Pith. sign in

REVIEW 4 major objections 4 minor 30 references

Learning Emotion-Invariant Speaker Representations for Speaker Verification

T0 review · 4 major / 4 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 Training with emotion-parallel pairs and energy-based masking lowers speaker verification EER from 5.65% to 4.56%, a relative 19.29% drop.

desk verdict A careful, well-ablated empirical study of a modest emotion-invariance recipe for speaker verification; the total gain is credible, but the smallest component (EM) rests on a single 0.14-point EER gap with no uncertainty estimate. read the letter →

arxiv 2505.18498 v1 pith:SCPLO6GH submitted 2025-05-24 cs.SD eess.AS

Jingguang Tian , Xinhui Hu , Xinkang Xu This is my paper · ORCID
classification cs.SDeess.AS
keywords speakerverificationemotion-invariantrepresentationsemotion-awaremaskingCopyPastedataaugmentationcosinesimilaritylossAAM-SoftmaxDushacorpusequalerrorrate
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 sets out to make speaker verification less vulnerable to emotion by making the learned speaker representation insensitive to emotional variation. It argues that with enough parallel data—two recordings of the same speaker said with different emotions—a Siamese-trained encoder can be pushed to ignore emotion. The proposed combination of CopyPaste-based augmentation, an ERL objective, and energy-based emotion-aware masking lowers equal error rate from 5.65% to 4.56% on the merged Dusha test set, a relative 19.29% drop. A sympathetic reader would care because emotional speech is common in real verification, and this route needs no emotion labels at test time.

What carries the argument

The scheme has three linked pieces. CopyPaste-based augmentation (S-CP, D-CP, and S+D-CP) builds new training utterances by concatenating two same-speaker clips, so each pair shares an identity but differs in emotion category or intensity. Emotion-invariant Representation Learning (ERL) runs such pairs through a Siamese encoder and sums AAM-Softmax classification loss with a cosine-similarity loss that pulls the two embeddings together, matching the cosine scoring used at test time. Emotion-aware masking (EM) normalizes per-frame RMS energy, classifies frames into high-energy, low-energy, and noise zones, and randomly masks $m=2$ blocks of $T=7$ frames in the dominant zone of the emotional input only, on the assumption that intense emotion lives in loud frames and subdued emotion in quiet ones.

What would settle it

Apply the same S+D-CP + ERL pipeline to a corpus where angry and happy speech is consistently quieter than neutral speech (for example, whispered anger or suppressed excitement). If EM still beats random masking there, the energy-masking story is not what drives the gain; if EM loses its advantage, the energy-emotion mapping is confirmed as the mechanism.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that emotion variability can be treated as a data problem rather than only a model problem: concatenating same-speaker utterances creates cheap emotion-parallel pairs, and training with a cosine-similarity loss on those pairs plus masking high- or low-energy frames removes much of the emotion-induced variance in speaker embeddings. The paper reports that this reduces EER from 5.65% to 4.56% on the merged cross-emotion and same-emotion test trials, with CopyPaste augmentation contributing a 7.26% relative gain, ERL a further 9.55%, and EM a further 2.48%.

Load-bearing premise

The load-bearing premise is that a frame's RMS energy marks where the emotion is, so masking the dominant-energy frames strips emotional content while leaving speaker identity intact.

Editorial extensions

If this is right

  • If correct, emotion-invariant speaker verification can be trained without collecting new parallel emotion data, since concatenation synthesizes it from labeled single-emotion clips.
  • The cosine-similarity loss aligns the training objective with the cosine scoring backend, so the gain should transfer to standard verification pipelines without changing the scoring stage.
  • Energy-based masking is a label-free way to suppress emotion: it needs no emotion annotation at inference, only the same encoder.
  • The method narrows the gap between cross-emotion and same-emotion trials, so the largest absolute gains appear on the hardest angry-to-neutral and angry-to-sad comparisons.
  • The reported 19.29% relative EER reduction on a 1,870-speaker corpus suggests the approach scales beyond the fewer-than-100-speaker datasets used in prior emotion speaker-verification work.

