Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Beyond One-Size-Fits-All: Personalized and Culturally Adaptive Emotional TTS via Interactive Optimization of Individual Emotion Perception Spaces

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that personalizing an emotional TTS's arousal–valence coordinates to each listener, through a short interactive genetic search, yields speech that listeners prefer over one-size-fits-all baselines, with personalized prefer

desk verdict Clever IGA-based emotional TTS personalization, but the main preference test is confounded and the objective metric is circular; needs a deconfounded evaluation. read the letter →

arxiv 2608.00998 v1 pith:SHBWDLQB submitted 2026-08-02 eess.AS

Wangzixi Zhou , Bagus Tris Atmaja , Sakriani Sakti This is my paper
classification eess.AS
keywords emotionaltext-to-speechinteractivegeneticalgorithmpersonalizationculturaladaptationarousal-valencemodelhuman-in-the-loopoptimizationspeechemotionrecognitioncross-culturalperception
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

Most emotional text-to-speech systems assume that one averaged arousal–valence coordinate describes how everyone hears a given emotion. This paper argues that this one-size-fits-all assumption is false, and that both individuals and cultural groups map emotions onto the A–V plane differently. It proposes a post-training personalization layer: an Interactive Genetic Algorithm that, in about three rounds of listening and picking, finds the A–V coordinate that matches a listener's perception, without retraining the acoustic model. Evaluations with Japanese, Chinese, and Indonesian listeners support the claim that personalized coordinates are preferred over a U.S.-averaged baseline 76% of the time, while culture-specific average coordinates are preferred 64.8–69.8% of the time. If right, this makes emotional adaptation in conversational AI cheap: a few clicks per user instead of large preference datasets or model retraining.

What carries the argument

The carrying mechanism is the closed loop between user choice and coordinate mutation. The Interactive Genetic Algorithm searches Russell's arousal–valence plane: each new A–V candidate $e_k$ is a random weighted average of the user-selected parents plus a uniform perturbation $\Delta \sim U(-M_g,M_g)$, with mutation strength $M_g$ decaying multiplicatively to a floor. What makes the loop cheap is that the high-dimensional acoustic representation is fixed; only the two scalar coordinates move. The companion Emotion Controller converts a chosen coordinate into prosody: a Gaussian Fourier feature map $\gamma(e)$ feeds an MLP that predicts a high-dimensional emotion feature vector from a pretra

What would settle it

Take a new group of listeners who never took part in the interactive search. For each emotion and each IGA-trained participant, synthesize speech from the participant's personalized A–V coordinate and from the U.S. average baseline, then have the new listeners choose which is the better emotional match. If the preference rate for personalized coordinates over the baseline is not significantly above 50%, the 76% result is at least partly an artifact of the same listeners re-encountering samples from their own search rather than a general improvement in perceptual alignment.

Watch

Extended reading notes

Core claim

The paper's central claim is that the same low-dimensional arousal–valence (A–V) space that existing emotional TTS systems use for coarse control can be reshaped, per listener, into a personalized emotion perception space that yields more perceptually aligned speech. Personalization is done by an Interactive Genetic Algorithm (IGA): the system synthesizes a population of candidate utterances from candidate A–V coordinates, the listener picks the sample that best matches the target emotion, and the chosen coordinates become parents for the next generation through weighted recombination and decaying uniform mutation. The authors report convergence within about three rounds, after which each pa

Load-bearing premise

The 76% personalization result assumes that the A/B preference of the same 15 participants who steered the IGA search is an unbiased measure of perceptual alignment rather than a familiarity or self-consistency effect from hearing samples produced by their own chosen coordinates.

Editorial extensions

If this is right

  • Per-user emotional control requires only about three rounds of listening and selecting, so the same backbone model can be adapted to many listeners without retraining or large preference datasets.
  • Culture-specific A–V averages computed from ten participants per group transfer some personalization benefit to new listeners from the same culture, suggesting a middle ground between fully individual and universal mappings.
  • Cross-cultural ranking shows listeners from all three groups prefer their own culture's A–V mapping (65–70%), implying that cultural context is a useful prior for emotional TTS.
  • Replacing a simple A–V embedding with the emotion controller improves naturalness (MOS 3.37→3.75), intelligibility (WER 21→17), and emotional similarity (CCC arousal 0.60→0.84, valence 0.64→0.77).
  • Because only the two A–V coordinates change during personalization, the framework is a post-training layer that can sit on top of any TTS exposing continuous emotion control.

