Pith. sign in

REVIEW 4 major objections 4 minor 49 references

Improving speaker verification robustness with synthetic emotional utterances

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Training a speaker verifier on CycleGAN-synthesized angry and happy utterances reduces equal error rate on emotional speech by up to 3.64% relative and shrinks the neutral-to-emotional gap from 1.30% to 0.94%.

desk verdict A plausible augmentation recipe with a load-bearing validity gap: the synthetic emotion isn't shown to be real, and the reporting hides the numbers needed to judge it. read the letter →

arxiv 2412.00319 v1 pith:6XCHLSE5 submitted 2024-11-30 cs.SD cs.AIeess.AS

Nikhil Kumar Koditala , Chelsea Jui-Ting Ju , Ruirui Li , Minho Jin , Aman Chadha , Andreas Stolcke This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords speakerverificationemotionalspeechCycleGANdataaugmentationvoiceconversiond-vectorembeddingGE2Elossequalerrorrate
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

Speaker verification systems stumble when a user is angry or happy instead of neutral, largely because emotional speech is scarce in training data. This paper tries to remove that scarcity by manufacturing emotion: two CycleGAN networks convert neutral utterances into angry and happy versions of the same speaker's voice, and those synthetic utterances are added to the training set of an LSTM d-vector verifier trained with GE2E loss. The authors report that augmented models consistently beat the baseline on emotional speech, with relative EER reductions up to 3.64%, and that the gap between neutral and emotional verification narrows from 1.30% to 0.94%. If the claim holds, emotion-robust speaker verification can be improved without collecting large amounts of real emotional speech.

What carries the argument

The carrying mechanism is a CycleGAN emotional voice converter trained without parallel data, applied to WORLD-vocoder spectral (MFCC) and prosody (F0) features. Two converters are trained, neutral-to-angry and neutral-to-happy, using a combined loss of adversarial, cycle-consistency, and identity terms; the synthetic emotional utterances are then spliced into the training set of a multi-layer LSTM d-vector speaker verifier trained with generalized end-to-end (GE2E) loss. The cycle-consistency loss is what is supposed to preserve speaker identity and linguistic content while the adversarial loss injects the target emotion.

What would settle it

Take the same neutral training utterances, run them through the CycleGAN pipeline without an emotion change (neutral-to-neutral), and add the same number of outputs to the SV training set. If neutral-to-neutral synthetic speech produces the same EER improvement on emotional test speech, the effect is not emotion-specific. Separately, have raters or an emotion classifier label the synthetic angry and happy utterances; if they are not reliably classified as angry or happy, the claim that the model learned emotion-invariant representations is unsupported.

Watch

Extended reading notes

Core claim

The paper's central claim is that CycleGAN-based emotional voice conversion is a working data augmentation strategy for speaker verification: synthetic angry and happy utterances, generated per speaker from neutral recordings, teach the verifier representations that generalize across emotional states. On a production-style LSTM d-vector speaker verifier, adding these synthetic utterances lowers equal error rate on emotional test utterances by 1.08% to 3.64% relative depending on the data mix, and the emotional-minus-neutral EER gap falls from 1.30% to 0.94%. The authors further claim the conversions preserve speaker identity: cosine similarity between neutral and synthetic angry embeddings (0.65) is higher than between neutral and authentic angry embeddings (0.51), and a t-SNE projection shows synthetic and authentic angry utterances overlapping for one speaker. The implied discovery is that the scarcity of labeled emotional speech, not the SV architecture, is the main obstacle to emotion robustness.

Load-bearing premise

The approach stands on the assumption that the CycleGAN-generated utterances are genuinely emotional (angry or happy) while preserving each speaker's identity, so the augmented training data teaches the verifier to ignore emotion rather than just adding noise or changing the data distribution.

