Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Fine-tune Before Structured Pruning: Towards Compact and Accurate Self-Supervised Models for Speaker Diarization

T0 review · 2 major / 5 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 Fine-tuning a WavLM encoder for speaker diarization before L0-structured pruning with distillation allows 80% of its parameters to be removed with no loss in diarization error.

desk verdict A useful, well-executed pruning recipe for diarization whose headline 'no degradation' claim needs variance estimates or softer wording. read the letter →

arxiv 2505.24111 v1 pith:XIPCVVYQ submitted 2025-05-30 eess.AS

Jiangyu Han , Federico Landini , Johan Rohdin , Anna Silnova , Mireia Diez , Jan Cernocky , Lukas Burget This is my paper
classification eess.AS
keywords speakerdiarizationstructuredpruningself-supervisedspeechmodelsWavLMknowledgedistillationL0regularizationmodelcompression
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 aims to show that a large self-supervised speech encoder used for speaker diarization can be shrunk to one fifth of its size without losing diarization accuracy, if it is first fine-tuned to the diarization task and then pruned with a distillation objective. On far-field single-channel AMI, AISHELL-4, and AliMeeting, removing 80% of WavLM parameters leaves macro DER at 15.2 vs 15.0 for Base+ and 14.1 vs 14.1 for Large, while single-GPU inference becomes 4.0x and 2.6x faster. The paper's distinctive claim is that the order matters: pruning a task-agnostic WavLM produces a student that stays similar to its teacher but cannot be fine-tuned back to competitive diarization performance, whereas pruning after task fine-tuning preserves accuracy. This matters because diarization systems must process long meetings in resource-constrained settings, and structured pruning gives actual speedups without specialized sparse-matrix hardware.

What carries the argument

The machinery is an $\ell_0$-regularized structured pruning objective in which each prunable group (a CNN kernel, an attention head, or an FFN intermediate dimension) is multiplied by a Hard Concrete random variable; the expected $\ell_0$ penalty is analytic and gradients come from reparameterization. An augmented Lagrangian term with learnable multipliers enforces a preset target sparsity, which is linearly warmed up over the first five epochs. Concurrently, a layer-wise distillation loss matches the student's hidden states at layers {0,4,8,12} (Base+) or {0,8,16,24} (Large) to a fixed teacher through learnable linear projections, using L1 plus negative cosine distance. The fine-tuned WavLM initializes both teacher and student, and after pruning the student is distilled toward the teacher for up to 20 epochs before the final diarization fine-tuning.

What would settle it

Run the full recipe at 80% sparsity five times with different random seeds on the same data. If any run's macro DER exceeds the unpruned baseline by more than 0.5 DER, the 'no performance degradation' claim is false; if all runs stay within roughly 0.2 DER, the claim is supported. A second check is to evaluate the 80%-pruned model on a far-field meeting corpus not used in training, where a large DER jump would reveal overfitting to the three datasets.

Watch

Extended reading notes

Core claim

The central discovery is that a WavLM encoder fine-tuned for speaker diarization contains a structured subnetwork that matches the full model's diarization error at 80% sparsity. The paper reports macro DER of 15.2% at 80% sparsity for WavLM Base+ against 15.0% unpruned, and 14.1% for WavLM Large against 14.1% unpruned, averaged over AMI, AISHELL-4, and AliMeeting; the pruned Large model has fewer parameters (63.3M) than the unpruned Base+ (94.4M) while achieving better DER. The claim is not that arbitrary pruning preserves performance, but that this specific recipe does: fine-tune, then prune with an L0-regularized distillation objective to a fixed target sparsity, then fine-tune again. The paper also establishes that this pre-fine-tuning step is load-bearing, since omitting it raises macro DER from 15.6 to 16.0 for the Base+ at the same sparsity.

Load-bearing premise

The strongest load-bearing premise is that the 80%-sparsity results are stable across training runs: the reported pruned-versus-unpruned gaps of 0.2 DER or less are treated as noise, but each configuration was run once with no seed variation.