Reading between the lines

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

  • The paper does not isolate which emotions benefit most from masking; a natural extension is to measure per-emotion EER with and without EM to see whether the gain comes mostly from high-arousal emotions such as anger.
  • If energy masking truly removes emotion, the same masking could be applied at test time to enrollment and probe utterances, something the paper only does during training; that is a testable extension.
  • The assumption that loudness tracks emotion is language- and recording-dependent; the method's transfer to whispered or low-effort emotional speech is an open question the paper does not address.
  • Because ERL needs parallel same-speaker pairs, its applicability to corpora without emotion labels depends on the CopyPaste synthesis step; the paper's S-CP and D-CP recipes are one possible instantiation, and other pairing rules could be explored.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a training scheme for speaker verification on emotional speech. The authors use CopyPaste-style augmentation to generate emotion-parallel utterances from the same speaker, combine AAM-Softmax with a cosine-similarity loss to pull same-speaker representations together, and add an energy-based emotion-aware masking (EM) strategy that masks frames in high- or low-energy regions. The model is pre-trained on VoxCeleb and fine-tuned on the Dusha Crowd dataset. On a merged same/cross-emotion test set, the reported EER drops from 5.65% (baseline) to 4.56% with the full system, a relative improvement of 19.29%. Ablations separate the contributions of CopyPaste augmentation, ERL, random masking, and EM.

Significance. If the reported result is statistically reliable, the paper makes a useful practical contribution: it shows that emotion-parallel data, generated via CopyPaste, combined with a cosine-similarity objective can improve speaker verification on a large emotional speech corpus. The ablations are logically ordered, the dataset and trial splits are made available, and the use of a publicly available corpus with hundreds of hours of data is a strength. The main caveat is that the entire empirical claim rests on point estimates of EER without confidence intervals, and the distinguishing EM component is supported by a single fine-tuning run. The core idea is plausible and the experimental framework is appropriate, but the statistical evidence is not yet at the level needed to support the component-wise attribution of the gain.

major comments (4)
  1. [Section 4.3.3 / Table 4 rows 15-17] The EM contribution is supported only by a single point estimate: EER moves from 4.70% (row 15) to 4.56% (row 17), an absolute drop of 0.14 percentage points, while random masking moves the EER to 4.75% (row 16). The test set contains only 84 speakers and over 400,000 highly correlated trials, so trial-level EER differences of this size can easily be within speaker-sampling or run-to-run noise. No confidence intervals, no multiple seeds, and no speaker-level bootstrap are reported. Because the paper attributes 2.48% (or 2.98%) of the total 19.29% improvement to EM, this component is load-bearing. The authors should report confidence intervals (e.g., via speaker-level bootstrap) or results across multiple random seeds, and ideally a significance test for the EM-versus-RM comparison.
  2. [Section 3.3 / hyperparameter selection] The hyperparameters alpha=1, m=2, T=7, the energy-zone thresholds 0.2/0.5, and the CopyPaste segment duration of 1 second are fixed without any mention of a validation set or sensitivity analysis. Since the test set is the only held-out evaluation described, there is a risk that some of these values were chosen on the basis of test-set performance, which would inflate the reported gains. The authors should either describe an explicit validation split and report the selection procedure, or provide a sensitivity study showing that the results are robust to reasonable variations in these values.
  3. [Section 2.3 / emotion-aware masking mechanism] The paper's premise is that RMS energy identifies frames carrying emotional content and that masking those frames removes emotional information while preserving speaker identity. This premise is asserted but not empirically supported. In particular, high-energy frames may also contain speaker-specific phonetic or prosodic information, and no analysis is shown that the masked frames align with emotion content (e.g., by comparing mask locations with emotion-label annotations or by measuring the effect of masking low- versus high-energy zones on downstream performance). Without such evidence, the mechanism underlying EM remains speculative even if the EER improvement were statistically significant.
  4. [Section 3.1 / Table 2] The construction of the test trials is under-specified. Table 2 reports the number of trials per emotion pair, but the text does not describe how target and non-target trials are formed, whether the utterances in a trial come from the same recording session or different sessions, whether each utterance is used in multiple trials, or how the emotion-label balance is maintained across the 84 speakers. This information is necessary to interpret the EER values and the same- versus cross-emotion comparisons. The authors should provide a precise description of the trial-generation process, including trial counts for target and impostor pairs.
minor comments (4)
  1. [Section 4.2 / Conclusions] The relative improvements attributed to ERL and EM are inconsistent between the ablation section and the conclusion: Section 4.3.2 reports a 10.31% ERL improvement from 5.24% to 4.70%, while the conclusion reports 9.55%; similarly Section 4.3.3 reports a 2.98% EM improvement from 4.70% to 4.56%, while the conclusion reports 2.48%. The two sets of numbers use different denominators (previous row versus original baseline); the text should state this explicitly to avoid confusion.
  2. [Throughout] The word 'trials' is frequently misspelled as 'trails' (e.g., Table 2 caption, Section 4.1). The manuscript should be proofread for this typo.
  3. [Section 2.2 / Equation (2)] Equation (2) is typeset ambiguously: the numerator and denominator are not clearly separated, and the notation for the representation vectors is difficult to read. Please rewrite the equation with explicit norms, e.g., cos(v_i_x, v_i_x') = (v_i_x · v_i_x') / (||v_i_x|| ||v_i_x'||).
  4. [Section 2.3] The sentence 'Once the emotional aspects of the speech are masked, it is considered to be devoid of emotion' is an overstatement. Masking selected frames cannot guarantee that all emotional information is removed, especially since emotion is conveyed prosodically across the whole utterance; a more cautious wording would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the ERL objective, EM masking, and end-to-end EER evaluation are separate; the reported improvements are empirical outcomes, not fitted constants or self-referential definitions.

