Y . Li, and G. Haffari, “Towards event extraction from speech with contextual clues,”CoRR, vol. abs/2401.15385, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2401.15385

    2024 doi arXiv

  24. [32]

    Speech commands: A public dataset for single-word speech recognition,

    P. Warden, “Speech commands: A public dataset for single-word speech recognition,”Dataset available from http://download. ten- sorflow. org/data/speech commands v0, vol. 1, 2017

    2017

  25. [33]

    V oxceleb: Large-scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxceleb: Large-scale speaker verification in the wild,”Computer Speech & Language, vol. 60, p. 101027, 2020

    2020

  26. [34]

    Iemocap: Interactive emotional dyadic motion capture database,

    C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, “Iemocap: Interactive emotional dyadic motion capture database,”Language resources and evaluation, vol. 42, pp. 335–359, 2008

    2008

  27. [35]

    Speech model pre-training for end-to-end spoken language understanding,

    L. Lugosch, M. Ravanelli, P. Ignoto, V . S. Tomar, and Y . Ben- gio, “Speech model pre-training for end-to-end spoken language understanding,”arXiv preprint arXiv:1904.03670, 2019

    1904 arXiv

  28. [38]

    Dark experience for general continual learning: a strong, sim- ple baseline,

    P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara, “Dark experience for general continual learning: a strong, sim- ple baseline,”Advances in neural information processing systems, vol. 33, pp. 15 920–15 930, 2020

    2020

  29. [39]

    Rie- mannian walk for incremental learning: Understanding forgetting and intransigence,

    A. Chaudhry, P. K. Dokania, T. Ajanthan, and P. H. Torr, “Rie- mannian walk for incremental learning: Understanding forgetting and intransigence,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 532–547

    2018

  30. [40]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,”arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  31. [2022]

    [Online]

    OpenReview.net, 2022. [Online]. Available: https: //openreview.net/forum?id=figzpGMrdD

    2022

  32. [2024]

    Available: https://doi.org/10.1109/TPAMI.2024

    [Online]. Available: https://doi.org/10.1109/TPAMI.2024. 3367329

    2024 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.