Editorial extensions

If this is right

  • At 80% sparsity, both WavLM Base+ and Large keep macro DER within 0.2 DER of their unpruned versions, so the pruned models can replace the originals in the tested far-field meeting conditions.
  • Single-GPU inference speeds up 4.0x (Base+) and 2.6x (Large) with no change to the inference pipeline, because structured pruning removes whole kernels, heads, and FFN dimensions rather than individual weights.
  • Pruned WavLM Large at 80% sparsity holds fewer parameters than unpruned Base+ yet scores better macro DER, indicating that larger SSL encoders are the better starting point until sparsity becomes extreme.
  • At 94% sparsity, WavLM Large drops to roughly the parameter count of 80%-pruned Base+ but performs clearly worse, so there is a per-model sweet spot beyond which pruning destroys accuracy.
  • Omitting the pre-pruning fine-tuning step costs 0.4 DER on the Base+ macro average, confirming that task fine-tuning before pruning is part of what makes high sparsity viable.

Reading between the lines

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

  • Editorial extension: the paper does not vary training seeds, so the strongest test of the recipe is seed stability; if the 80% parity reproduces across seeds, the same fine-tune-prune-distill order should transfer to other SSL encoders and other tasks where a pretrained encoder is task-tuned.
  • Editorial extension: the learned mask pattern (middle CNN kernels kept, Transformer layers 9-10 removed, some attention heads never pruned) suggests the pruning process doubles as a data-driven relevance map of WavLM; a natural next experiment is to check whether the same mask transfers across corpora and acoustic conditions.
  • Editorial extension: the 94%-pruned-Large versus 80%-pruned-Base+ comparison implies that model selection for a fixed parameter budget should be a joint search over starting model size and sparsity rather than always pruning the largest available model.
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

2 major / 5 minor

Summary. The paper proposes a three-stage compression pipeline for WavLM-based speaker diarization systems: first fine-tune the SSL model on the diarization task, then perform structured pruning (CNN kernels, attention heads, FFN dimensions) with an L0 regularization and knowledge-distillation objective, and finally fine-tune the pruned model again. Experiments on far-field single-channel AMI, AISHELL-4, and AliMeeting report that WavLM Base+ and WavLM Large can be pruned to 80% sparsity while maintaining diarization performance, with 4.0x and 2.6x single-GPU inference speedups. The paper also studies sparsity levels up to 94%, ablates the importance of fine-tuning before pruning, and analyzes cosine similarity between student and teacher.

Significance. If the central claim is confirmed, the paper is practically valuable: it gives a concrete recipe for compressing a large SSL encoder for diarization without specialized sparse kernels, on multiple realistic far-field datasets, with released code. The fine-tune-before-prune insight is clearly demonstrated by the triangles in Figure 4, and the ablation in Table 2 is informative. The main weakness is evidential rather than conceptual: the claim of 'without any performance degradation' is stronger than the reported DER numbers support, and no variance estimates are provided. The teacher/student shared initialization is a design choice rather than a circular error, because the final claims are anchored to external diarization benchmarks.

major comments (2)
  1. [§3.3.1, Table 1] The abstract and conclusion claim that pruning can remove 80% of parameters 'without any performance degradation', but the reported results at 80% sparsity show nonzero degradation on every dataset for Base+ (macro DER 15.2 vs 15.0; AMI 15.7 vs 15.6; AISHELL-4 12.1 vs 11.8; AliMeeting 17.9 vs 17.7) and on AMI for Large (15.1 vs 14.8). The Large macro DER stays at 14.1 only because AliMeeting improves by 0.5 while AMI worsens by 0.3. No confidence intervals, seed repetitions, or paired significance tests are reported anywhere. This is load-bearing because the 'no degradation' wording appears in the abstract, Section 1, and Section 4; please either add multi-seed or paired-comparison evidence, or rephrase the claim to 'on par' / 'comparable'.
  2. [§3.3.3, Table 2] The full pipeline row ('+ further distill') that corresponds to the proposed method also shows a systematic loss relative to the unpruned reference: AISHELL-4 12.1 vs 11.8 and AliMeeting 17.9 vs 17.7, with macro DER 15.2 vs 15.0. Thus the no-degradation issue is not confined to a single exploratory table row; the final configuration itself exhibits small degradations. The manuscript should acknowledge these numbers and connect them to the variance question raised for Table 1.
minor comments (5)
  1. [§2.2, Eq. (2)] The relation between the distillation loss and the original diarization/powerset loss during the pruning phase is ambiguous; please state explicitly whether Eq. (2) is the only loss used during pruning and, if a weighted combination is used, give the weights.
  2. [§3.2] The sentence 'the target sparsity is linearly increased to the pre-defined values' should specify the schedule (e.g., per epoch, per step) and the exact endpoint values used in each experiment.
  3. [Table 1] The header 'collar=0s MacroSparsity #Params. Speedup' appears to be misformatted, making it hard to tell which columns are grouped; please fix the table layout.
  4. [§3.3.1] The speedup measurements use 'the input batch size optimized for maximum GPU utilization'; please report the actual batch sizes and state whether the same batch is used for both the unpruned and pruned models.
  5. [Figure 2] The caption says the averaged values are over 5 runs, but no such averaging is reported for the DER numbers in Table 1; please clarify why timing has repeated runs while the central DER results do not.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pruning results are empirical, benchmarked on external diarization datasets, and the teacher/student same-initialization is a design choice rather than a derivational loop.

