Pith. sign in

REVIEW 3 major objections 5 minor 44 references

ZeSTA: Zero-Shot TTS Augmentation with Domain-Conditioned Training for Data-Efficient Personalized Speech Synthesis

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read ZeSTA claims that conditioning a TTS model on a binary real/synthetic domain label, plus oversampling real utterances, lets low-resource personalized TTS exploit abundant zero-shot synthetic speech without losing the target speaker's voice.

desk verdict A simple, useful recipe for making ZS-TTS augmentation work in low-resource personalization; evidence is consistent but the domain-label mechanism needs a placebo control. read the letter →

arxiv 2603.04219 v3 pith:ISEWEFLJ submitted 2026-03-04 cs.SD cs.AIeess.AS

Youngwon Choi , Jinwoo Oh , Hwayeon Kim , Hyeonyu Kim This is my paper
classification cs.SDcs.AIeess.AS
keywords zero-shotTTSdomain-conditionedtraininglow-resourceadaptationspeakersimilaritysyntheticdataaugmentationoversamplingpersonalizedspeechsynthesisdata-efficientvoicecloning
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

ZeSTA tries to show that the speaker-similarity drop caused by augmenting low-resource TTS fine-tuning with zero-shot synthetic speech can be undone by a simple domain label. The method adds a small embedding marking each training sample as real or synthetic and, at inference, forces the label to real; it also oversamples the few real recordings. In their numbers, this restores speaker embedding similarity to nearly the real-only level (0.815 vs 0.818) while keeping the intelligibility gain (WER 10.563 vs 12.520). If right, zero-shot TTS becomes a reliable data source for personalized voice building with only 10% real recordings.

What carries the argument

The central object is a small domain embedding, appended to the acoustic generator's conditioning input, representing d in {real, synthetic}. It is used alongside the speaker embedding; the model is trained with each sample labeled by its origin, and at inference d is fixed to real. The paper also repeats each real utterance three times (oversampling). The work of the domain embedding is to let the text encoder remain speaker-agnostic and shared across domains while the acoustic module is told which acoustic style to expect, so synthetic data contributes linguistic and phonetic diversity without pulling the voice statistics away.

What would settle it

Train ZeSTA on a third zero-shot generator that produces speech with markedly lower speaker similarity (SECS well below 0.7) or on speaker-mismatched synthetic data; the paper's own matched-versus-mismatched ablation (SECS 0.807 vs 0.792) shows the method cannot overcome a large domain gap. If the binary domain label fails to restore similarity in such a setting, the claim that a single domain label separates the acoustic shift from speaker identity is refuted.

Watch

Extended reading notes

Core claim

On its own terms, ZeSTA's central claim is that a binary domain label—'real' or 'synthetic'—injected into the acoustic generation stage of a TTS model during fine-tuning, together with a modest oversampling of real utterances, removes most of the speaker-identity drift caused by mixing large amounts of zero-shot TTS speech into a low-resource personalized training set, while keeping the intelligibility gains. In the main setup (10% real + 90% synthetic), speaker embedding similarity reaches 0.815 with the proposed method versus 0.765 for naive mixing and 0.818 for real-only training; WER improves from 12.520 (real-only) to 10.563. Subjective ABX tests also show listeners prefer the condition

Load-bearing premise

The load-bearing premise is that a single binary label 'real' versus 'synthetic' captures enough of the acoustic domain shift to remove it at inference, with speaker identity and text content untouched; if the domain discrepancy is multidimensional or entangled with content, conditioning on d=real will not fully restore the target speaker.

Editorial extensions

If this is right

  • Low-resource TTS fine-tuning can safely use large amounts of zero-shot synthetic speech; the target voice is preserved with no architectural change to the base TTS model.
  • The domain-conditioning plus oversampling combination is more effective than oversampling alone, and it works with two different zero-shot generators, suggesting the recipe is generator-agnostic.
  • Intelligibility is not sacrificed: WER with domain conditioning and oversampling stays near naive-mixing levels, below real-only, while speaker similarity is restored.
  • Additional synthetic data beyond the 90% ratio can be added if filtered for low word error rate, improving CER and WER with only a small similarity cost.
  • The domain embedding should be moderate in size (about 64 units); very small embeddings hurt intelligibility, and large ones slightly reduce speaker similarity.