Editorial extensions

If this is right

  • Adding synthetic happy utterances improves both neutral and emotional verification, with overall EER down 0.61% at 10 per speaker and 1.44% at 20 per speaker.
  • Adding synthetic angry utterances improves emotional EER, up to 21.61% relative for angry speech at 50 per speaker, but degrades neutral EER by up to -3.77%, so the mix must be tuned.
  • The neutral-to-emotional EER gap shrinks from 1.30% to 0.94% on the full training set when 15 angry and 15 happy synthetic utterances per speaker are added.
  • The gain is not just from more data: a 60-neutral plus 20-angry configuration does not beat 50-neutral plus 20-angry, and over-augmenting with angry data hurts.
  • Adding synthetic data keeps media-speech FAR below the 3% target, so this augmentation does not obviously open a spoofing hole.
  • The method works within a production-constrained LSTM baseline, not a state-of-the-art verifier, so the reported gains are a lower-bound demonstration of the augmentation's value.

Reading between the lines

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

  • The paper validates identity preservation but not emotion realization: there is no emotion classifier, human listening test, or acoustic emotion metric on the synthetic utterances, so the reported gains might come from any prosodic or spectral perturbation rather than from the specific target emotion.
  • The converters are trained on public-speaker emotional corpora and applied to internal speakers; a direct test on same-domain speakers would show whether the identity-preservation result transfers when source and target speakers overlap.
  • If the mechanism is genuinely emotion-invariant embeddings, the same augmentation should transfer to other verifier architectures such as x-vectors or ECAPA, which the paper does not test.
  • The neutral-to-synthetic angry cosine similarity being higher than the neutral-to-authentic angry value (0.65 vs 0.51) is ambiguous: it may indicate stronger identity preservation, or it may indicate the synthetic emotion is weaker than the real thing, and a supervised emotion-strength measure would disambiguate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 using CycleGAN-based emotional voice conversion as a data augmentation strategy for speaker verification (SV). Two CycleGAN networks convert neutral utterances into angry and happy utterances while attempting to preserve speaker identity; these synthetic emotional utterances are added to the training data of an LSTM d-vector SV model trained with the GE2E loss. Experiments on an internal, consent-collected dataset report relative EER improvements from incorporating synthetic emotional data, with up to 3.64% relative EER reduction on emotional utterances and a narrowing of the neutral-to-emotional performance gap from 1.30% to 0.94%. The paper also includes a spoofing-resilience experiment using media speech as a proxy.

Significance. If the central claim holds, the paper would provide a practical and broadly applicable data-augmentation recipe for improving SV robustness to emotional speech, a problem of real-world importance. The paper deserves credit for including a data-size control experiment in Section 4.2 to separate augmentation signal from corpus growth, and for consistently positive emotional-subset trends in Table 3 across several augmentation configurations. However, the load-bearing mechanism—that the CycleGAN outputs are genuinely emotional while preserving speaker identity—is not convincingly validated, and the main evidence table contains anomalies. The proprietary internal dataset and the reporting of only relative EER further limit reproducibility and comparability.