Reading between the lines

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

  • My inference: if the 76% preference persists when naive listeners judge the same personalized coordinates, the IGA loop is not just finding a sample the user happened to pick; it is recovering a stable point in the user's emotion space, making the procedure a reusable zero-retraining personalization primitive for conversational agents.
  • My inference: the cultural clustering in the A–V space suggests a 'culture bias' vector could be learned from group-level differences and applied to a universal A–V mapping; the paper reports group averages but does not test such a vectorized transfer, so this is an extension.
  • My inference: because the dataset A–V values are estimated by one SER model rather than human annotation, the CCC gains may partly reflect that model's internal label conventions; an independent human-annotation check would tell whether the objective gains correspond to perceived emotion shifts.
  • My inference: the same interactive-optimization pattern could be applied to any low-dimensional subjective control in synthetic speech—politeness, formality, perceived age—by substituting a different control space for A–V and keeping the IGA loop unchanged.
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 / 4 minor

Summary. This paper proposes a framework for personalized and culturally adaptive emotional text-to-speech. It uses an Interactive Genetic Algorithm (IGA) to adapt arousal–valence (A–V) coordinates for each listener while keeping the acoustic backbone (Grad-TTS) fixed. An Emotion Controller maps A–V coordinates to SER-derived acoustic features. The authors report improved MOS/WER/CCC over a Grad-TTS baseline, show that personalized A–V coordinates differ from averaged U.S.-based values across Chinese, Indonesian, and Japanese listeners, and report A/B preference rates of 76% for personalized speech and 64.8–69.8% for culture-specific speech over the U.S. baseline. The central claim is that personalization and cultural adaptation improve emotional alignment beyond a one-size-fits-all model.

Significance. The idea of adapting a low-dimensional emotion representation with minimal user feedback is timely and practically relevant, and the IGA instantiation is transparent and reproducible. If the empirical claims were validated with independent listeners and a non-circular objective metric, the paper would make a useful contribution to controllable and personalized TTS. The current evidence, however, does not yet support the headline claim: the personalization A/B test is confounded by using the same participants whose selections defined the coordinates, and the CCC metric is computed with the same SER model that provides the training targets. The cultural comparisons are underpowered and lack inferential statistics. Strengths of the paper include the explicit algorithm, the cross-cultural participant pool, and the public demo page; these make the methodological issues particularly worth addressing in revision.

major comments (3)
  1. [§5.3 (personalization A/B)] The personalization A/B test is confounded. The same 15 participants who completed the IGA loop in §4.2 later compared speech generated from their personalized A–V coordinate against the U.S.-based baseline. Their personalized coordinate is the endpoint of their own interactive search, and the preferred sample may be one they explicitly selected during the IGA. The 76% preference rate can thus reflect familiarity, self-consistency, or demand characteristics rather than improved perceptual alignment for new listeners. No hold-out listener condition, significance test, or confidence interval is reported; with n=15, a 76% result has a wide binomial CI. This confound directly undermines the paper's central claim that IGA personalization yields more perceptually aligned emotional speech.
  2. [§4.3 and Table 1 (CCC metric)] The CCC metric is circularly defined. The same pre-trained SER model [26] is used to (i) assign A–V labels to the training dataset (§4.1), (ii) provide the latent training targets for the Emotion Feature Predictor (§3.2), and (iii) predict A–V values for the synthesized and reference speech in the evaluation (§4.3). The high CCC values (0.84 arousal, 0.77 valence) therefore largely measure how well the generator reproduces the SER model's own representation, not an independent degree of perceived emotional similarity. An evaluation with human A–V ratings or with a different, held-out SER model is needed before claiming 'closer alignment between synthesized and reference emotional characteristics'.
  3. [§5.3 (culture-specific and cross-cultural)] The culture-specific A/B evaluation avoids the self-selection confound but is severely underpowered and lacks inferential statistics. The text says 15 new participants were used, implying 5 per culture, and the reported preference rates of 64.8–69.8% are given without significance tests or confidence intervals. The cross-cultural ranking results (65%, 67%, 70%) have the same problem. In addition, the participant accounting is unclear: §4.2 says 30 participants completed personalization, while §5.3 says only 15 of those took the personalization A/B test; the relationship between the personalization pool and the A/B pools is not explained. These issues make the cross-cultural claims suggestive at best.
minor comments (4)
  1. [§5.1] The '23.5% relative reduction in WER' is incorrect: (21−17)/21 ≈ 19.0%. The calculation should use the baseline value in the denominator.
  2. [§3.2] The hyperparameters D_f and D_feat are not specified. Please provide these values to make the Emotion Feature Predictor reproducible.
  3. [Algorithm 1] The stopping condition 'until user satisfaction' is not formally defined. Please report the number of IGA generations typically needed for convergence.
  4. [§5.3] The description of the A/B tests as 'blind' lacks procedural details: randomization, counterbalancing of A/B order, and whether participants could replay samples. These details matter for interpreting preference results.

