Pith. sign in

REVIEW 4 major objections 5 minor 31 references

The paper claims that a WavLM-TDNN speaker-verification model, trained in two stages on mostly unlabeled classroom audio, consistently beats both an off-the-shelf and a classroom-adapted ECAPA-TDNN, with average relative EER reductions of 2

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

On a private 18-classroom dataset, WavLM-TDNN with two-stage self-supervised then supervised training achieves the lowest equal error rates among the classroom speaker verification systems tested.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible applied result for classroom speaker verification, but the evaluation protocol is under-specified and possible train/test leakage makes the headline numbers hard to trust as-is. the 4 major comments →

arxiv 2608.03623 v1 pith:XCA34EJB submitted 2026-08-04 eess.AS

Speaker Verification Under Real Classroom Conditions for English Speech

Saba Tabatabaee , Jing Liu , Megh Krishnaswamy , Carol Espy-Wilson This is my paper
classification eess.AS
keywords speaker verificationclassroom environmentschildren's speechself-supervised learningWavLM-TDNNmomentum contrasttwo-stage trainingequal error rate
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 is trying to establish that speaker verification can work in real classrooms, where children and adults talk over each other, if the model is built from a large self-supervised speech representation (WavLM) plus a TDNN backend and trained in two stages. It uses 18 English-speaking 6th to 8th grade classrooms, mostly unlabeled, and shows that first pre-training with momentum-contrast self-supervision and then fine-tuning on a small labeled subset beats SSL alone by an average relative EER reduction of 13.39%. Against the standard ECAPA-TDNN baseline trained on adult speech, the proposed WavLM-TDNN achieves a 23.99% average relative EER reduction; against an ECAPA-TDNN trained on the same classroom data, 6.32%. The result matters because annotation of classroom audio is expensive, and a model that learns from unlabeled classroom recordings plus a little supervision could make classroom analytics feasible.

Core claim

On the paper's terms, the central discovery is that a WavLM-TDNN architecture—WavLM-Large as frontend with a learnable weighted sum over its 25 transformer layers, followed by two dense branches, five TDNN blocks, and attention pooling—consistently outperforms ECAPA-TDNN for classroom speaker verification when both are trained on the same EDSI classroom data. Across five folds, WavLM-TDNN reaches an average EER of 15.40% versus 16.44% for classroom-adapted ECAPA-TDNN and 20.26% for the off-the-shelf ECAPA-TDNN baseline, corresponding to average relative reductions of 6.32% and 23.99%. The paper also establishes that two-stage training—MoCo SSL pre-training on unlabeled classroom audio follow

What carries the argument

The load-bearing object is the WavLM-TDNN model: WavLM-Large extracts frame-level representations from raw audio; a learnable softmax-weighted sum fuses its 25 layer outputs; two dense branches feed a dilated TDNN stack and a GELU branch, and both are aggregated by attention pooling into a 256-dimensional speaker embedding. During training, the WavLM encoder is not frozen, so the whole network adapts to classroom acoustics. The training machinery is MoCo-style self-supervision with a momentum-updated key encoder, InfoNCE loss, a 65,536-entry queue, and augmentations (noise, reverberation, random chunks), optionally followed by supervised fine-tuning with cross-entropy, AAM-Softmax, and tripl

Load-bearing premise

The load-bearing premise is that the five folds are independent measurements; with 18 classrooms and four test classrooms per fold, the test sets are not disjoint, and the paper reports no seed or per-fold variance, so the consistency and average reductions are weaker than five independent confirmations.

