Pith. sign in

REVIEW 3 major objections 6 minor 38 references

FT-Boosted SV: Towards Noise Robust Speaker Verification for English Speaking Classroom Environments

T0 review · 3 major / 6 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 speaker-verification models on augmented children's speech roughly halves equal-error rates in real English-speaking classroom recordings.

desk verdict A useful domain-adaptation study for classroom SV, but the children's evaluation likely has speaker leakage and the classroom numbers lack uncertainty; deserves a major-revision review, not a desk reject. read the letter →

arxiv 2505.20222 v1 pith:W2IDICEO submitted 2025-05-26 eess.AS

Saba Tabatabaee , Jing Liu , Carol Espy-Wilson This is my paper · ORCID
classification eess.AS
keywords SpeakerVerificationClassroomChildren'sSpeechDataAugmentationx-vectorECAPA-TDNNBabbleNoiseFine-tuning
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 claims that fine-tuning pretrained speaker-verification models on children's speech mixed with classroom-like noise makes the models work in real English-speaking classrooms, a setting they were not trained for. The authors combine three children's corpora covering conversational, prompted/spontaneous, and read speech into a single development set, then add background noise, simulated children's babble, and reverberation before fine-tuning x-vector and ECAPA-TDNN embeddings. On averaged classroom error rates, augmented fine-tuning cuts the ECAPA-TDNN error on MPT classrooms roughly in half, from 10.33% to 5.35%, and reduces the x-vector error on NCTE classrooms from 21.97% to 14.32%. If this holds, it offers a practical recipe for adapting speaker verification to noisy educational settings without requiring a large in-domain classroom corpus.

What carries the argument

The mechanism is a two-stage fine-tuning pipeline on a combined children's development set. The development set merges conversational, spontaneous/prompted, and read children's speech to cover diverse ages and speaking styles; classroom-like conditions are injected by mixing background noises and simulated children's babble at 5 to 15 dB signal-to-noise ratios and by convolving with reverberant room impulse responses. Training uses triplet loss with online hard batch mining on the speaker embeddings, and verification is performed by cosine similarity with score normalization.

What would settle it

Count overlapping speaker IDs between the development and test partitions of the combined children's corpus; if any overlap exists, remove those speakers' test utterances and recompute the equal error rates, or evaluate on a fully held-out children's corpus, and check whether the fine-tuned gains persist.

Watch

Extended reading notes

Core claim

The central discovery is that domain-specific fine-tuning with augmented children's data transfers to real classrooms. Starting from pretrained x-vector and ECAPA-TDNN models, the authors fine-tune all parameters on 446 hours of children's speech assembled from three corpora, then repeat fine-tuning with the same data mixed with background noises, simulated babble from 12 to 25 randomly selected children, and room impulse responses at signal-to-noise ratios from 5 to 15 dB. The resulting FT-Boosted ECAPA-TDNN reduces the average equal error rate on MPT classrooms from 10.33% to 5.35% and on NCTE classrooms from 8.81% to 5.44%, while the FT-Boosted x-vector reduces the average NCTE error from 21.97% to 14.32%; both models also improve on children's test sets. An ablation shows that combining all three children's corpora beats any subset, and the best model retains strong performance on adult speech test sets.

Load-bearing premise

The reported children's-test improvements assume the fine-tuning and test partitions contain no shared speakers; if a child appears in both, the gains are optimistically biased.

Editorial extensions

If this is right

  • A usable speaker-verification system for English-speaking classrooms can be built by fine-tuning a pretrained model on a few hundred hours of children's speech rather than collecting a large in-domain classroom corpus.
  • For classroom settings, a development set that mixes read, conversational, and spontaneous children's speech outperforms any single children's corpus, so corpus diversity is part of the recipe.
  • ECAPA-TDNN appears to be the stronger choice among the two architectures tested, especially for far-field microphones and classrooms with heavy babble.
  • Adult speech performance is not sacrificed by children's-data fine-tuning, which matters because teachers' voices share the same classroom audio.