major comments (4)
  1. [§4.1, Table 2] The validation of the emotion-conversion mechanism is insufficient and the interpretation of Table 2 is ambiguous. The cosine similarity between neutral and synthetic angry utterances (0.65 ± 0.06) is higher than between neutral and authentic angry utterances (0.51 ± 0.10), which the authors read as evidence of speaker-identity preservation. However, this result is equally consistent with the synthetic utterances remaining closer to neutral, i.e., the conversion may not have fully realized the target emotion. The t-SNE plot in Figure 1 covers a single speaker and is qualitative. Because the abstract and conclusion attribute the EER gains specifically to synthetic emotional utterances, the paper needs a quantitative check that the generated samples are perceived or acoustically classified as the intended emotion (e.g., an emotion classifier, prosodic feature comparison, or listening test). Without such evidence, the central mechanism is unestablished.
  2. [Table 4] Table 4 contains suspiciously identical values across rows: the Happy column is 6.16% and the Angry column is 1.87% in all three augmentation configurations, and the Sad/Calm values repeat (0.90 appears twice, and -0.51/-1.02 appear in patterns) despite different amounts of synthetic data. This is inconsistent with independently retrained models and suggests copy-paste errors or a different aggregation than the table caption implies. Since Table 4 supports the headline claim that the neutral-to-emotional gap narrows from 1.30% to 0.94%, these anomalies undermine the evidential basis of that claim and must be resolved.
  3. [§4.2, Tables 3 and 4] The paper reports only relative EER changes and does not provide absolute EER values, confidence intervals, or significance tests. In particular, the claim of a 3.64% relative EER reduction for the 15 angry + 15 happy configuration is presented without an estimate of variability, so the reader cannot judge whether the differences between configurations (e.g., 1.08% vs. 3.64%) are meaningful. Given that the authors state they are restricted from disclosing absolute EER, the reporting of paired error bars, bootstrap intervals, or at least a statistical comparison across configurations is necessary to support the quantitative strength of the central claim.
  4. [§4.2.1] The spoofing-resilience analysis does not actually compare models trained with and without synthetic data. The claim that adding synthetic utterances did not negatively affect media-speech FAR is supported only by the statement that FAR remained below a 3% target, with no baseline FAR on the same media-speech evaluation set. Without a no-synthetic control, the experiment cannot establish the absence of an adverse effect, which is the stated conclusion of this subsection.
minor comments (4)
  1. [Equation (1)] The adversarial loss in Equation (1) is likely missing a logarithm in the first expectation: it is written as Ey[DY(y)] rather than Ey[log DY(y)], which is the standard form for the non-saturating GAN loss described in the text.
  2. [§3.2.1] The sentence "During the training phase of the CycleGAN network, the input consists of source (neutral) and target (emotional) utterances from same speaker" is confusing because Section 3.1 states that the CycleGAN training data come from public datasets with a limited number of speakers, while the internal data are used for the SV model. Please clarify which speaker pool is used for CycleGAN training and how the same-speaker pairing is obtained in a non-parallel setting.
  3. [Throughout] There are several typographical errors that should be corrected: "gaussain" (Section 1), "Mel-spectogram" (Section 2.1), "oppurtunity" (Section 3.2.1), "consine" (Section 3.2.1), and inconsistent spacing in "WORLD V ocoder" and "V oice conversion" throughout the references.
  4. [§4.2] The data-size control argument would be stronger if the compared configurations were matched in total utterance count; the 60 neutral + 20 angry configuration adds 10 real neutral utterances plus 20 synthetic ones, so it does not isolate data size from synthetic-data proportion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EER gains are measured on held-out authentic emotional utterances, and the synthetic data are produced from external public emotion corpora, not from the evaluation set.

full rationale

The claimed derivation chain is empirical rather than definitional: (1) CycleGAN emotional voice converters are trained on public corpora (ESD, EmoV, Ravdess) to map neutral utterances to angry or happy while preserving speaker identity; (2) the trained converters are applied to internal training speakers' neutral utterances; (3) LSTM/GE2E speaker verification models are trained on neutral plus synthetic emotional utterances; and (4) EER is measured on a held-out internal evaluation set containing authentic emotional utterances. No step defines its output in terms of its input. The synthetic data are generated from external emotion datasets and applied to different internal speakers, so the reported relative EER improvements (up to 3.64%) are not fitted predictions of the evaluation labels. The data-size control experiments in Section 4.2 attempt to separate the augmentation signal from mere dataset growth, which further supports that the result is not forced by construction. The only self-citation is [32], used to justify the 40-dimensional mel-spectrogram input to the LSTM; it is not load-bearing for the central claim, and the EER improvements are independently measured. The weak validation of whether synthetic utterances truly realize the target emotion — Table 2's cosine similarities and the single-speaker t-SNE plot — is a correctness or validity concern, not a circularity. The paper's own limitations (no absolute EER disclosure, baseline not SOTA, ongoing spoofing considerations) do not create a circular derivation. Therefore, no circular step is identifiable.