full rationale

This paper is an empirical compression study with no claimed first-principles derivation. The central results (Table 1) are measured diarization error rates on AMI, AISHELL-4, and AliMeeting for the authors' fine-tuned unpruned and pruned WavLM models. The pruning objective is the standard L0-regularized distillation loss from prior work, and the teacher being the fine-tuned WavLM with the student initialized from the same weights is a self-referential design choice, but it is not used to derive a prediction; it is an experimental setup. No quantity is fitted and then renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is repackaged under new coordinates. The only self-citation is the DiariZen pipeline, which is code-released and externally evaluated, so it is not a load-bearing unverified claim. The concern that the 'without any performance degradation' claim is under-supported by Table 1 (e.g., Base+ macro DER moving from 15.0 to 15.2 at 80% sparsity) is about statistical evidence and missing error bars, not circularity. Therefore, no circular step is identified.

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

The central result rests on task fine-tuning before pruning, L0 plus distillation optimization, and the DiariZen evaluation pipeline. The hand-set numbers that can move the result are the target sparsity, the distilled layer set, and the distillation loss weights; all are empirically chosen. No new entities, forces, or dimensions are introduced.

free parameters (3)
  • target sparsity t = 0.8, 0.9, 0.94, plus intermediate values in Figure 3
    User-chosen target sparsity; the 80% no-loss claim is conditional on this target and on the linear schedule over the first 5 epochs.
  • distillation layer subset S = {0, 4, 8, 12} for Base+, {0, 8, 16, 24} for Large
    Hand-selected subset of WavLM layers used in Equation 2; no sensitivity analysis is provided.
  • distillation loss weights = 0.5 L1 + 0.5 cosine
    Equal weighting in Equation 2 follows prior work [6, 18]; no ablation over these weights is reported.
assumptions (3)
  • domain assumption WavLM representations can be transferred to diarization and remain structured-prunable after task fine-tuning.
    This is the central premise of the pipeline; it is supported only by the empirical results in Tables 1 and Table 2, not by a theoretical argument.
  • standard math The Hard Concrete L0 relaxation and augmented Lagrangian optimization find sparse subnetworks that can be fine-tuned back to full performance.
    Borrowed from [9, 10]; the paper does not measure the relaxation gap or verify that the optimized sparse structure is globally good.
  • domain assumption The DiariZen EEND plus pyannote clustering pipeline gives a valid diarization error rate, including its assumptions of at most 4 speakers and 2 overlapping speakers.
    Taken from [3]; if this pipeline misrepresents diarization quality, all DER comparisons inherit the error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-tune Before Structured Pruning: Towards Compact and Accurate Self-Supervised Models for Speaker Diarization." pith.science (2026). https://pith.science/paper/XIPCVVYQ

@misc{pith2026250524111,
  author       = {Pith},
  title        = {Pith review of: Fine-tune Before Structured Pruning: Towards Compact and Accurate Self-Supervised Models for Speaker Diarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XIPCVVYQ}},
  note         = {Machine review of arXiv:2505.24111}
}
read the original abstract

Self-supervised learning (SSL) models like WavLM can be effectively utilized when building speaker diarization systems but are often large and slow, limiting their use in resource constrained scenarios. Previous studies have explored compression techniques, but usually for the price of degraded performance at high pruning ratios. In this work, we propose to compress SSL models through structured pruning by introducing knowledge distillation. Different from the existing works, we emphasize the importance of fine-tuning SSL models before pruning. Experiments on far-field single-channel AMI, AISHELL-4, and AliMeeting datasets show that our method can remove redundant parameters of WavLM Base+ and WavLM Large by up to 80% without any performance degradation. After pruning, the inference speeds on a single GPU for the Base+ and Large models are 4.0 and 2.6 times faster, respectively. Our source code is publicly available.