Reading between the lines

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

  • Editorial inference: If the same augmented fine-tuning recipe transfers to other low-resource noisy speech tasks, the key design choice to copy is generating babble from the target population rather than relying only on generic noise corpora.
  • Editorial inference: The paper's per-classroom results suggest that microphone configuration and group-discussion babble, not just overall noise level, drive the hardest cases; a direct test would be to rank classrooms by measured babble-to-speech ratio and correlate it with equal error rate.
  • Editorial inference: Because the classroom labels come from automatic diarization and transcription, the reported error rates include label noise; a cleaner test of the core claim would re-evaluate on a small manually transcribed classroom subset.
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

3 major / 6 minor

Summary. The paper proposes fine-tuning two pretrained speaker verification models, x-vector and ECAPA-TDNN, on augmented children's speech data (MyST, CSLU, CMU, collectively MCC) to improve robustness in English-speaking classroom environments. The augmentation includes MUSAN noises, simulated children's babble, and reverberation. The authors report that fine-tuning with augmented MCC data reduces equal error rate (EER) on both classroom test sets (NCTE, MPT) and on children's speech test sets (MCC/MyST/CMU/CSLU), with the largest gains for ECAPA-TDNN on MPT and for x-vector on NCTE. An ablation study compares combinations of the three children's datasets and concludes that the full MCC combination is optimal.

Significance. If the reported gains are reliable, the paper would be a useful practical contribution: it is the first study, to my knowledge, targeting speaker verification in real English-speaking classrooms, it evaluates on actual classroom recordings rather than simulated conditions, and it provides a concrete recipe (fine-tuning with in-domain augmented children's speech) that could transfer to related low-resource classroom domains. The inclusion of adult test sets (LibriSpeech, VoxCeleb1) as a control for speaker-age trade-off is a strength, as is the use of two widely used architectures. The central direction is plausible and the classroom results are suggestive. However, the reliability of the quantitative claims is weakened by apparent speaker overlap in the children's evaluation split, by test-set-based selection in the ablation study, and by very small classroom evaluation sets with automatic labels and no uncertainty estimates.