Circularity Check

2 steps flagged · score 6.0 of 10

Personalization A/B test is evaluated on the same participants whose IGA clicks defined the personalized coordinates, making the 76% preference partly forced by construction; the CCC metric is also self-referential to the SER model used for training targets.

  1. fitted input called prediction [Section 4.2 / Section 5.3, Figure 3 (left bars)]
    "Participants who completed the IGA personalization process performed blind A/B comparisons between speech generated using their personalized A–V values and speech generated using the U.S.-based reference A–V baseline. As shown in Fig. 3 (left bars), participants were unaware of which sample corresponded to their personalized mapping, yet personalized speech achieved a preference rate of 76%."

    The personalized A–V coordinate is the endpoint of the same participant's IGA selections (Algorithm 1). Re-presenting that coordinate's synthesized speech to the same listener and asking whether it beats a fixed baseline is a manipulation check on the optimization, not an independent test of perceptual alignment. If the IGA converged, the listener should prefer their own endpoint by construction; the 76% preference is therefore partly an artifact of fitting each coordinate to that listener's prior choices and cannot by itself support generalization to new listeners. Blindness does not remove familiarity or self-selection confounds.

  2. other [Section 3.2 / Section 4.1 / Section 4.3 / Section 5.1]
    "Training targets h_gt are obtained from a pre-trained Speech Emotion Recognition (SER) model [26]... The model is trained using an L1 loss between h_pred and h_gt. ... CCC to measure emotional similarity by comparing SER-predicted A–V values between generated and reference speech."

    The Emotion Feature Predictor is trained to regress the SER model's own internal emotion features (h_gt), and the CCC evaluation then uses the same SER model to predict A–V values from both generated and reference speech. High CCC therefore partly measures how well the generator reproduces the SER model's feature space—exactly the training objective—rather than an independent, listener-based measure of emotional similarity. The metric is self-referential to the model that produced the training targets.

full rationale

The central personalization claim rests on the A/B preference test in Section 5.3. That test uses the same 15 participants whose IGA clicks optimized their personalized A–V coordinates (Section 4.2). Because the personalized coordinate is literally fitted to each participant's stated preferences, the 76% preference is a partially forced consequence of the optimization procedure; it confirms that the optimization found the participant's own preferred region, but it does not independently establish perceptual alignment for new listeners. This is a fitted-input-called-prediction pattern, not merely a statistical weakness. The culture-specific comparison uses new participants and is less confounded, though it is underpowered and lacks inferential statistics—a correctness concern, not circularity. Separately, the objective CCC in Section 5.1 is computed with the same SER model that provided training targets for the Emotion Feature Predictor, making the metric partially self-referential. Citations to [30] and [26] are not load-bearing self-citations; the issue is the training/evaluation model overlap, not author self-citation. Overall, the strongest empirical evidence for personalization reduces partly by construction, so a score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or latent entities are introduced; the individualized perception space is a re-coordinatization of Russell's A-V axes, not a new entity. The ledger focuses on the hand-set IGA hyperparameters, unreported search parameters, and the domain assumptions about the A-V model and SER-derived ground truth that the central claim depends on.