Figures

Figures reproduced from arXiv: 2505.24111 by the authors.

Figure 1
Figure 1. (a) system for WavLM fine-tuning; (b) illustration of structured pruning. The light-colored rectangles will be pruned. 2.1. Speaker diarization pipeline The speaker diarization pipeline is used for WavLM fine-tuning. We utilized DiariZen [3], a framework with local end-to-end neural diarization (EEND) followed by speaker embedding clustering using pyannote [25], to build our diarization pipeline. The EEND module int… view at source ↗
Figure 2
Figure 2. Inference time comparisons of CNNs and Transformer layers. The averaged values out of 5 runs are reported. results in the same number of parameters (18.8M) as the pruned WavLM Base+ at 80% sparsity, performs significantly worse. Our results suggest that overly aggressive pruning can com￾pletely undermine a model’s performance. Thus, for scenar￾ios requiring a fast and compact model, pruning from a smaller model may … view at source ↗
Figure 3
Figure 3. Performance comparisons between WavLM Base+ and WavLM Large under different pruning setups [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Visualizations of CNN channels, attention heads, and FFN intermediate dimensions from the pruned WavLM Base+. The original sizes are 512, 12, and 3072, respectively. The pruning sparsity is set to {60%, 70%, 80%, 85%, 90%}. 5. As observed, CNN layers in the middle are …

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. The tttAI System for the TSA-ASR Task of the SmartGlasses Challenge 2026

    eess.AS 2026-07 conditional novelty 4.0 of 10

    A cascaded smart-glasses TSA-ASR system with a dominant-speaker overlap fallback achieved 7.10% tcpCER on two-person dialogues and 34.04% on multi-party meetings, ranking second on the meeting track.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Introduction Recently, several speaker diarization systems [1–3] incorporate WavLM [4], one of the state-of-the-art self-supervised learning (SSL) models, resulting in excellent performance. However, the pre-trained SSL models often have high computational and memory costs at inference time, as well as large storage require- ments, making them impractical...

    work page

  2. [2]

    Methods Our method consists of three steps: (1) building a diarization system on top of pre-trained WavLM and fine-tuning it for the diarization task; (2) pruning the fine-tuned model; and (3) fur- ther fine-tuning the pruned model using the same diarization pipeline. Figure 1 shows the framework of our method, where the prunable units are individual kern...

    arXiv 2025

  3. [3]

    Datasets We follow the data setups in [3] to use the far-field single- channel data from AMI [21, 22], AISHELL-4 [23], and Al- iMeeting [24], for system evaluation

    Experiments 3.1. Datasets We follow the data setups in [3] to use the far-field single- channel data from AMI [21, 22], AISHELL-4 [23], and Al- iMeeting [24], for system evaluation. Our model is trained us- ing a combination of the three training sets, with their corre- sponding development sets combined and used for validation. 3.2. Configurations For fi...

    work page

  4. [4]

    The experiments on multiple real far-field datasets prove that our method achieves strong performance

    Conclusion In this study, we propose compressing SSL models like WavLM through structured pruning, with an application to speaker di- arization. The experiments on multiple real far-field datasets prove that our method achieves strong performance. After prun- ing, our method can keep only 20% of the parameters while maintaining the same diarization perfor...

    work page

  5. [5]

    For Transformer layers, the 9th and 10th are almost completely removed after pruning

    As observed, CNN layers in the middle are more relevant. For Transformer layers, the 9th and 10th are almost completely removed after pruning. For most prunable units, higher sparsity generally results in more elements being eliminated proportion- ally. However, for a few units, such as the 1st and last two layers of MHAs, the structure remains unchanged ...

    work page

  6. [6]

    Linguistics, Artificial Intelligence and Language and Speech Technologies: from Research to Applications