What would settle it

Rerun the five-fold protocol with strictly disjoint test classrooms (e.g., four test folds of four classrooms plus one fold of two, or a leave-classrooms-out split with a fixed seed) and compute average EER. If WavLM-TDNN no longer beats the classroom-trained ECAPA-TDNN by roughly 6.32% relative, or if the two-stage advantage over SSL-only disappears, the paper's main claim fails. Also check whether the two reused test classrooms carry the reported gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the two-stage gain holds, speaker-verification systems for classrooms can be built from mostly unlabeled recordings plus a small labeled sample, cutting annotation effort.
  • If WavLM-TDNN's advantage holds, the standard adult-trained ECAPA-TDNN is a weak proxy for classroom performance; classroom-domain data should be part of evaluation.
  • The 18.85% average relative EER improvement from adapting ECAPA-TDNN to classroom data implies domain adaptation alone accounts for much of the gain.
  • The higher EER in folds with more group-discussion babble implies noise conditions, not just speaker age, determine classroom SV difficulty, so systems should be tuned to babble-heavy conditions.
  • The cross-lingual classroom evaluation suggests the WavLM-TDNN benefit may transfer to non-English classrooms, though the test set is small (120 utterances, 29 minutes).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next experiment is to vary the amount of labeled data in fine-tuning (e.g., one session vs several per classroom) to map the annotation-cost curve implied by the two-stage result.
  • Since babble from children is generated from a separate children's corpus, the same augmentation recipe could be tested on other classroom corpora with different microphone layouts to see if the WavLM-TDNN advantage is robust to recording hardware.
  • The English-trained WavLM-TDNN's improvement on the Spanish-English classroom session suggests the layer-wise fusion of WavLM features may be capturing language-independent speaker traits; this could be tested by holding out all Spanish labels and measuring EER as a function of labeled Spanish data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a WavLM-TDNN speaker verification (SV) system for real classroom speech containing children and adults. Using an in-house EDSI dataset of 18 English-speaking classrooms with partial speaker labels, the authors compare two training strategies: MoCo-based SSL alone and a two-stage approach (SSL pre-training followed by supervised fine-tuning). They also compare against an off-the-shelf SpeechBrain ECAPA-TDNN (base) model and an ECAPA-TDNN trained on EDSI with the same two-stage procedure. Five-fold classroom-level cross-validation reports average EERs of 15.40% for two-stage WavLM-TDNN versus 17.78% for SSL-only WavLM-TDNN, 20.26% for ECAPA-TDNN (base), and 16.44% for EDSI-trained ECAPA-TDNN, corresponding to relative improvements of 13.39%, 23.99%, and 6.32%, respectively. A small multilingual classroom session (120 utterances, 29 minutes) is used to claim cross-lingual generalization, with WavLM-TDNN achieving 15.89% EER versus 22.43% and 16.82% for the baselines.

Significance. If the results hold, the paper would make a useful empirical contribution: it is apparently the first SV evaluation on a real classroom corpus containing both children and adults with pre-trained speech models, and it demonstrates that SSL pre-training plus limited supervised fine-tuning is beneficial in this setting. Strengths include the use of a real-world classroom dataset, classroom-level splits to avoid speaker overlap, enrollment data held out from training, and comparison against both an external baseline and an in-domain retrained baseline. The main claims are falsifiable and the tables are internally consistent, but the evaluation protocol has gaps that currently weaken the strength of the central claims.

major comments (4)
  1. [§2.3.1–2.3.2, Tables 1–2] The manuscript does not state whether the MoCo SSL pre-training is performed per fold with the four test classrooms excluded. Section 2.3.2 says pre-training is done 'on the EDSI dataset, including both EDSI (WO-ID) and EDSI (W-ID)' with no mention of excluding test classrooms. If pre-training runs once on all 18 classrooms, then every fold's test classrooms contribute to the learned WavLM representations, breaking the holdout assumption and biasing the comparison in favor of WavLM-TDNN, especially against the off-the-shelf ECAPA-TDNN (base). This is load-bearing for the paper's central generalization claim and must be clarified; if the pre-training is fold-specific, the exact per-fold procedure and data split must be described.
  2. [§2.1, five-fold protocol] The protocol 'randomly selecting four classrooms from EDSI (W-ID) as the test set in each fold' consumes 20 test-classroom slots from 18 classrooms, so at least two classrooms must appear in more than one test fold. No seed, no disjoint split, and no per-fold classroom composition are reported. Consequently, the 'consistently outperforms across all five folds' language and the average relative EER reductions are not equivalent to five independent confirmations. The paper should report the exact split, whether test classrooms are reused, and ideally use disjoint test sets or account for repeated classrooms in the analysis.
  3. [Tables 1–3] The evaluation reports only point estimates of EER. There are no standard deviations, confidence intervals, bootstrap resamples, or significance tests. Some reported differences are small (e.g., 16.01 vs 16.81 in Fold 3, Table 2) and may be within trial-level noise. The authors should provide uncertainty measures, such as paired bootstrap over speakers/classrooms or at least per-fold trial counts and score distributions, to support the claimed consistent improvements.
  4. [§3.3, Table 3] The multilingual generalization claim is based on a single session of 120 utterances and 29 minutes of speech. No speaker count, trial count, or confidence interval is given, and the text calls this a 'multi-lingual classrooms dataset' although it is one annotated session. Such a tiny sample cannot support the conclusion that the model 'generalizes effectively to non-English classroom speech.' The authors should either substantially expand this evaluation or substantially soften the claim to a preliminary observation.
minor comments (5)
  1. [§1] Typo: 'with their optimization is being equivalent up to' should read 'with their optimization being equivalent up to'.
  2. [§2.1] The random selection of 18 classrooms and the five-fold selection are not reproducible without a seed or detailed split listing.
  3. [§2.3.1] Reproducibility details are missing: batch size, number of GPUs, audio sampling rate, chunk size, and the exact distribution of SNR/augmentation parameters are not reported.
  4. [§3.1] The explanation that EER variation across folds is caused by group discussion versus instructional settings is post hoc and not supported by any quantitative measure of babble or discussion activity.
  5. [§3.3] The phrase 'multi-lingual classrooms dataset' overstates what is a single 120-utterance session; rephrase to 'a multi-lingual classroom session' and provide basic descriptive statistics.