full rationale

I walked the claimed derivation chain: the training objective in Eq. (1)-(2) combines AAM-Softmax and a cosine similarity loss, while the reported result is EER computed on held-out merged test trials using cosine distance scoring. The cosine similarity loss and the cosine scoring metric are aligned, but the evaluation is a held-out error rate over 421,122 trials, not a re-statement of the training loss; no parameter fitted to the test set is renamed as a prediction. The emotion-aware masking step in Sec. 2.3 uses RMS energy of the speech signal (Eq. 3) to select mask positions; this is a physical acoustic feature that does not encode the test labels or the speaker labels, and the paper's claim is the testable empirical statement that masking these regions improves EER. The assumption that energy-dominant regions carry emotional content is an input premise, not a circular reduction of the result. CopyPaste augmentation constructs x' from x, but the contribution of such parallel data is measured by EER change in Table 4, which is an independent outcome rather than a construction-level identity. No load-bearing self-citation appears: references [18], [19], [20], [22], and [23] are external datasets, prior augmentation work, a standard loss, and prior speaker-encoder work by other authors; no uniqueness theorem or ansatz is imported from the authors' own prior work. The inconsistent ablation attributions in Sec. 4.2 and the conclusion (9.55%/2.48% vs 10.31%/2.98%) and the absence of confidence intervals for the small 4.70%-to-4.56% EM gain are statistical and correctness risks, not circularity. I therefore find no significant circularity and assign score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-set hyperparameters and two domain assumptions about emotion localization via speech energy and about CopyPaste preserving speaker identity. No new physical entities are introduced.

free parameters (5)
  • alpha (cosine loss weight) = 1
    Eq. (1) and Section 3.3; balances AAM-Softmax and cosine similarity loss, directly controls ERL contribution.
  • m (number of masks) = 2
    Section 3.3; number of randomly chosen mask centers in the emotion-dominant energy zone.
  • T (mask span in frames) = 7
    Section 3.3; each mask covers T frames, determining how much emotional content is removed.
  • energy zone thresholds = 0.2 and 0.5
    Section 2.3; partition normalized RMS energy into high, low, and noise zones; these boundaries are chosen by hand.
  • CopyPaste segment duration = 1 second
    Section 3.3; determines the length of each concatenated segment and thus the parallel sample construction.
assumptions (4)
  • domain assumption RMS energy is a reliable proxy for emotional intensity and location in speech.
    Invoked in Section 2.3: 'This energy, indicative of the speaker's vocal loudness and intensity, can serve as a reflection of their emotional state [21].'
  • domain assumption Masking high- or low-energy frames leaves the utterance emotion-free while preserving speaker identity.
    Section 2.3: 'Once the emotional aspects of the speech are masked, it is considered to be devoid of emotion.' No verification that speaker identity survives.
  • domain assumption CopyPaste concatenation of two same-speaker utterances produces a valid speech sample with the same speaker label and natural acoustic properties.
    Section 2.1: 'we argue that having a large amount of emotional parallel data can force the speaker encoders to focus more on the intrinsic characteristics...' No artifact check on concatenated samples.
  • domain assumption The four emotion categories in the Dusha Crowd subset and the constructed trial pairs represent emotional speaker variability adequately.
    Section 3.1 uses Crowd train/test with angry, positive, neutral, and sad only; results may not transfer to other emotions or languages.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Emotion-Invariant Speaker Representations for Speaker Verification." pith.science (2026). https://pith.science/paper/SCPLO6GH