free parameters (6)
  • Initial mutation strength M1 = 0.20
    Set by hand in Eq. 2; controls IGA search step size and convergence behavior.
  • Mutation decay rate gamma = 0.90
    Set by hand in Eq. 2; controls the exploration-exploitation trade-off.
  • Minimum mutation strength M_min = 0.05
    Set by hand in Eq. 2; prevents premature convergence.
  • IGA population size N = not reported
    Algorithm 1 depends on N, but the experiments never state N or the number of generations per emotion, which directly affects search quality and user burden.
  • Initialization A-V coordinates = not stated
    Algorithm 1 initializes candidate coordinates without specifying whether they come from SER-estimated U.S. averages or another source, which affects comparison with the baseline.
  • Fourier feature dimension D_f and emotion feature dimension D_feat = not stated
    The Gaussian Fourier feature mapping and the Emotion Feature Predictor MLP depend on these dimensions, which are unspecified.
assumptions (5)
  • domain assumption Russell's arousal-valence circumplex is a sufficient and meaningful control space for emotional expression in TTS.
    The entire personalization operates in a 2D A-V space; if emotion perception is not reducible to arousal and valence, optimizing these coordinates cannot deliver the claimed alignment (Sections 1 and 3).
  • domain assumption A pretrained SER model (Wagner et al. [26]) provides valid continuous A-V annotations and high-level emotion features for American English emotional speech.
    Dataset A-V labels, training targets for the Emotion Feature Predictor, and the objective CCC all rely on this single SER model (Sections 3.2 and 4.1).
  • domain assumption The combined EXPRESSO, EmoV-DB, and ESD recordings form a coherent single-speaker emotional TTS dataset.
    Three datasets with different speakers and recording conditions are concatenated into a nine-hour American English female dataset without addressing speaker or domain mismatch (Section 4.1).
  • domain assumption Listener preferences expressed during the IGA loop are stable, honest, and reflect their genuine perception of the target emotion.
    The optimized coordinate is the output of an interactive search; if preferences are inconsistent or anchored by the sample population, the returned coordinate is not a stable personal map (Section 4.2, Algorithm 1).
  • ad hoc to paper Eq. 1's random weighted combination plus uniform mutation explores the 2D A-V space well enough to approach each listener's optimum within a few IGA generations.
    The paper states convergence typically occurs within three rounds but provides no convergence analysis or proof that the stochastic search covers the relevant A-V region (Section 3.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond One-Size-Fits-All: Personalized and Culturally Adaptive Emotional TTS via Interactive Optimization of Individual Emotion Perception Spaces." pith.science (2026). https://pith.science/paper/SHBWDLQB

@misc{pith2026260800998,
  author       = {Pith},
  title        = {Pith review of: Beyond One-Size-Fits-All: Personalized and Culturally Adaptive Emotional TTS via Interactive Optimization of Individual Emotion Perception Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SHBWDLQB}},
  note         = {Machine review of arXiv:2608.00998}
}
read the original abstract

The rise of conversational AI has increased interest in emotional Text-to-Speech (TTS). Most systems rely on discrete emotion labels, which fail to capture the nuanced nature of human affect. Recent models employ dimensional representations such as Russell's arousal-valence (A-V) model, offering finer control. However, emotional perception varies across individuals and cultures, which may cause mismatches between modeled and perceived emotions. We propose a personalized and culturally adaptive emotional TTS framework that performs interactive optimization of individualized A-V perception spaces using an Interactive Genetic Algorithm. By adapting emotion representations to each listener, the system produces speech with more perceptually aligned emotional expression than models using averaged A-V values. Evaluations with Japanese, Chinese, and Indonesian participants highlight the importance of personalization and cultural adaptation for moving beyond one-size-fits-all emotional TTS.

Figures

Figures reproduced from arXiv: 2608.00998 by the authors.