Circularity Check

0 steps flagged

No significant circularity: the paper reports measured EER comparisons against an external baseline; no claimed result reduces to its own inputs.

full rationale

The paper's central claims are empirical: WavLM-TDNN with two-stage training achieves lower EER than ECAPA-TDNN baselines on an in-house classroom corpus. The EER numbers are obtained by evaluating trained models on classroom-level held-out test folds, not derived from the training objective or from any fitted parameter renamed as a prediction. The baselines include an off-the-shelf ECAPA-TDNN from SpeechBrain trained on VoxCeleb, which is external to this work. The only self-citation, [5], appears in a routine literature-review sentence about limited prior classroom SV work; it is not load-bearing and does not define or force any result. The SSL and two-stage training procedures are conventional (MoCo, fine-tuning with CE/ArcFace/triplet losses) and are not circular: the SSL objective does not incorporate the test EER, and the supervised fine-tuning labels are speaker IDs, not the evaluation metric. A possible concern that SSL pre-training may have included test-classroom data would be a data-leakage / experimental-validity issue, not a circularity issue under the definitions used here. No equation, fitted value, or self-citation is shown to be equivalent to the claimed outcomes. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on empirical assumptions about transferability of WavLM, annotation quality, speaker independence between classrooms, and the SSL objective. Hyperparameters are numerous and unablated. No new theoretical entities are introduced.

free parameters (5)
  • MoCo contrastive hyperparameters = m=0.996, tau=0.07, queue K=65536
    Hand-set; control the SSL objective and embedding quality; no ablation reported.
  • Augmentation policy = noise+reverb probability 0.6, SNR 0-15 dB, babble from 5-20 speakers
    Hand-designed to make embeddings invariant to noise; performance of both training strategies depends on it.
  • Optimization schedule = SSL 250 epochs, lr 5e-3 to 5e-5; fine-tune lr 5e-5 to 5e-7; SGD momentum 0.9, weight decay 1e-4
    Chosen by hand; the two-stage vs SSL comparison could change with different schedules.
  • Five-fold test-set selection = 4 classrooms per fold, random selection, no seed
    With 18 classrooms this implies overlapping test sets across folds; the average EER and 'consistent' claim depend on this random choice.
  • Architecture dimensions = TDNN dims (512,512,512,512,1500), embedding 256, dual branches 512/256
    Hand-designed; no ablation or comparison against other capacities reported.
axioms (5)
  • domain assumption WavLM-Large's pre-trained representations are speaker-discriminative and robust to classroom babble for child and adult speech.
    Section 2.2 builds the whole model on WavLM-Large features; if the representations do not transfer to this domain, the architecture's advantage collapses.
  • domain assumption The EDSI (W-ID) annotator-assigned speaker labels are correct.
    Section 2.1: two annotators labeled utterances using seating charts and video, but no inter-annotator agreement is reported; all EER measurements use these labels as ground truth.
  • domain assumption No speaker appears in more than one classroom, so classroom-level splitting guarantees speaker-independent train and test sets.
    Section 2.1 states 'there is no speaker overlap between classrooms' but provides no verification; if false, enrollment or test speakers could leak into training.
  • domain assumption MoCo/InfoNCE on unlabeled utterances learns speaker-discriminative embeddings.
    Section 2.3.1 adopts this from prior literature; the paper does not analyze whether positive pairs (augmented chunks of the same utterance) are sufficient to isolate speaker identity in multi-speaker babble.
  • ad hoc to paper The single 29-minute, 120-utterance multilingual session is representative of cross-lingual classroom conditions.
    Section 3.3 uses this one session to claim cross-lingual generalization; the sample is far too small to support the claim.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Speaker Verification Under Real Classroom Conditions for English Speech." pith.science (2026). https://pith.science/paper/XCA34EJB