major comments (3)
  1. [Section 2, MCC development/test split] The reported speaker counts are internally inconsistent. Section 2 states that the MCC development set contains 2,544 speakers and the test set contains 2,540 speakers, but the filtered MyST, CSLU, and CMU corpora together contain only 1,354 + 1,100 + 74 = 2,528 speakers. Since the development set alone exceeds the total pool, the two sets cannot be speaker-disjoint. The description 'stratification based on speaker utterances' suggests an utterance-level split in which the same speaker appears on both sides of the boundary. Under that protocol, the triplet-loss fine-tuning has already seen the test speakers' voices, so the MCC/MyST/CMU/CSLU EERs in Table 1 and the abstract's claim of improved children's speech verification are optimistically biased. This needs to be fixed by a speaker-disjoint split (or by a clear, justified cross-validation scheme) and by re-reporting the children's results.
  2. [Section 4.3, Table 3] The ablation study selects the development set based on the lowest EER on the NCTE and MPT classroom test sets. Section 4.3 states that each fine-tuned model is 'evaluating its performance on both the MPT and NCTE classroom datasets' and that MCC is chosen because it 'achieves the lowest EER for both the NCTE and MPT datasets.' This is a form of test-set model selection: the same classroom recordings are used to choose the best development set and then to report the final classroom EERs for that best configuration. As a result, the reported improvements for the MCC configuration are selection-biased. The authors should either hold out the classroom test sets during development-set selection, or use a nested/leave-one-classroom-out procedure, or at minimum disclose the selection protocol and report the variance across the eight NCTE and two MPT classrooms.
  3. [Section 2.0.6, Section 2.0.7, Tables 1 and 2] The classroom evaluation is very small and the reported EERs are single point estimates without confidence intervals or significance tests. NCTE contributes 2.78 hours and 2,918 utterances from 8 classrooms, and MPT contributes 1.02 hours and 465 utterances from 2 classrooms. The speaker labels for NCTE and MPT come from an automatic Pyannote/TranscribeMe pipeline with no reported validation, and the MPT recordings use different microphone configurations that are averaged. Given the small test sets and the noisy labels, the absolute differences between baseline and fine-tuned EERs in Table 2 may be within the range of sampling variability. I ask the authors to report per-classroom bootstrap confidence intervals or another measure of uncertainty, and to state the expected label-error rate of the diarization pipeline.
minor comments (6)
  1. [Abstract and Section 4.1] The phrases 'a 5% improvement' and 'an 8% average improvement' are ambiguous: Table 1 shows absolute EER reductions of about 4.98 and 7.65 percentage points for ECAPA-TDNN (MPT) and x-vector (NCTE), respectively, which correspond to much larger relative improvements. Please consistently distinguish percentage points from relative percentages.
  2. [Section 2] The sentence describing the split says 'stratification based on speaker utterances'; this is unclear. Please state explicitly whether the split is speaker-disjoint or utterance-disjoint, and how speaker IDs were handled across the MyST, CSLU, and CMU corpora.
  3. [Table 1 and Table 2] The column headers in Table 2 mix classroom identifiers (31, 201, 230, etc.) with dataset names (NCTE, MPT) without a legend. Please add a note explaining that these are recording IDs and which rows refer to which classroom.
  4. [Section 3, Methodology] The text says 'computing the cosine distance between speaker embeddings'; speaker verification typically uses cosine similarity, and the EER threshold is applied to similarity scores. Please correct the wording.
  5. [Section 4.3, Table 3] The ablation table reports averages over classrooms for each development set, but with only two MPT classrooms and eight NCTE classrooms, a per-classroom breakdown would help the reader judge whether the MCC advantage is consistent or driven by a single recording.
  6. [General] For reproducibility, please report the number of training epochs actually used under early stopping, the learning-rate schedule parameters, and the random seed(s). The paper currently lists only the optimizer, patience, and decay values.

Circularity Check

2 steps flagged · score 6.0 of 10

Apparent MCC dev/test speaker overlap and Table 3 test-set selection make headline EER gains partly circular

  1. fitted input called prediction [Section 2 (Dataset Description) and Section 3 (Methodology), with results in Table 1]
    "Each dataset was divided into 85% for development (70% training, 15% validation) and 15% for testing, with stratification based on speaker utterances to address data imbalance. ... the MCC development set included 161,722 number of utterances and 446.46 hours of speech from 2,544 speakers, and the testing set contained 28,540 number of utterances and 78.36 hours of speech from 2,540 speakers. ... We unfroze all model parameters (x-vector: 4,211,604; ECAPA-TDNN: 20,767,552) and finetuned them using the MCC development set in two stages."

    The three corpora contain only about 2,528 unique speakers (1,354 MyST + 1,100 CSLU + 74 CMU). A 2,544-speaker development set and a 2,540-speaker test set cannot both be speaker-disjoint subsets of that pool; each count already exceeds the total. The described 'stratification based on speaker utterances' therefore places the same speakers' voices on both sides of the train/test boundary. Finetuning on the MCC development set is thus fitted to the very speakers whose utterances appear in the MCC/MyST/CMU/CSLU test sets, so Table 1's children's-speech EER reductions are optimistically biased by construction and are not predictions for unseen speakers.

  2. fitted input called prediction [Section 4.3 (Analysis of the choice of development set), Table 3]
    "In each subsequent step, we finetuned the ECAPA-TDNN (FT-Boosted) model with different combinations of datasets, evaluating its performance on both the MPT and NCTE classroom datasets. ... Finally, the results demonstrate that finetuning with the combination of MyST, CSLU, and CMU datasets (referred to as MCC) achieves the lowest EER for both the NCTE and MPT datasets."