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

The central claim depends mainly on the quality and validity of CycleGAN-generated emotional speech, the reliability of emotion labels, and the chosen hyperparameters. No new physical or theoretical entities are introduced.

free parameters (5)
  • CycleGAN loss weights lambda_CY and lambda_ID = 10 and 5 respectively
    Hand-chosen weights in Equation (4) control the adversarial/cycle/identity trade-off and therefore the quality of synthetic emotional speech.
  • CycleGAN discriminator head-start k = Not specified
    Number of initial training steps with the discriminator frozen is chosen by hand to stabilize GAN training (Section 3.2.1).
  • CycleGAN training set size per emotion = 2000 utterances
    Authors sampled 2000 utterances per emotion from public datasets to train each conversion network (Section 3.2.1).
  • SV augmentation counts per speaker = 10/20/50 angry or happy synthetic utterances
    The experimental configurations in Tables 3 and 4 sweep the amount of synthetic data; conclusions about improvement depend on these chosen counts.
  • SV model hyperparameters (N, M, learning rate, decay) = N=32, M=5, eta=1e-6, beta=0.98
    Standard GE2E/LSTM training hyperparameters chosen by hand (Section 3.2.2).
assumptions (5)
  • domain assumption The public emotional datasets (Emotional Speech Dataset, EmoV, Ravdess) provide valid angry and happy speech for CycleGAN training.
    The conversion networks are trained on these datasets (Section 3.1); if their labels or audio quality are not representative, the synthetic utterances may not carry the intended emotion.
  • domain assumption WORLD vocoder F0 and MFCC features are sufficient to represent both emotion and speaker identity for conversion and verification.
    Both the CycleGAN conversion and the SV model operate on these features (Sections 2.1 and 2.2), so any information lost in feature extraction limits the method.
  • domain assumption Cycle-consistency and identity losses preserve linguistic content and speaker identity during conversion.
    The paper relies on the CycleGAN objective (Equations 2 and 3) to keep speaker identity, citing [31]; this is inherited from prior work rather than verified here.
  • domain assumption Majority-vote labels from at least three annotators on anonymized internal utterances are reliable ground truth for emotion and speaker.
    The internal dataset labels are produced this way (Section 3.1) and are used to train and evaluate the SV models.
  • domain assumption The LSTM d-vector with GE2E loss is an adequate SV baseline for measuring augmentation benefit.
    The authors acknowledge the baseline is not SOTA (Section 2.1), so improvements may not transfer to modern SV systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving speaker verification robustness with synthetic emotional utterances." pith.science (2026). https://pith.science/paper/6XCHLSE5

@misc{pith2026241200319,
  author       = {Pith},
  title        = {Pith review of: Improving speaker verification robustness with synthetic emotional utterances},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6XCHLSE5}},
  note         = {Machine review of arXiv:2412.00319}
}
read the original abstract

A speaker verification (SV) system offers an authentication service designed to confirm whether a given speech sample originates from a specific speaker. This technology has paved the way for various personalized applications that cater to individual preferences. A noteworthy challenge faced by SV systems is their ability to perform consistently across a range of emotional spectra. Most existing models exhibit high error rates when dealing with emotional utterances compared to neutral ones. Consequently, this phenomenon often leads to missing out on speech of interest. This issue primarily stems from the limited availability of labeled emotional speech data, impeding the development of robust speaker representations that encompass diverse emotional states. To address this concern, we propose a novel approach employing the CycleGAN framework to serve as a data augmentation method. This technique synthesizes emotional speech segments for each specific speaker while preserving the unique vocal identity. Our experimental findings underscore the effectiveness of incorporating synthetic emotional data into the training process. The models trained using this augmented dataset consistently outperform the baseline models on the task of verifying speakers in emotional speech scenarios, reducing equal error rate by as much as 3.64% relative.

Figures

Figures reproduced from arXiv: 2412.00319 by the authors.