@misc{pith2026260803623,
  author       = {Pith},
  title        = {Pith review of: Speaker Verification Under Real Classroom Conditions for English Speech},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XCA34EJB}},
  note         = {Machine review of arXiv:2608.03623}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Developing speaker verification (SV) models that are robust to classroom noise and effective across both children and adult speakers is critical for AI tools supporting educational environments. In this study, we use a real-world English-speaking classrooms dataset containing partial speaker identity annotations, with most recordings remaining unlabeled. We adapt the WavLM-TDNN model for classroom SV, achieving average relative reductions in Equal Error Rate (EER) of 23.99% and 6.32% compared to the ECAPA-TDNN baseline and the ECAPA-TDNN model trained on classroom data, respectively. Additionally, we investigate two training strategies for SV in classroom settings: self-supervised learning (SSL) and a two-stage approach that first pre-trains with SSL and then fine-tunes with limited annotated data. Five-fold cross-validation demonstrates that the two-stage strategy consistently outperforms the SSL-only approach, achieving an average relative EER reduction of 13.39%.

Figures

Figures reproduced from arXiv: 2608.03623 by Carol Espy-Wilson, Jing Liu, Megh Krishnaswamy, Saba Tabatabaee.

Figure 1
Figure 1. Figure 1: Proposed WavLM-TDNN model architecture. LSSL = 1 N XN i=1 − log exp  xqi · x + ki /τ PK j=0 exp xqi · xki,j /τ (2) where xqi denotes a query sample (i.e., anchor) and x + ki is its corresponding positive key. The negative keys x − ki,j are en￾coded by the key (momentum) encoder. In Equation 2, the sum is over one positive pair and K negative pairs, with τ = 0.07 as the temperature hyperparameter. To inc… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages · 1 internal anchor

  1. [1]

    Introduction Robust speaker verification (SV) systems for classroom envi- ronments are essential for enabling intelligent educational tech- nologies that facilitate the monitoring of students participation and teacher–student interactions. However, most existing SV research has focused primarily on adult speech [1–4], with com- paratively limited attentio...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Dataset description We used an in-house multi-modal classrooms dataset, referred to as the EDSI dataset

    Methodology 2.1. Dataset description We used an in-house multi-modal classrooms dataset, referred to as the EDSI dataset. EDSI consists of audio and video record- ings from 6th to 8th grade mathematics classrooms, along with relevant student information such as demographic information, student achievement data, seating charts, and psychometric data collec...

    work page arXiv

  3. [3]

    Results and discussion 3.1. Comparison of the training methods for the WavLM- TDNN Table 1 compares the SSL and two-stage training strategies ap- plied to the WavLM-TDNN model across five folds, evaluated using EER. The two-stage approach consistently outperforms the SSL-only method across all five folds, achieving an average relative reduction of 13.39% ...

    work page

  4. [4]

    Conclusions and future work In this study, we investigate speaker verification (SV) in real English-speaking classroom settings, which include both chil- dren and adults and challenging acoustic conditions such as babble noise during group discussions. We demonstrate the ef- fectiveness of a two-stage training strategy for SV , in which the model is first...

    work page

  5. [5]

    Self-supervised speaker verification with relational mask predic- tion,

    J.-h. Kim, H.-S. Heo, B.-J. Lee, Y . Kwon, M. Lee, and H.-J. Yu, “Self-supervised speaker verification with relational mask predic- tion,” inProc. Interspeech 2024, 2024, pp. 2655–2659

    work page 2024

  6. [6]

    Disentangling speaker and content in pre-trained speech models with latent diffusion for robust speaker verification,

    Z. Li, M.-W. Mak, J.-T. Chien, M. Pilanci, Z. Jin, and H. Meng, “Disentangling speaker and content in pre-trained speech models with latent diffusion for robust speaker verification,” inProc. In- terspeech, 2025, pp. 1108–1112

    work page 2025

  7. [7]

    Idir: Identifying and distilling informative relations for speaker verification,

    C.-X. Gan, Z. Li, Z. Jin, Z. Huang, M.-W. Mak, and K. A. Lee, “Idir: Identifying and distilling informative relations for speaker verification,” inProc. Interspeech 2025, 2025, pp. 5758–5762

    work page 2025

  8. [8]

    Noise-robust speaker verifica- tion with attenuated speech restoration and consistency training,

    S. Han, Y . Ahn, and J. W. Shin, “Noise-robust speaker verifica- tion with attenuated speech restoration and consistency training,” IEEE Transactions on Audio, Speech and Language Processing, 2025

    work page 2025

  9. [9]

    FT-Boosted SV: To- wards Noise Robust Speaker Verification for English Speaking Classroom Environments ,

    S. Tabatabaee, J. Liu, and C. Espy-Wilson, “FT-Boosted SV: To- wards Noise Robust Speaker Verification for English Speaking Classroom Environments ,” inInterspeech 2025, 2025, pp. 2815– 2819

    work page 2025

  10. [10]