    The choice of MCC as the 'optimal development set' is made by minimizing EER on the exact NCTE and MPT classroom test sets that are later reported as the method's classroom results (Tables 1 and 2). No independent validation set is used for this selection. The claim that MCC 'achieves the lowest EER' is therefore the selection criterion itself: reporting the best-of-four test-set minimum as the FT-Boosted classroom improvement is test-set model selection, which forces optimistic point estimates rather than providing independent confirmation.

full rationale

The paper's headline claim has two legs. The children's-speech leg is not an independent evaluation: under the paper's own speaker counts, both the MCC development set (2,544) and the MCC test set (2,540) exceed the total unique speakers in MyST+CSLU+CMU (about 2,528), so a speaker-disjoint split is impossible; finetuning on this development set means the children's test EERs in Table 1 are measured on voices already seen during training. The classroom leg uses external NCTE/MPT recordings and is not affected by that speaker overlap, but Section 4.3 selects the MCC development set by comparing EER on those same classroom test sets (Table 3) and then reports those numbers as the method's classroom performance, making the 'MCC is optimal' conclusion a test-set minimum rather than a held-out finding. The adult LS/Vox results are external and not the paper's main contribution, so they do not validate the headline. These issues are data-split leakage and test-set selection rather than self-citation or definitional identity, but they mean the reported reductions are partly forced by the evaluation protocol.

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

The central claim rests on assumptions about data representativeness, label quality, and the transferability of pretrained models. The main free parameter is the choice of the MCC development set, which was selected via the classroom test sets.

free parameters (4)
  • Finetuning dataset combination = MCC (MyST + CSLU + CMU)
    Selected by comparing EER on NCTE and MPT test sets (Section 4.3, Table 3); effectively a free choice tuned on test data.
  • Noise SNR range = 5 to 15 dB
    Chosen by hand; no sensitivity analysis shown (Section 3).
  • Babble mix size = 12 to 25 children
    Chosen by hand; no sensitivity analysis shown (Section 3).
  • Training hyperparameters (early stopping, LR schedule) = patience 8, decay 0.5
    Standard choices, but not fully specified (learning rate, batch size, triplet margin omitted), hampering replication (Section 3).
assumptions (4)
  • domain assumption Pretrained VoxCeleb x-vector and ECAPA-TDNN are suitable starting points for classroom SV.
    The paper fine-tunes these models without investigating alternatives; assumes VoxCeleb pretraining transfers.
  • domain assumption Pyannote diarization and TranscribeMe outputs can serve as ground truth speaker labels for classroom evaluation.
    Used to create labels and timestamps for NCTE and MPT (Sections 2.0.6-2.0.7); errors would bias EER.
  • domain assumption The three children's datasets (MyST, CSLU, CMU) together represent classroom speech.
    Justified by diversity of age and speech style (Section 4.3), but not empirically validated.
  • domain assumption Simulated babble noise and MUSAN noises approximate real classroom noise.
    The paper proposes this equivalence (Section 3) without measuring distribution shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FT-Boosted SV: Towards Noise Robust Speaker Verification for English Speaking Classroom Environments." pith.science (2026). https://pith.science/paper/W2IDICEO

@misc{pith2026250520222,
  author       = {Pith},
  title        = {Pith review of: FT-Boosted SV: Towards Noise Robust Speaker Verification for English Speaking Classroom Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W2IDICEO}},
  note         = {Machine review of arXiv:2505.20222}
}
read the original abstract