Figure 1
Figure 1. Overview of the proposed framework: (a) human-in-the-loop personalization using an Interactive Genetic Algorithm (IGA); (b) architecture of the emotional TTS generator with the proposed emotion controller. 2. Related Work 2.1. Controllable Emotional TTS Emotional speech synthesis has evolved from categorical emo￾tion representations (e.g., “happy,” “sad”) [8, 9, 10, 11] to di￾mensional emotion representations such a… view at source ↗
Figure 2
Figure 2. A–V distributions from the personalization process for (a) Chinese, (b) Indonesian, and (c) Japanese participants. The • indicates participant-selected A–V values, and × indicates averaged reference dataset values [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. A/B preference test results. source of each sample. Preference rates were 64.8%, 69.8%, and 65.6%, respectively, suggesting improved emotional align￾ment. Cross-cultural evaluation: In addition, we also conducted a cross-cultural evaluation in which participants ranked samples generated using Chinese, Indonesian, and Japanese A–V val￾ues. Participants from all three cultures consistently preferred samples generated … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 27 canonical work pages

  1. [26]

    Hierarchical genera- tive modeling for controllable speech synthesis,

    W.-N. Hsu, Y . Zhang, R. J. Weiss, H. Zen, Y . Wu, Y . Wang, Y . Cao, Y . Jia, Z. Chen, J. Shenet al., “Hierarchical genera- tive modeling for controllable speech synthesis,”arXiv preprint arXiv:1810.07217, 2018

    arXiv 2018

  2. [1]

    Beyond One-Size-Fits-All: Personalized and Culturally Adaptive Emotional TTS via Interactive Optimization of Individual Emotion Perception Spaces

    Introduction Human emotion is inherently complex, shaped by subtle phys- iological, psychological, and social factors. Rather than exist- ing as simple categorical states, emotional experiences lie on a continuum influenced by personality, gender, culture, and sit- uational context [1]. Capturing this variability is essential for speech synthesis systems ...

    work page Pith review arXiv 2026

  3. [2]

    happy,” “sad

    Related Work 2.1. Controllable Emotional TTS Emotional speech synthesis has evolved from categorical emo- tion representations (e.g., “happy,” “sad”) [8, 9, 10, 11] to di- mensional emotion representations such as arousal, valence, and dominance [12, 13, 14]. Dimensional representations en- able smoother and more fine-grained affect modulation com- pared ...

    work page

  4. [3]

    The overall archi- tecture is illustrated in Fig

    Proposed Framework We propose a personalized emotional TTS framework with human-in-the-loop preference learning. The overall archi- tecture is illustrated in Fig. 1 and consists of two compo- nents: (a) emotion preference learning that personalizes the arousal–valence (A–V) emotion space through user feedback, and (b) an emotion-controllable speech genera...

    work page

  5. [4]

    Dataset and Training Setup Dataset:We constructed a 9-hour American English fe- male emotional speech dataset (22.05 kHz) by combining EX- PRESSO [27], EmoV-DB [28], and ESD [9]

    Experimental Setup 4.1. Dataset and Training Setup Dataset:We constructed a 9-hour American English fe- male emotional speech dataset (22.05 kHz) by combining EX- PRESSO [27], EmoV-DB [28], and ESD [9]. Since these datasets do not provide continuous arousal–valence (A–V) an- notations, we estimated A–V values using a pre-trained Speech Emotion Recognition...

    work page 2000

  6. [5]

    Speech Quality and Emotional Similarity Table 1 summarizes the performance of the emotional speech generator before human-in-the-loop IGA personalization

    Experimental Results 5.1. Speech Quality and Emotional Similarity Table 1 summarizes the performance of the emotional speech generator before human-in-the-loop IGA personalization. We compare the baseline Grad-TTS with emotion embedding against Grad-TTS with the proposed emotion controller. (a)Chinese (b)Indonesian (c)Japanese Figure 2:A–V distributions f...

    work page

  7. [6]

    Conclusion We presented a novel personalized emotional TTS framework that incorporates human-in-the-loop optimization with an Inter- active Genetic Algorithm to address the subjectivity of emotion perception. Our results show that the proposed emotional con- troller improves both speech quality and emotional similarity, while personalization highlights in...

    work page

  8. [7]

    Acknowledgement Part of this work was supported by JSPS KAKENHI Grant Numbers JP24K0296 and JP25H01139, as well as JST NEXUS (JPMJNX25C1)

    work page

Show all 38 references
  1. [8]

    Generative AI Use Disclosure Generative AI tools were used only for minor language editing, such as grammar checking and wording refinement to improve clarity. All manuscript content and scientific work—including problem formulation, methodology development, experimental desig...