Figure 1
Figure 1. t-SNE plot of utterances from a single speaker in neu [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 48 canonical work pages

  1. [1]

    Improving speaker verification robustness with synthetic emotional utterances

    Introduction Speaker verification (SV) systems are designed to discern whether a given audio stream belongs to the speaker who has been previously enrolled, based on various acoustic character- istics of the speaker such as pitch, tone and intonation [1]. SV system typically consists of two phases – an enrollment phase and a verification phase. During the...

    work page Pith review arXiv 2024

  2. [2]

    Methods 2.1. Speaker Verification Model The SV model is a multi-layer LSTM network which uses a 40-dimensional Mel-spectogram as input [32] and outputs an n-dimensional d-vector or deep vector, which is the average of activations derived from the last hidden layer of the LSTM [33]. d-vector helps in speaker verification by encapsulating the neu- ral embed...

    work page

  3. [3]

    Experiment Settings 3.1. Data To train the CycleGAN model, we used non-parallel emotional utterances drawn from three distinct open source datasets: Emo- tional Speech Dataset [35], EmoV [36], and Ravdess [37]. Emotional Speech dataset consists of 350 utterances spoken by 10 English speakers in different emotional states. EmoV dataset is a collection of u...

    work page

  4. [4]

    50 neutral + 10 angry

    Results 4.1. Emotion Conversion As discussed in Section 3.2.1, we trained two different Cycle- GAN networks to transform neutral-toned utterances into those expressing anger and happiness. To assess the efficacy of this emotional modulation, we conducted a comprehensive evalu- ation. Firstly, we utilized the cycleGAN model to synthesize angry utterances f...

    work page

  5. [5]

    Conclusion In this work, we have introduced a pioneering approach uti- lizing the CycleGAN framework to significantly improve SV systems through the innovative use of data augmentation. This technique uniquely maintains the individual vocal traits of speakers while generating synthetic emotional speech samples, thus enhancing the training process and impr...

    work page

  6. [6]

    Acknowledgments We would like to thank Brecht Desplanques for his contribution in reviewing this work and providing valuable feedback

    work page

  7. [7]

    Hamid Aghajan, Juan Augusto Wrede, and R.L-C Del- gado, Human-Centric Interfaces for Ambient Intelligence, 01 2010

    work page 2010

  8. [8]

    A tutorial on text-independent speaker ver- ification,

    Bimbot Fr ´ed´eric, Jean-Franc ¸ois Bonastre, Corinne Fredouille, Guillaume Gravier, and Ivan Magrin- Chagnolleau, “A tutorial on text-independent speaker ver- ification,” EURASIP Journal on Advances in Signal Pro- cessing, 2004

    work page 2004

Show all 49 references
  1. [9]

    Speaker indexing in large audio databases using anchor models,

    D.E. Sturim, D.A. Reynolds, E. Singer, and J.P. Camp- bell, “Speaker indexing in large audio databases using anchor models,” in 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), 2001, vol. 1, pp. 429–432 vol.1

    2001

  2. [10]

    Speaker verification using support vector machines and high-level features,

    William M. Campbell, Joseph P. Campbell, Terry P. Glea- son, Douglas A. Reynolds, and Wade Shen, “Speaker verification using support vector machines and high-level features,” IEEE Transactions on Audio, Speech, and Lan- guage Processing, vol. 15, no. 7, pp. 2085–2094, 2007

    2007

  3. [11]

    Rawnet: Advanced end-to-end deep neural network using raw waveforms for text-independent speaker verification,

    Jee weon Jung, Hee-Soo Heo, Ju ho Kim, Hye jin Shim, and Ha jin Yu, “Rawnet: Advanced end-to-end deep neural network using raw waveforms for text-independent speaker verification,” in Interspeech, 2019

    2019

  4. [12]

    Speaker diariza- tion with lstm,

    Quan Wang, Carlton Downey, Li Wan, Philip Andrew Mansfield, and Ignacio Lopz Moreno, “Speaker diariza- tion with lstm,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 5239–5243

    2018

  5. [13]

    An overview of automatic speaker ver- ification system,

    Ravika Naika, “An overview of automatic speaker ver- ification system,” in Intelligent Computing and Infor- mation and Communication , Subhash Bhalla, Vikrant Bhateja, Anjali A. Chandavale, Anil S. Hiwale, and Suresh Chandra Satapathy, Eds., Singapore, 2018, pp. 603–610, Sprin...