Creating Speaker Verification (SV) systems for classroom settings that are robust to classroom noises such as babble noise is crucial for the development of AI tools that assist educational environments. In this work, we study the efficacy of finetuning with augmented children datasets to adapt the x-vector and ECAPA-TDNN to classroom environments. We demonstrate that finetuning with augmented children's datasets is powerful in that regard and reduces the Equal Error Rate (EER) of x-vector and ECAPA-TDNN models for both classroom datasets and children speech datasets. Notably, this method reduces EER of the ECAPA-TDNN model on average by half (a 5 % improvement) for classrooms in the MPT dataset compared to the ECAPA-TDNN baseline model. The x-vector model shows an 8 % average improvement for classrooms in the NCTE dataset compared to its baseline.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Studies have primarily focused on developing SV systems for adult speech [1, 2, 3, 4, 5]

    Introduction Developing speaker verification (SV) systems for classroom en- vironments is crucial for fostering a more equitable learni ng at- mosphere, while also providing teachers with a valuable too l to improve their teaching methods. Studies have primarily focused on developing SV systems for adult speech [1, 2, 3, 4, 5]. How- ever, these systems may...

    work page

  2. [2]

    Each dataset w as divided into 85% for development (70% training, 15% valida- tion) and 15% for testing, with stratification based on speak er utterances to address data imbalance

    Dataset Description For finetuning x-vector and ECAPA-TDNN models, we used three children’s speech datasets: My Science Tutor (MyST), Center for Spoken Language Understanding (CSLU) kids speech, and Carnegie Mellon University (CMU) kids corpus, collectively referred to as MCC In this study. Each dataset w as divided into 85% for development (70% training, ...

    work page

  3. [3]

    Methodology We finetuned two pretrained models of x-vector and ECAPA- TDNN that are available in the SpeechBrain toolkit [28] . The se models were originally trained on the V oxCeleb1 and V ox- Celeb2 datasets [29], which contain over 1,150,000 utteran ces from approximately 7,363 celebrities, including backgrou nd noise, laughter, and overlapping speech. ...

    work page

  4. [4]

    Analysis of finetuning x-vector and ECAPA-TDNN Table 1 shows the EER across all test folds for the x-vector an d ECAPA-TDNN models

    Results and Discussion 4.1. Analysis of finetuning x-vector and ECAPA-TDNN Table 1 shows the EER across all test folds for the x-vector an d ECAPA-TDNN models. To evaluate the SV systems on chil- dren’s datasets, we use the MCC, which is the combination of test sets from the MyST, CSLU, and CMU datasets. Addition- ally, we report the EER for each individua...

    work page

  5. [5]

    Conclusions And Future Work In this study, we explored speaker verification in real Engli sh- speaking classrooms, where both children and adults are present. We demonstrated the effectiveness of finetuning pr e- trained models using a domain-specific, augmented dataset f or speaker verification tasks in noisy and challenging environ - ments, such as classroo...

    work page

  6. [6]

    Dynamic fully - connected layer for large-scale speaker verification,

    Z. Song, L. He, B. Zhao, M. Xu, and Y . Zheng, “Dynamic fully - connected layer for large-scale speaker verification,” in Proc. IN- TERSPEECH 2023, 2023, pp. 2003–2007

    work page 2023

  7. [7]

    Adversarial data augmentation for robust speaker verification,

    Z. Zhou, J. Chen, N. Wang, L. Li, and D. Wang, “Adversarial data augmentation for robust speaker verification,” in Proceedings of the 2023 9th International Conference on Communication a nd Information Processing, 2023, pp. 226–230

    work page 2023

  8. [8]

    Self-supervised speaker verification with relational mas k predic- tion,

    J.-h. Kim, H.-S. Heo, B.-J. Lee, Y . Kwon, M. Lee, and H.-J. Y u, “Self-supervised speaker verification with relational mas k predic- tion,” in Proc. Interspeech 2024, 2024, pp. 2655–2659

    work page 2024

Show all 38 references
  1. [9]

    Densely connected time delay neur al net- work for speaker verification

    Y .-Q. Y u and W.-J. Li, “Densely connected time delay neur al net- work for speaker verification.” in Interspeech, 2020, pp. 921–925