Reading between the lines

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

  • The binary domain label is likely a coarse proxy; the paper's own t-SNE and speaker-mismatch experiments show the shift is larger for mismatched speakers, so per-source or continuous domain conditioning may be needed for generators with a larger domain gap.
  • The intelligibility gain from synthetic data may come partly from lower acoustic variability, which the paper itself suggests; if so, the same conditioning trick could transfer to other tasks where synthetic speech is used as augmentation, such as ASR or speaker verification.
  • The method's success depends on speaker-consistent synthetic data; the paper shows mismatched augmentation fails even with domain conditioning, so practitioners should pair conditioning with speaker-matched generation or filtering.
  • A natural next step is to replace the binary domain label with a domain-adversarial objective to close the residual gap visible in the t-SNE overlap, potentially pushing the 0.815 speaker similarity figure closer to the real-only 0.818 baseline.
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 / 5 minor

Summary. The paper proposes ZeSTA, a fine-tuning recipe for low-resource personalized TTS in which a large amount of ZS-TTS synthetic speech is mixed with a small set of real target-speaker recordings. To prevent the synthetic data from degrading speaker similarity, the method adds a binary domain embedding (real vs. synthetic) to the conditioning stack of a VITS-based multi-speaker TTS model and oversamples the scarce real utterances during fine-tuning. At inference, the model is conditioned on the real domain. Experiments on LibriTTS and an in-house dataset (YoBind) with two ZS-TTS generators (Fish-Speech and CosyVoice 2) show that Domain Conditioning (DC) plus oversampling (OS) improves SECS over naive mixing by 0.026–0.05, with ABX preferences of 60–71% over the naive baseline, while largely retaining the intelligibility gains of synthetic augmentation.

Significance. If the central claim holds, ZeSTA is a simple, practical contribution: a plug-in conditioning label and an oversampling factor that let practitioners use abundant ZS-TTS data without sacrificing speaker fidelity in few-shot TTS adaptation. The paper has clear strengths: it tests two different ZS-TTS generators and two corpora, includes both objective and human evaluation, and reports ablations on embedding size and on speaker-matched vs. speaker-mismatched synthetic data. The method itself is architecturally minimal and easy to reproduce. The main weakness is that the causal role of the domain label is not isolated from the added model capacity, and the objective tables lack variance information. If the placebo concern is resolved, the contribution would be solid for a speech/audio venue.

major comments (3)
  1. [§4.2, Table 3] The central comparison between DC and no DC (e.g., LibriTTS/Fish-Speech SECS 0.807 vs. 0.765) conflates two changes: the addition of a 64-dimensional conditioning vector and the semantic mapping of that vector to real/synthetic origin. The paper never includes a placebo condition in which the domain label is randomized, reversed, or held constant while preserving the same parameter count. Without such a control, the SECS gain cannot be uniquely attributed to the real/synthetic distinction as claimed in the abstract. The authors should add a randomized-label or null-label ablation to Table 3 or Table 5; if the gain persists, the mechanistic interpretation in §3.2 needs revision.
  2. [Table 3, §4.2] Objective results are reported as single averages over three random seeds, with no standard deviations or confidence intervals. Several comparisons are numerically small (e.g., LibriTTS/CV2 extra-synth SECS 0.814 vs. 0.815), and without error bars it is impossible to assess whether the claimed improving trends are stable. The subjective ABX results are significant, but the objective table is a primary source of evidence for the 'preserving intelligibility' claim. Please provide per-seed results or variance estimates.
  3. [§1, §3.2, Figure 1] The paper repeatedly states the method works 'without modifying the base architecture,' but it actually reduces the speaker-embedding size from 256 to 64 and adds a domain embedding. This is a modification to the model's conditioning mechanism, even if the main backbone is unchanged. The wording is misleading and should be clarified, otherwise readers may assume zero architectural change.
minor comments (5)
  1. [§4.1, 'Datasets'] The choice of the longest real utterance as the ZS-TTS reference prompt is motivated by [30], but the potential interaction with utterance-level prosodic variation is not discussed. A sentence or two on why this is preferred over multiple reference prompts would help.
  2. [Table 3, '+Extra Synth'] The amount of extra synthetic data is described only as '800 randomly sampled transcripts from the VCTK corpus.' Please state the resulting number of Synth utterances per speaker and the total ratio of synthetic to real data for this condition.
  3. [§4.3, Figure 2] The t-SNE visualizations are qualitative. The 'moderate domain shift' claim would be more convincing with a quantitative metric, e.g., the centroids' Euclidean distance or a simple classifier accuracy on the latent features.
  4. [§4.3, Table 5] The embedding-size analysis is useful, but it is only run on LibriTTS with Fish-Speech. The authors could note whether the same non-monotonic trend was observed for the other condition, or justify why a single configuration suffices.
  5. [General] Minor typos: 'V oice conversion' in §2, 'V AD' in §4.1, and 'CosyV oice 2' in Table 2. Also, the audio samples link is mentioned in the abstract but not provided in the submission; please add the URL.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: ZeSTA's gains are empirical, evaluated on held-out test data with external metrics; the domain label is a training conditioning input, not a renamed evaluation target.