    Acknowledgements The work was supported by Ministry of Education, Youth and Sports of the Czech Republic (MoE) through the OP JAK project "Linguistics, Artificial Intelligence and Language and Speech Technologies: from Research to Applications" (ID:CZ.02.01.01/00/23_020/0008518), by European Defence Fund project ARCHER, and Horizon 2020 Marie Sklodowska- ...

    work page 2020

  7. [7]

    NTT speaker diarization system for CHiME-7: multi-domain, multi- microphone end-to-end and vector clustering diarization,

    N. Tawara, M. Delcroix, A. Ando, and A. Ogawa, “NTT speaker diarization system for CHiME-7: multi-domain, multi- microphone end-to-end and vector clustering diarization,” inProc. ICASSP. IEEE, 2024, pp. 11 281–11 285

    work page 2024

  8. [8]

    Mamba-based segmentation model for speaker diariza- tion,

    A. Plaquet, N. Tawara, M. Delcroix, S. Horiguchi, A. Ando, and S. Araki, “Mamba-based segmentation model for speaker diariza- tion,” arXiv preprint arXiv:2410.06459, 2024

    arXiv 2024

Show all 34 references
  1. [9]

    Leveraging self-supervised learning for speaker diarization,

    J. Han, F. Landini, J. Rohdin, A. Silnova, M. Diez, and L. Burget, “Leveraging self-supervised learning for speaker diarization,” in Proc. ICASSP. IEEE, 2025, pp. 1–5

    2025

  2. [10]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  3. [11]

    A survey on deep neural net- work pruning: Taxonomy, comparison, analysis, and recommen- dations,

    H. Cheng, M. Zhang, and J. Q. Shi, “A survey on deep neural net- work pruning: Taxonomy, comparison, analysis, and recommen- dations,” IEEE Transactions on Pattern Analysis and Machine In- telligence, 2024

    2024

  4. [12]

    DistilHuBERT: Speech representation learning by layer-wise distillation of hidden-unit bert,

    H.-J. Chang, S.-w. Yang, and H.-y. Lee, “DistilHuBERT: Speech representation learning by layer-wise distillation of hidden-unit bert,” in Proc. ICASSP. IEEE, 2022, pp. 7087–7091

    2022

  5. [13]

    Deep ver- sus wide: An analysis of student architectures for task-agnostic knowledge distillation of self-supervised speech models,

    T. Ashihara, T. Moriya, K. Matsuura, and T. Tanaka, “Deep ver- sus wide: An analysis of student architectures for task-agnostic knowledge distillation of self-supervised speech models,” inProc. Interspeech, 2022, pp. 411–415

    2022

  6. [14]

    Distil-Whisper: Robust knowledge distillation via large-scale pseudo labelling,

    S. Gandhi, P. von Platen, and A. M. Rush, “Distil-Whisper: Robust knowledge distillation via large-scale pseudo labelling,” arXiv preprint arXiv:2311.00430, 2023

    2023 arXiv

  7. [15]

    Learning sparse neu- ral networks through L0 regularization,

    C. Louizos, M. Welling, and D. P. Kingma, “Learning sparse neu- ral networks through L0 regularization,” in Proc. ICLR, 2018

    2018

  8. [16]

    Structured pruning of large language models,

    Z. Wang, J. Wohlwend, and T. Lei, “Structured pruning of large language models,” in Proc. EMNLP, 2020, pp. 6151–6162

    2020

  9. [17]

    Structured pruning learns com- pact and accurate models,

    M. Xia, Z. Zhong, and D. Chen, “Structured pruning learns com- pact and accurate models,” in Proc. ACL, 2022, pp. 1513–1528

    2022

  10. [18]

    Sheared LLaMA: Ac- celerating language model pre-training via structured pruning,

    M. Xia, T. Gao, Z. Zeng, and D. Chen, “Sheared LLaMA: Ac- celerating language model pre-training via structured pruning,” in Proc. ICLR, 2024

    2024

  11. [19]

    Movement pruning: Adaptive sparsity by fine-tuning,

    V . Sanh, T. Wolf, and A. Rush, “Movement pruning: Adaptive sparsity by fine-tuning,” Advances in neural information process- ing systems, vol. 33, pp. 20 378–20 389, 2020

    2020

  12. [20]

    Sparse progres- sive distillation: Resolving overfitting under pretrain-and-finetune paradigm,