    2020

  2. [10]

    A n effective deep embedding learning architecture for speake r verifi- cation

    Y . Jiang, Y . Song, I. McLoughlin, Z. Gao, and L.-R. Dai, “A n effective deep embedding learning architecture for speake r verifi- cation.” in INTERSPEECH, 2019, pp. 4040–4044

    2019

  3. [11]

    Comparison of speaker verification performance for adult a nd child speech,

    S. Safavi, M. Najafian, A. Hanani, M. Russell, and P . Janˇ c oviˇ c, “Comparison of speaker verification performance for adult a nd child speech,” 2014

    2014

  4. [12]

    Acoustics of children’s speech: developmental changes of temporal an d spectral parameters

    S. W. Lee, A. Potamianos, and S. S. Narayanan, “Acoustics of children’s speech: developmental changes of temporal an d spectral parameters.” The Journal of the Acoustical Society of America, vol. 105 3, pp. 1455–68, 1999. [Online]. Available: https://api.semanticscholar.org/Co...

    1999

  5. [13]

    Speaker ver- ification experiments for adults and children using shared e mbed- ding spaces,

    T. Kaseva, H. Kathania, A. Rouhe, and M. Kurimo, “Speaker ver- ification experiments for adults and children using shared e mbed- ding spaces,” in Nordic Conference on Computational Linguistics. Link¨ oping University Electronic Press, 2021, pp. 86–93

    2021

  6. [14]

    An attention-based backend allowing efficient fine- tuning of transformer models for speaker verification,

    J. Peng, O. Plchot, T. Stafylakis, L. Moˇ sner, L. Burget, and J. ˇCernock` y, “An attention-based backend allowing efficient fine- tuning of transformer models for speaker verification,” in 2022 IEEE Spoken Language Technology W orkshop (SLT) . IEEE, 2023, pp. 555–562

    2022

  7. [15]

    Efficient adapter tuning of pre -trained speech models for automatic speaker verification,

    M. Sang and J. H. Hansen, “Efficient adapter tuning of pre -trained speech models for automatic speaker verification,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Spe ech and Signal Processing (ICASSP) . IEEE, 2024, pp. 12 131– 12 135

    2024

  8. [16]

    Fine-tune pre - trained models with multi-level feature fusion for speaker verifi- cation,

    S. Peng, W. Guo, H. Wu, Z. Li, and J. Zhang, “Fine-tune pre - trained models with multi-level feature fusion for speaker verifi- cation,” in Proc. Interspeech 2024, 2024, pp. 2110–2114

    2024

  9. [17]

    Deep speaker embeddings for speaker verification of children,

    M. H. Abed and D. Sztah´ o, “Deep speaker embeddings for speaker verification of children,” in International Conference on Text, Speech, and Dialogue. Springer, 2024, pp. 58–69

    2024

  10. [18]

    Childaugm ent: Data augmentation methods for zero-resource children’s sp eaker verification,

    V . P . Singh, M. Sahidullah, and T. Kinnunen, “Childaugm ent: Data augmentation methods for zero-resource children’s sp eaker verification,” The Journal of the Acoustical Society of America , vol. 155, no. 3, pp. 2221–2232, 2024

    2024

  11. [19]

    In - domain and out-of-domain data augmentation to improve chil - dren’s speaker verification system in limited data scenario ,

    S. Shahnawazuddin, W. Ahmad, N. Adiga, and A. Kumar, “In - domain and out-of-domain data augmentation to improve chil - dren’s speaker verification system in limited data scenario ,” in ICASSP 2020-2020 IEEE International Conference on Acousti cs, Speech and Signal Processing (...