    2018

  6. [14]

    Ap- plications of speaker recognition,

    Nilu Singh, Prof. Raees Khan, and Raj Shree Pandey, “Ap- plications of speaker recognition,” Procedia Engineering, vol. 38, pp. 3122–3126, 12 2012

    2012

  7. [15]

    Emotional speaker identification using a novel capsule nets model,

    Ali Bou Nassif, Ismail Shahin, Ashraf Elnagar, Divya Ve- layudhan, Adi Alhudhaif, and Kemal Polat, “Emotional speaker identification using a novel capsule nets model,” Expert Systems with Applications , vol. 193, pp. 116469, 2022

    2022

  8. [16]

    Wearable emotion recognition using heart rate data from a smart bracelet,

    Lin Shu, Yang Yu, Wenzhuo Chen, Haoqiang Hua, Qin Li, Jianxiu Jin, and Xiangmin Xu, “Wearable emotion recognition using heart rate data from a smart bracelet,” Sensors (Basel, Switzerland), vol. 20, 2020

    2020

  9. [17]

    Physiologi- cal changes associated with emotion,

    Purves D, Augustine GJ, and Fitzpatrick D, “Physiologi- cal changes associated with emotion,” Neuroscience. 2nd edition, 2001

    2001

  10. [18]

    X-vectors meet emo- tions: A study on dependencies between emotion and speaker recognition,

    Raghavendra Pappagari, Tianzi Wang, Jes ´us Villalba, Nanxin Chen, and Najim Dehak, “X-vectors meet emo- tions: A study on dependencies between emotion and speaker recognition,” in Proc. IEEE ICASSP , 05 2020, pp. 7169–7173

    2020

  11. [19]

    Achieving fair speech emotion recognition via perceptual fairness,

    Woan-Shiuan Chien and Chi-Chun Lee, “Achieving fair speech emotion recognition via perceptual fairness,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  12. [20]

    Gender-related dif- ferences in the production and perception of emotion,

    M.G.J. Swerts and E.J. Krahmer, “Gender-related dif- ferences in the production and perception of emotion,” in Proceedings of the international conference on spoken language processing (Interspeech 2008) . 2008, pp. 334– 337, ISCA

    2008

  13. [21]

    Demographic fair- ness in biometric systems: What do the experts say?,

    Christian Rathgeb, Pawel Drozdowski, Dinusha C. Frings, Naser Damer, and Christoph Busch, “Demographic fair- ness in biometric systems: What do the experts say?,” IEEE Technology and Society Magazine , vol. 41, no. 4, pp. 71–82, 2022

    2022

  14. [22]

    Effects of emotional valence and arousal on the voice perception network,

    Patricia E. G. Bestelmeyer, Sonja A. Kotz, and Pascal Be- lin, “Effects of emotional valence and arousal on the voice perception network,” in Soc Cogn Affect Neurosci, 2017, pp. 1351–1358

    2017

  15. [23]

    Unveiling the acoustic properties that describe the valence dimension,

    Carlos Busso and Tauhidur Rahman, “Unveiling the acoustic properties that describe the valence dimension,” in Proc. Interspeech, 09 2012, pp. 1179–1182

    2012

  16. [24]

    Generalized end-to-end loss for speaker verification,

    Li Wan, Quan Wang, Alan Papir, and Ignacio Moreno, “Generalized end-to-end loss for speaker verification,” in Proc. IEEE ICASSP, 04 2018, pp. 4879–4883

    2018

  17. [25]

    V oice conversion based on maximum-likelihood estimation of spectral parameter trajectory,