    Nresnet: nested residual network based on channel and frequency domain attention mechanism for speaker verification in classroom,

    Q. Zheng, Z. Chen, X. Jiang, M. Lin, M. Wang, and Y .-L. Lu, “Nresnet: nested residual network based on channel and frequency domain attention mechanism for speaker verification in classroom,”Multimedia Tools and Applications, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:270549454

    work page 2024

  11. [11]

    Automatic speaker verification system substantiating children’s dialects in school settings,

    V . Kadyan, P. Bawa, R. Choudhary, and B. Dua, “Automatic speaker verification system substantiating children’s dialects in school settings,”2022 Seventh International Conference on Parallel, Distributed and Grid Computing (PDGC), pp. 744– 748, 2022. [Online]. Available: https://api.semanticscholar.org/ CorpusID:257261238

    work page 2022

  12. [12]

    Momentum con- trast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum con- trast for unsupervised visual representation learning,” inProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    work page 2020

  13. [13]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bo- janowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 9650–9660

    work page 2021

  14. [14]

    Self-supervised learning with cluster-aware-dino for high-performance robust speaker verifica- tion,

    B. Han, Z. Chen, and Y . Qian, “Self-supervised learning with cluster-aware-dino for high-performance robust speaker verifica- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 529–541, 2023

    work page 2023

  15. [15]

    Contrastive self-supervised speaker embedding with sequential disentanglement,

    Y . Tu, M.-W. Mak, and J.-T. Chien, “Contrastive self-supervised speaker embedding with sequential disentanglement,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2704–2715, 2024

    work page 2024

  16. [16]

    On the duality between contrastive and non-contrastive self-supervised learning,

    Q. Garrido, Y . Chen, A. Bardes, L. Najman, and Y . Lecun, “On the duality between contrastive and non-contrastive self-supervised learning,”arXiv preprint arXiv:2206.02574, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Front-end factor analysis for speaker verification,

    N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,”IEEE Trans- actions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2010

    2010

  18. [18]

    X-vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-vectors: Robust dnn embeddings for speaker recognition,” in2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5329–5333

    2018

  19. [19]

    Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    Pith/arXiv arXiv 2005

  20. [20]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  21. [21]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  22. [22]

    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

  23. [23]

    Large-scale self-supervised speech representation learning for automatic speaker verification,

    Z. Chen, S. Chen, Y . Wu, Y . Qian, C. Wang, S. Liu, Y . Qian, and M. Zeng, “Large-scale self-supervised speech representation learning for automatic speaker verification,” inICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6147–6151

    work page 2022

  24. [24]

    Multi-feature integration for speaker embed- ding extraction,

    S. Sankalaet al., “Multi-feature integration for speaker embed- ding extraction,” inICASSP 2022-2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7957–7961

    work page 2022

  25. [25]

    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

    Pith/arXiv arXiv 2021

  26. [26]

    Robust speaker recognition with transformers using wav2vec 2.0,

    S. Novoselov, G. Lavrentyeva, A. Avdeeva, V . V olokhov, and A. Gusev, “Robust speaker recognition with transformers using wav2vec 2.0,”arXiv preprint arXiv:2203.15095, 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Layer- aware tdnn: Speaker recognition using multi-layer features from pre-trained models,

    J. S. Kim, H. J. Park, W. Shin, J. Yun, and S. W. Han, “Layer- aware tdnn: Speaker recognition using multi-layer features from pre-trained models,”arXiv preprint arXiv:2409.07770, 2024

    work page arXiv 2024

  28. [28]

    In- terspeech 2021 deep noise suppression challenge,

    C. K. Reddy, H. Dubey, K. Koishida, A. Nair, V . Gopal, R. Cut- ler, S. Braun, H. Gamper, R. Aichner, and S. Srinivasan, “In- terspeech 2021 deep noise suppression challenge,” inINTER- SPEECH, 2021

    work page 2021

  29. [29]

    My sci- ence tutor (myst)–a large corpus of children’s conver- sational speech,

    S. Pradhan, R. A. Cole, and W. H. Ward, “My sci- ence tutor (myst)–a large corpus of children’s conver- sational speech,” inInternational Conference on Lan- guage Resources and Evaluation, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:262458961

    work page 2023

  30. [30]

    Speechbrain: A general-purpose speech toolkit,

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong et al., “Speechbrain: A general-purpose speech toolkit,”arXiv preprint arXiv:2106.04624, 2021

    Pith/arXiv arXiv 2021

  31. [31]

    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

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.