    2020

  12. [20]

    Musan: A music, speech , and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech , and noise corpus,” ArXiv, vol. abs/1510.08484, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:15676318

    2015 arXiv

  13. [21]

    Nresnet: nested residual network based on channel and frequency domain attention mechanism for speaker verificat ion 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 verificat ion in classroom,” Multimedia Tools and Applications , 2024. [Online]. Available: https://api.semantics...

    2024

  14. [22]

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

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

    2022

  15. [23]

    Spoken language recognition using x-vector s,

    D. Snyder, D. Garcia-Romero, A. McCree, G. Sell, D. Pove y, and S. Khudanpur, “Spoken language recognition using x-vector s,” in The Speaker and Language Recognition W orkshop , 2018. [Online]. Available: https://api.semanticscholar.org/ CorpusID: 51862809

    2018

  16. [24]

    Ecapa -tdnn: Emphasized channel attention, propagation and aggregatio n in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa -tdnn: Emphasized channel attention, propagation and aggregatio n in tdnn based speaker verification,” in Interspeech, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:218630075

    2020

  17. [25]

    Pyannote.audio 2.1 speaker diarization pi peline: Principle, benchmark, and recipe,

    H. Bredin, “Pyannote.audio 2.1 speaker diarization pi peline: Principle, benchmark, and recipe,” in Interspeech, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:260906702

    2023

  18. [26]

    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,” ArXiv, vol. abs/2310.13025, 2023. [Online]. Available: https: //api.semanticscholar.org/CorpusID:260919244

    2023 arXiv

  19. [27]

    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,” in International Conference on Lan- guage Resources and Evaluation , 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:262458961

    2023

  20. [28]

    Cslu: Kids’ speec h ver- sion 1.1,

    K. Shobaki, J.-P . Hosom, and R. Cole, “Cslu: Kids’ speec h ver- sion 1.1,” Linguistic Data Consortium, 2007

    2007

  21. [29]

    The cmu kids corpu s,

    M. Eskenazi, J. Mostow, and D. Graff, “The cmu kids corpu s,” Linguistic Data Consortium, vol. 11, 1997

    1997

  22. [30]

    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,” in 2015 IEEE international conference on acoustics, speech an d signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  23. [31]

    V oxce leb: Large-scale speaker verification in the wild,

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

    2020

  24. [32]

    The ncte transcripts: A dataset of elementary math classroom transcripts,

    D. Demszky and H. Hill, “The ncte transcripts: A dataset of elementary math classroom transcripts,” arXiv preprint arXiv:2211.11772, 2022

    2022 arXiv

  25. [33]

    Speechbrain: A general-purpose speech toolkit,

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

    2021 arXiv

  26. [34]

    V oxceleb2: D eep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: D eep speaker recognition,” in Interspeech, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:49211906

    2018

  27. [35]

    Facenet: A unified embedding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding for face recognition and clustering,” in Proceedings of the IEEE conference on computer vision and pattern recognit ion, 2015, pp. 815–823

    2015

  28. [36]

    End-to-end text-independen t speaker verification with triplet loss on short utterances

    C. Zhang and K. Koishida, “End-to-end text-independen t speaker verification with triplet loss on short utterances.” in Interspeech, 2017, pp. 1487–1491

    2017

  29. [37]

    Text-independent speaker veri fication based on triplet loss,

    J. He, J. He, and L. Zhu, “Text-independent speaker veri fication based on triplet loss,” in 2020 IEEE 4th Information Technology, Networking, Electronic and Automation Control Conference (IT- NEC), vol. 1. IEEE, 2020, pp. 2385–2388

    2020

  30. [38]

    Msrane t: Learning discriminative embeddings for speaker verificati on via channel and spatial attention mechanism in alterable scena rios,

    Q. Zheng, Z. Chen, H. Liu, Y . Lu, J. Li, and T. Liu, “Msrane t: Learning discriminative embeddings for speaker verificati on via channel and spatial attention mechanism in alterable scena rios,” Expert Systems with Applications , vol. 217, p. 119511, 2023

    2023

Pith tools

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