    Tomoki Toda, Alan Black, and Keiichi Tokuda, “V oice conversion based on maximum-likelihood estimation of spectral parameter trajectory,” Audio, Speech, and Lan- guage Processing, IEEE Transactions on , vol. 15, pp. 2222 – 2235, 12 2007

    2007

  18. [26]

    Algorithms for non- negative matrix factorization,

    Daniel Lee and H. Sebastian Seung, “Algorithms for non- negative matrix factorization,” in Advances in Neural In- formation Processing Systems, T. Leen, T. Dietterich, and V . Tresp, Eds. 2000, vol. 13, MIT Press

    2000

  19. [27]

    Exemplar-based sparse representation with residual compensation for voice conversion,

    Zhizheng Wu, Tuomas Virtanen, Eng Chng, and Haizhou Li, “Exemplar-based sparse representation with residual compensation for voice conversion,” Audio, Speech, and Language Processing, IEEE/ACM Transactions on , vol. 22, pp. 1506–1521, 10 2014

    2014

  20. [28]

    V oice conversion using deep neural networks with layer-wise generative training,

    Ling-Hui Chen, Zhen-Hua Ling, Li-Juan Liu, and Lirong Dai, “V oice conversion using deep neural networks with layer-wise generative training,” Audio, Speech, and Lan- guage Processing, IEEE/ACM Transactions on , vol. 22, pp. 1859–1872, 12 2014

    2014

  21. [29]

    Spectral mapping using artificial neu- ral networks for voice conversion,

    Srinivas Desai, Alan Black, B. Yegnanarayana, and Kishore Prahallad, “Spectral mapping using artificial neu- ral networks for voice conversion,” Audio, Speech, and Language Processing, IEEE Transactions on, vol. 18, pp. 954 – 964, 08 2010

    2010

  22. [30]

    V oice conversion in high-order eigen space using deep belief nets,

    Toru Nakashika, Ryoichi Takashima, Tetsuya Takiguchi, and Yasuo Ariki, “V oice conversion in high-order eigen space using deep belief nets,” in Proc. Interspeech, 08 2013, pp. 369–372

    2013

  23. [31]

    On the use of i-vectors and average voice model for voice conversion without parallel data,

    Jie Wu, Zhizheng Wu, and Lei Xie, “On the use of i-vectors and average voice model for voice conversion without parallel data,” in 2016 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA), 2016, pp. 1–6

    2016

  24. [32]

    Non-parallel voice conversion using variational autoencoders conditioned by phonetic poste- riorgrams and d-vectors,

    Yuki Saito, Yusuke Ijima, Kyosuke Nishida, and Shinno- suke Takamichi, “Non-parallel voice conversion using variational autoencoders conditioned by phonetic poste- riorgrams and d-vectors,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASS...

    2018

  25. [33]

    V oice conversion from non-parallel corpora using variational auto-encoder,

    Chin-Cheng Hsu, Hsin-Te Hwang, Yi-Chiao Wu, Yu Tsao, and Hsin-Min Wang, “V oice conversion from non-parallel corpora using variational auto-encoder,” 2016

    2016

  26. [34]

    V oice Conversion from Un- aligned Corpora using Variational Autoencoding Wasser- stein Generative Adversarial Networks,

    Chin-Cheng Hsu, Hsin-Te Hwang, Yi-Chiao Wu, Yu Tsao, and Hsin-Min Wang, “V oice Conversion from Un- aligned Corpora using Variational Autoencoding Wasser- stein Generative Adversarial Networks,”arXiv e-prints, p. arXiv:1704.00849, Apr. 2017

    2017 arXiv

  27. [35]

    On the study of generative adversarial net- works for cross-lingual voice conversion,

    Berrak Sisman, Mingyang Zhang, Minghui Dong, and Haizhou Li, “On the study of generative adversarial net- works for cross-lingual voice conversion,” in 2019 IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU), 2019, pp. 144–151

    2019

  28. [36]

    Cyclegan-vc: Non-parallel voice conversion using cycle-consistent ad- versarial networks,