@misc{pith2026250518498,
  author       = {Pith},
  title        = {Pith review of: Learning Emotion-Invariant Speaker Representations for Speaker Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCPLO6GH}},
  note         = {Machine review of arXiv:2505.18498}
}
read the original abstract

In recent years, the rapid progress in speaker verification (SV) technology has been driven by the extraction of speaker representations based on deep learning. However, such representations are still vulnerable to emotion variability. To address this issue, we propose multiple improvements to train speaker encoders to increase emotion robustness. Firstly, we utilize CopyPaste-based data augmentation to gather additional parallel data, which includes different emotional expressions from the same speaker. Secondly, we apply cosine similarity loss to restrict parallel sample pairs and minimize intra-class variation of speaker representations to reduce their correlation with emotional information. Finally, we use emotion-aware masking (EM) based on the speech signal energy on the input parallel samples to further strengthen the speaker representation and make it emotion-invariant. We conduct a comprehensive ablation study to demonstrate the effectiveness of these various components. Experimental results show that our proposed method achieves a relative 19.29\% drop in EER compared to the baseline system.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 27 canonical work pages

  1. [1]

    Currently, SV systems that use low- dimensional speaker representations extracted from deep learning- based speaker encoders have become the dominant approach in this field

    INTRODUCTION Speaker verification (SV) aims to determine if two speech samples come from the same person. Currently, SV systems that use low- dimensional speaker representations extracted from deep learning- based speaker encoders have become the dominant approach in this field. The performance of SV experienced a significant boost as researchers explored...

    work page

  2. [2]

    Learning Emotion-Invariant Speaker Representations for Speaker Verification

    LEARNING EMOTION-INV ARIANT SPEAKER REPRESENTA TIONS In this section, we present a scheme for learning emotion-invariant speaker representations in detail. A diagram of the scheme can be found in Fig. 1. For convenience, the data that fed into speaker encoder in the training stage are represented as{(xi, yi)}N i=1, where xi refers to input utterance i, yi...

    work page Pith review arXiv 2025

  3. [3]

    Datasets Our models are first pre-trained on the V oxCeleb [22] and then fine- tuned on the Dusha [18] dataset to evaluate the performance of SV

    EXPERIMENTAL SETUP 3.1. Datasets Our models are first pre-trained on the V oxCeleb [22] and then fine- tuned on the Dusha [18] dataset to evaluate the performance of SV . V oxCeleb is a dataset from YouTube with diverse speakers and speech in varied acoustic environments. It has two subsets: V ox- Celeb1 and V oxCeleb2, mainly containing English speech da...

    work page

  4. [4]

    Performance of the baseline system The performance of the pre-trained model on V oxCeleb1 is presented in Table 3, with the results of the ECAPA model referenced from [3]

    RESULTS 4.1. Performance of the baseline system The performance of the pre-trained model on V oxCeleb1 is presented in Table 3, with the results of the ECAPA model referenced from [3]. On the V oxCeleb1-O and V oxCeleb1-E test sets, the performance of ResNet34-TSDP is comparable to that of ECAPA. Additionally, a relative improvement of 6.5% is observed on...

    work page

  5. [5]

    We first verified that emotional utterances degrade SV performance, with cross-emotion test trails performed worse than same-emotion test trails