  2. [9]

    Gender and culture differences in emotion

    A. H. Fischer, P. M. Rodriguez Mosquera, A. E. Van Vianen, and A. S. Manstead, “Gender and culture differences in emotion.” Emotion, vol. 4, no. 1, p. 87, 2004

    2004

  3. [10]

    Plutchik and H

    R. Plutchik and H. Kellerman,Theories of emotion. Academic press, 2013, vol. 1

    2013

  4. [11]

    A circumplex model of affect

    J. A. Russell, “A circumplex model of affect.”Journal of person- ality and social psychology, vol. 39, no. 6, p. 1161, 1980

    1980

  5. [12]

    Matcha-tts: A fast tts architecture with conditional flow match- ing,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-tts: A fast tts architecture with conditional flow match- ing,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 341–11 345

    2024

  6. [13]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. JianZhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2025...

    2025

  7. [14]

    Naturalspeech 2: Latent diffusion models are natu- ral and zero-shot speech and singing synthesizers,

    K. Shen, Z. Ju, X. Tan, Y . Liu, Y . Leng, L. He, T. Qin, S. Zhao, and J. Bian, “Naturalspeech 2: Latent diffusion models are natu- ral and zero-shot speech and singing synthesizers,”arXiv preprint arXiv:2304.09116, 2023

    2023 arXiv

  8. [15]

    Styletts 2: Towards human-level text-to-speech through style dif- fusion and adversarial training with large speech language mod- els,

    Y . A. Li, C. Han, V . Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to-speech through style dif- fusion and adversarial training with large speech language mod- els,”Advances in neural information processing systems, vol. 36, pp. 19 594–19 621, 2023

    2023

  9. [16]

    Deep learning-based expressive speech synthesis: a systematic review of approaches, challenges, and resources,

    H. Barakat, O. Turk, and C. Demiroglu, “Deep learning-based expressive speech synthesis: a systematic review of approaches, challenges, and resources,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2024, no. 1, p. 11, 2024

    2024

  10. [17]

    Emotional voice con- version: Theory, databases and esd,

    K. Zhou, B. Sisman, R. Liu, and H. Li, “Emotional voice con- version: Theory, databases and esd,”Speech Communication, vol. 137, pp. 1–18, 2022

    2022

  11. [18]

    Emospeech: Guiding fast- speech2 towards emotional text to speech,

    D. Diatlova and V . Shutov, “Emospeech: Guiding fast- speech2 towards emotional text to speech,”arXiv preprint arXiv:2307.00024, 2023

    2023 arXiv

  12. [19]

    Emo-dpo: Controllable emotional speech synthesis through direct preference optimization,

    X. Gao, C. Zhang, Y . Chen, H. Zhang, and N. F. Chen, “Emo-dpo: Controllable emotional speech synthesis through direct preference optimization,” inICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  13. [20]

    Emosphere-tts: Emotional style and intensity modeling via spherical emotion vector for controllable emotional text-to- speech,

    D.-H. Cho, H.-S. Oh, S.-B. Kim, S.-H. Lee, and S.-W. Lee, “Emosphere-tts: Emotional style and intensity modeling via spherical emotion vector for controllable emotional text-to- speech,”arXiv preprint arXiv:2406.07803, 2024

    2024 arXiv

  14. [21]

    Emo- sphere++: Emotion-controllable zero-shot text-to-speech via emotion-adaptive spherical vector,

    D.-H. Cho, H.-S. Oh, S.-B. Kim, and S.-W. Lee, “Emo- sphere++: Emotion-controllable zero-shot text-to-speech via emotion-adaptive spherical vector,”IEEE Transactions on Affec- tive Computing, 2025

    2025

  15. [22]

    Ece-tts: A zero-shot emotion text-to-speech model with simplified and precise control,

    S. Liang, R. Zhou, and Q. Yuan, “Ece-tts: A zero-shot emotion text-to-speech model with simplified and precise control,”Applied Sciences, vol. 15, no. 9, p. 5108, 2025

    2025

  16. [23]

    Train- ing language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Rayet al., “Train- ing language models to follow instructions with human feedback,” Advances in neural information processing systems, vol. 35, pp. 27 730–27 744, 2022