    Takuhiro Kaneko and Hirokazu Kameoka, “Cyclegan-vc: Non-parallel voice conversion using cycle-consistent ad- versarial networks,” in 2018 26th European Signal Pro- cessing Conference (EUSIPCO), 2018, pp. 2100–2104

    2018

  29. [37]

    Transform- ing spectrum and prosody for emotional voice conversion with non-parallel training data,

    Kun Zhou, Berrak Sisman, and Haizhou Li, “Transform- ing spectrum and prosody for emotional voice conversion with non-parallel training data,” in Odyssey Speaker and Language Recognition Workshop, 05 2020

    2020

  30. [38]

    Fusion of embeddings net- works for robust combination of text dependent and in- dependent speaker recognition,

    Ruirui Li, Chelsea Ju, Zeya Chen, Hongda Mao, Oguz Elibol, and Andreas Stolcke, “Fusion of embeddings net- works for robust combination of text dependent and in- dependent speaker recognition,” in Proc. Interspeech, 08 2021, pp. 4593–4597

    2021

  31. [39]

    Deep neural networks for small footprint text-dependent speaker verifi- cation,

    Ehsan Variani, Xin Lei, Erik McDermott, Ignacio Lopez Moreno, and Javier Gonzalez-Dominguez, “Deep neural networks for small footprint text-dependent speaker verifi- cation,” in 2014 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2014, pp...

    2014

  32. [40]

    World: A vocoder-based high-quality speech synthesis system for real-time applications,

    Masanori MORISE, Fumiya YOKOMORI, and Kenji Ozawa, “World: A vocoder-based high-quality speech synthesis system for real-time applications,”IEICE Trans- actions on Information and Systems , vol. E99.D, pp. 1877–1884, 07 2016

    2016

  33. [41]

    Seen and unseen emotional style transfer for voice con- version with a new emotional speech dataset,

    Kun Zhou, Berrak Sisman, Rui Liu, and Haizhou Li, “Seen and unseen emotional style transfer for voice con- version with a new emotional speech dataset,” in Proc. IEEE ICASSP, 02 2021

    2021

  34. [42]

    The emotional voices database: Towards controlling the emotion dimension in voice generation systems,

    Adaeze Adigwe, No ´e Tits, Kevin El Haddad, Sarah Os- tadabbas, and Thierry Dutoit, “The emotional voices database: Towards controlling the emotion dimension in voice generation systems,” 06 2018

    2018

  35. [43]

    The ryer- son audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,

    Steven R. Livingstone and Frank A. Russo, “The ryer- son audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,” PLOS ONE, vol. 13, no. 5, pp. 1–35, 05 2018

    2018

  36. [44]

    Lutz Prechelt, Early Stopping — But When? , Springer Berlin Heidelberg, Berlin, Heidelberg, 2012

    2012

  37. [45]

    Towards principled methods for training generative adversarial networks,

    Martin Arjovsky and Leon Bottou, “Towards principled methods for training generative adversarial networks,” in International Conference on Learning Representations , 2017

    2017

  38. [46]

    Generative adversarial nets,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversarial nets,” in Ad- vances in neural information processing systems , 2014, pp. 2672–2680

    2014

  39. [47]

    Adam: A method for stochastic optimization,

    Diederik P. Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” in 3rd International Con- ference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Pro- ceedings, Yoshua Bengio and Yann LeCun, Eds., 2015

    2015

  40. [48]

    Visualiz- ing data using t-sne,

    Laurens van der Maaten and Geoffrey Hinton, “Visualiz- ing data using t-sne,” Journal of Machine Learning Re- search, vol. 9, no. 86, pp. 2579–2605, 2008

    2008

  41. [49]

    Identifying source speakers for voice conversion based spoofing attacks on speaker verification systems,

    Danwei Cai, Zexin Cai, and Ming Li, “Identifying source speakers for voice conversion based spoofing attacks on speaker verification systems,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.