    CONCLUSIONS In this paper, we propose a scheme for learning speaker representa- tions that are invariant to emotions. We first verified that emotional utterances degrade SV performance, with cross-emotion test trails performed worse than same-emotion test trails. Then to alleviate this problem, we propose three improved techniques to train the speaker enc...

    work page

  6. [6]

    X-vectors: Robust dnn em- beddings for speaker recognition,

    David Snyder, Daniel Garcia-Romero, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur, “X-vectors: Robust dnn em- beddings for speaker recognition,” in 2018 IEEE interna- tional conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5329–5333

    work page 2018

  7. [7]

    But system description to vox- celeb speaker recognition challenge 2019,

    Hossein Zeinali, Shuai Wang, Anna Silnova, Pavel Mat ˇejka, and Old ˇrich Plchot, “But system description to vox- celeb speaker recognition challenge 2019,” arXiv preprint arXiv:1910.12592, 2019

    arXiv 2019

  8. [8]

    ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,

    Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck, “ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,” in Proc. Interspeech 2020, 2020, pp. 3830–3834

    work page 2020

Show all 30 references
  1. [9]

    MFA-Conformer: Multi-scale Feature Aggregation Con- former for Automatic Speaker Verification,

    Yang Zhang, Zhiqiang Lv, Haibin Wu, Shanshan Zhang, Pengfei Hu, Zhiyong Wu, Hung yi Lee, and Helen Meng, “MFA-Conformer: Multi-scale Feature Aggregation Con- former for Automatic Speaker Verification,” in Proc. Inter- speech 2022, 2022, pp. 306–310

    2022

  2. [10]

    Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,

    Xu Xiang, Shuai Wang, Houjun Huang, Yanmin Qian, and Kai Yu, “Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,” in 2019 Asia- Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEE...

    2019

  3. [11]

    In Defence of Metric Learn- ing for Speaker Recognition,

    Joon Son Chung, Jaesung Huh, Seongkyu Mun, Minjae Lee, Hee-Soo Heo, Soyeon Choe, Chiheon Ham, Sunghwan Jung, Bong-Jin Lee, and Icksang Han, “In Defence of Metric Learn- ing for Speaker Recognition,” inProc. Interspeech 2020, 2020, pp. 2977–2981

    2020

  4. [12]

    Multi-query multi-head attention pooling and inter-topk penalty for speaker verification,

    Miao Zhao, Yufeng Ma, Yiwei Ding, Yu Zheng, Min Liu, and Minqiang Xu, “Multi-query multi-head attention pooling and inter-topk penalty for speaker verification,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022,...

    2022

  5. [13]

    Explor- ing binary classification loss for speaker verification,

    Bing Han, Zhengyang Chen, and Yanmin Qian, “Explor- ing binary classification loss for speaker verification,” in ICASSP 2023-2023 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  6. [14]

    Nplda: A deep neural plda model for speaker verification,

    Shreyas Ramoji, Prashant Krishnan, and Sriram Ganapathy, “Nplda: A deep neural plda model for speaker verification,” in Proc. Odyssey 2020 The Speaker and Language Recognition Workshop, 2020, pp. 202–209

    2020

  7. [15]

    Scoring of Large-Margin Embeddings for Speaker Verification: Cosine or PLDA?,

    Qiongqiong Wang, Kong Aik Lee, and Tianchi Liu, “Scoring of Large-Margin Embeddings for Speaker Verification: Cosine or PLDA?,” in Proc. Interspeech 2022, 2022, pp. 600–604

    2022

  8. [16]

    Attention back-end for automatic speaker verification with multiple enrollment utterances,

    Chang Zeng, Xin Wang, Erica Cooper, Xiaoxiao Miao, and Junichi Yamagishi, “Attention back-end for automatic speaker verification with multiple enrollment utterances,” in ICASSP 2022-2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE,...