    2022

  17. [24]

    A survey of reinforcement learning from human feedback,

    T. Kaufmann, P. Weng, V . Bengs, and E. H ¨ullermeier, “A survey of reinforcement learning from human feedback,” 2024

    2024

  18. [25]

    Towards creative evolutionary systems with interac- tive genetic algorithm,

    S.-B. Cho, “Towards creative evolutionary systems with interac- tive genetic algorithm,”Applied Intelligence, vol. 16, no. 2, pp. 129–138, 2002

    2002

  19. [27]

    Natural tts synthesis by conditioning wavenet on mel spectrogram pre- dictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerrv-Ryanet al., “Natural tts synthesis by conditioning wavenet on mel spectrogram pre- dictions,” in2018 IEEE international conference on acoustics, speech and signal processing...

    2018

  20. [28]

    Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,” inInternational confer- ence on machine learning. PMLR, 2022, pp. 2709–2720

    2022

  21. [29]

    A review on genetic algorithm: past, present, and future,

    S. Katoch, S. S. Chauhan, and V . Kumar, “A review on genetic algorithm: past, present, and future,”Multimedia tools and appli- cations, vol. 80, no. 5, pp. 8091–8126, 2021

    2021

  22. [30]

    Optimization of sound of au- tonomous sensory meridian response with interactive genetic al- gorithm,

    M. Fukumoto and K. Hijikuro, “Optimization of sound of au- tonomous sensory meridian response with interactive genetic al- gorithm,” in2020 Joint 11th International Conference on Soft Computing and Intelligent Systems and 21st International Sympo- sium on Advanced Intelligent ...

    2020

  23. [31]

    Evolutionary process indi- cators for active igas applied to weight tuning in unit selection tts synthesis,

    L. Formiga, F. Al ´ıas, and X. Llor `a, “Evolutionary process indi- cators for active igas applied to weight tuning in unit selection tts synthesis,” inIEEE Congress on Evolutionary Computation. IEEE, 2010, pp. 1–8

    2010

  24. [32]

    Grad-tts: A diffusion probabilistic model for text-to-speech,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, and M. Kudinov, “Grad-tts: A diffusion probabilistic model for text-to-speech,” in International conference on machine learning. PMLR, 2021, pp. 8599–8608

    2021

  25. [33]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Y . Ren, C. Hu, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y . Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” arXiv preprint arXiv:2006.04558, 2020

    2006 arXiv

  26. [34]

    Dawn of the trans- former era in speech emotion recognition: Closing the valence gap,

    J. Wagner, A. Triantafyllopoulos, H. Wierstorf, M. Schmitt, F. Burkhardt, F. Eyben, and B. W. Schuller, “Dawn of the trans- former era in speech emotion recognition: Closing the valence gap,”IEEE Transactions on Pattern Analysis and Machine Intel- ligence, pp. 1–13, 2023

    2023

  27. [35]

    Ex- presso: A benchmark and analysis of discrete expressive speech resynthesis,

    T. A. Nguyen, W.-N. Hsu, A. d’Avirro, B. Shi, I. Gat, M. Fazel- Zarani, T. Remez, J. Copet, G. Synnaeve, M. Hassidet al., “Ex- presso: A benchmark and analysis of discrete expressive speech resynthesis,”arXiv preprint arXiv:2308.05725, 2023

    2023 arXiv

  28. [36]

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

    A. Adigwe, N. Tits, K. E. Haddad, S. Ostadabbas, and T. Du- toit, “The emotional voices database: Towards controlling the emotion dimension in voice generation systems,”arXiv preprint arXiv:1806.09514, 2018

    2018 arXiv

  29. [37]

    HiFi-GAN: Generative adversar- ial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversar- ial networks for efficient and high fidelity speech synthesis,”Adv. Neural Inf. Process. Syst., vol. 2020-December, no. NeurIPS, 2020

    2020

  30. [38]

    Evaluation of error- And correlation- based loss functions for multitask learning dimensional speech emotion recognition,

    B. T. Atmaja and M. Akagi, “Evaluation of error- And correlation- based loss functions for multitask learning dimensional speech emotion recognition,”J. Phys. Conf. Ser ., vol. 1896, no. 1, p. 012004, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.