    S. Huang, D. Xu, I. E. Yen, Y . Wang, S.-E. Chang, B. Li, S. Chen, M. Xie, S. Rajasekaran, H. Liu et al., “Sparse progres- sive distillation: Resolving overfitting under pretrain-and-finetune paradigm,” in Proc. ACL, 2022, pp. 190–200

    2022

  13. [21]

    Block prun- ing for faster transformers,

    F. Lagunas, E. Charlaix, V . Sanh, and A. M. Rush, “Block prun- ing for faster transformers,” in Proc. EMNLP, 2021, pp. 10 619– 10 629

    2021

  14. [22]

    SUPERB: Speech processing universal performance benchmark,

    S.-w. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Linet al., “SUPERB: Speech processing universal performance benchmark,” in Proc. Interspeech, 2021, pp. 1194–1198

    2021

  15. [23]

    Task- agnostic structured pruning of speech representation models,

    H. Wang, S. Wang, W.-Q. Zhang, H. Suo, and Y . Wan, “Task- agnostic structured pruning of speech representation models,” in Proc. Interspeech, 2023, pp. 231–235

    2023

  16. [24]

    DPHuBERT: Joint distillation and pruning of self-supervised speech models,

    Y . Peng, Y . Sudo, S. Muhammad, and S. Watanabe, “DPHuBERT: Joint distillation and pruning of self-supervised speech models,” in Proc. Interspeech, 2023, pp. 62–66

    2023

  17. [25]

    Structured pruning of self-supervised pre-trained models for speech recog- nition and understanding,

    Y . Peng, K. Kim, F. Wu, P. Sridhar, and S. Watanabe, “Structured pruning of self-supervised pre-trained models for speech recog- nition and understanding,” in Proc. ICASSP. IEEE, 2023, pp. 1–5

    2023

  18. [26]

    Accurate and structured pruning for effi- cient automatic speech recognition,

    H. Jiang, L. L. Zhang, Y . Li, Y . Wu, S. Cao, T. Cao, Y . Yang, J. Li, M. Yang, and L. Qiu, “Accurate and structured pruning for effi- cient automatic speech recognition,” in Proc. Interspeech, 2023, pp. 4104–4108

    2023

  19. [27]

    The AMI meeting corpus: A pre-announcement,

    J. Carletta, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, V . Karaiskos, W. Kraaij, M. Kronenthalet al., “The AMI meeting corpus: A pre-announcement,” in Interna- tional workshop on machine learning for multimodal interaction. Springer, 2005, pp. 28–39

    2005

  20. [28]

    The AMI meet- ing corpus,

    W. Kraaij, T. Hain, M. Lincoln, and W. Post, “The AMI meet- ing corpus,” in Proc. International Conference on Methods and Techniques in Behavioral Research, 2005, pp. 1–4

    2005

  21. [29]

    AISHELL-4: An open source dataset for speech enhancement, separation, recognition and speaker diariza- tion in conference scenario,

    Y . Fu, L. Cheng, S. Lv, Y . Jv, Y . Kong, Z. Chen, Y . Hu, L. Xie, J. Wu, H. Bu et al., “AISHELL-4: An open source dataset for speech enhancement, separation, recognition and speaker diariza- tion in conference scenario,” in Proc. Interspeech , 2021, pp. 3665–3669

    2021

  22. [30]

    M2MeT: The ICASSP 2022 multi-channel multi-party meeting transcription challenge,

    F. Yu, S. Zhang, Y . Fu, L. Xie, S. Zheng, Z. Du, W. Huang, P. Guo, Z. Yan, B. Ma et al., “M2MeT: The ICASSP 2022 multi-channel multi-party meeting transcription challenge,” in Proc. ICASSP. IEEE, 2022, pp. 6167–6171

    2022

  23. [31]

    pyannote. audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,

    H. Bredin, “pyannote. audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,” in Proc. Interspeech, 2023, pp. 1983–1987

    2023

  24. [32]

    Conformer: Convolution- augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu et al. , “Conformer: Convolution- augmented transformer for speech recognition,” in Proc. Inter- speech, 2020, pp. 5036–5040

    2020

  25. [33]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” in Proc. Interspeech, 2023, pp. 3222–3226

    2023

  26. [34]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” in Proc. ICLR, 2019

    2019

Pith tools

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