    2022

  9. [17]

    Prob- abilistic Spherical Discriminant Analysis: An Alternative to PLDA for length-normalized embeddings,

    Niko Brummer, Albert Swart, Ladislav Mosner, Anna Silnova, Oldrich Plchot, Themos Stafylakis, and Lukas Burget, “Prob- abilistic Spherical Discriminant Analysis: An Alternative to PLDA for length-normalized embeddings,” in Proc. Inter- speech 2022, 2022, pp. 1446–1450

    2022

  10. [18]

    A study of speaker verification performance with expressive speech,

    Srinivas Parthasarathy, Chunlei Zhang, John HL Hansen, and Carlos Busso, “A study of speaker verification performance with expressive speech,” in 2017 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2017, pp. 5540–5544

    2017

  11. [19]

    x-vectors meet emotions: A study on dependencies between emotion and speaker recognition,

    Raghavendra Pappagari, Tianzi Wang, Jesus Villalba, Nanxin Chen, and Najim Dehak, “x-vectors meet emotions: A study on dependencies between emotion and speaker recognition,” in ICASSP 2020-2020 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP)...

    2020

  12. [20]

    Emotion attribute projection for speaker recognition on emo- tional speech,

    Huanjun Bao, Ming-Xing Xu, and Thomas Fang Zheng, “Emotion attribute projection for speaker recognition on emo- tional speech,” in Proc. Interspeech 2007, 2007, pp. 758–761

    2007

  13. [21]

    Segment- Level Effects of Gender, Nationality and Emotion Informa- tion on Text-Independent Speaker Verification,

    Kai Li, Masato Akagi, Yibo Wu, and Jianwu Dang, “Segment- Level Effects of Gender, Nationality and Emotion Informa- tion on Text-Independent Speaker Verification,” in Proc. In- terspeech 2020, 2020, pp. 2987–2991

    2020

  14. [22]

    Instance-based Temporal Normalization for Speaker Verifi- cation,

    Thanathai Lertpetchpun and Ekapol Chuangsuwanich, “Instance-based Temporal Normalization for Speaker Verifi- cation,” in Proc. INTERSPEECH 2023, 2023, pp. 3172–3176

    2023

  15. [23]

    Hy- brid Dataset for Speech Emotion Recognition in Russian Lan- guage,

    Vladimir Kondratenko, Nikolay Karpov, Artem Sokolov, Nikita Savushkin, Oleg Kutuzov, and Fyodor Minkin, “Hy- brid Dataset for Speech Emotion Recognition in Russian Lan- guage,” in Proc. INTERSPEECH 2023, 2023, pp. 4548–4552

    2023

  16. [24]

    Copypaste: An augmentation method for speech emotion recognition,

    Raghavendra Pappagari, Jes ´us Villalba, Piotr ˙Zelasko, Lau- reano Moro-Velazquez, and Najim Dehak, “Copypaste: An augmentation method for speech emotion recognition,” in ICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEE...

    2021

  17. [25]

    Arcface: Additive angular margin loss for deep face recogni- tion,

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou, “Arcface: Additive angular margin loss for deep face recogni- tion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4690–4699

    2019

  18. [26]

    Sur- vey on speech emotion recognition: Features, classification schemes, and databases,

    Moataz El Ayadi, Mohamed S Kamel, and Fakhri Karray, “Sur- vey on speech emotion recognition: Features, classification schemes, and databases,” Pattern recognition, vol. 44, no. 3, pp. 572–587, 2011

    2011

  19. [27]

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

    Arsha Nagrani, Joon Son Chung, Weidi Xie, and Andrew Zis- serman, “V oxceleb: Large-scale speaker verification in the wild,” Computer Speech & Language , vol. 60, pp. 101027, 2020

    2020

  20. [28]

    Re- visiting the statistics pooling layer in deep speaker embedding learning,

    Shuai Wang, Yexin Yang, Yanmin Qian, and Kai Yu, “Re- visiting the statistics pooling layer in deep speaker embedding learning,” in 2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2021, pp. 1–5

    2021

  21. [29]

    Mu- san: A music, speech, and noise corpus,

    David Snyder, Guoguo Chen, and Daniel Povey, “Mu- san: A music, speech, and noise corpus,” arXiv preprint arXiv:1510.08484, 2015

    2015 arXiv

  22. [30]

    A study on data augmen- tation of reverberant speech for robust speech recognition,

    Tom Ko, Vijayaditya Peddinti, Daniel Povey, Michael L Seltzer, and Sanjeev Khudanpur, “A study on data augmen- tation of reverberant speech for robust speech recognition,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017, pp...

    2017

Pith tools

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