full rationale

The paper's central claim is an empirical one: domain-conditioned training (DC) plus real-data oversampling (OS) improves speaker similarity over naive synthetic augmentation while preserving intelligibility. There is no derivation chain in which a fitted parameter or input is renamed as a prediction. The conditional model p(y|x,d) with d in {real,synthetic} is trained with a standard likelihood objective and evaluated on held-out test utterances using external metrics (ECAPA-TDNN for SECS, Whisper for CER/WER, human MOS/ABX). Setting d=real at inference is a design choice, not an identity; the model must still learn to generate the target speaker's voice, and the improvement over no-DC in Table 3 is an experimentally measured contrast, not a consequence of the label by construction. The only citation to the authors' prior work, [7], supplies pretraining hyperparameters ('we follow the training setup described in [7]'), which is not load-bearing for the main result. The use of Whisper both to filter extra synthetic training data and to measure WER is a mild evaluation-pipeline coupling, but the filter applies to training samples, not to the test outputs, so it does not force the reported generalization results. The absence of a placebo domain-label control is a causal-identification concern, not a circularity. Overall, the method is self-contained and testable against external benchmarks, so no significant circularity is present.

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

The ledger is light: no new physical or architectural entities are postulated, and there is no derivation-based circularity. The load-bearing items are the sufficiency of a single binary domain tag, the transferability of the VITS text encoder, and a handful of hand-set hyperparameters (embedding size 64, oversampling factor 3, 600 epochs, WER<5% filter) that are not ablated. The in-house dataset and code are the main reproducibility obstacles.

free parameters (4)
  • domain_embedding_size = 64
    Chosen from {16, 64, 256} on LibriTTS (Table 5); 16 and 64 give identical SECS (0.807), 64 is selected for best CER/WER trade-off. A sensitivity pick rather than a derived quantity.
  • oversampling_factor_OS = 3
    Set by hand in §3.3 ('modestly emphasizing real samples... by a small factor') and used in all main experiments; no ablation or sensitivity analysis is reported, so the central results may depend on this value.
  • WER_filter_threshold = 5%
    In the (+Extra Synth) setting, synthesized utterances are retained only if Whisper-medium WER < 5% (§4.2); the threshold is chosen without sensitivity analysis and uses the same ASR model as the evaluation.
  • fine_tuning_epochs = 600
    Fine-tuning is run for 600 epochs for all speakers 'to ensure stable convergence' (§4.1); no early stopping or per-speaker selection is described, making this a hand-set hyperparameter.
assumptions (5)
  • ad hoc to paper A single binary domain label d ∈ {real, synthetic} is sufficient to capture the real-vs-synthetic domain shift during fine-tuning, with inference at d=real.
    Core modeling choice in §3.2 (model ŷ = g(h_ling, d)). The paper's own §4.4 shows the domain gap changes with speaker consistency, and Table 2 shows different ZS-TTS sources have different similarity, so one binary tag may not cover all domain shifts.
  • domain assumption ZS-TTS generated speech preserves the target speaker's voice and the input text's linguistic content well enough to serve as beneficial augmentation.
    Invoked in §3.1; the paper measures SECS of 0.756–0.794 for synthetic speech (Table 2), but this is itself a fitted measurement and may not hold for other ZS-TTS models or reference-selection strategies.
  • domain assumption The VITS architecture trained on VCTK is a representative base for personalized TTS, and its text encoder is speaker-agnostic while the acoustic module can be modulated by an extra embedding.
    Assumed in §3.2 and used throughout; only VITS is tested, so generality to other TTS backbones is asserted rather than demonstrated.
  • domain assumption ECAPA-TDNN cosine similarity and Whisper-based CER/WER are valid proxies for speaker similarity and intelligibility.
    Evaluation protocol in §4.1; these are standard external models, but they are specific choices that influence all reported numbers.
  • domain assumption The low-resource protocol (10% of train data, ~13–37 real utterances per speaker) and the 14 selected speakers are representative of real personalization scenarios.
    Dataset construction in §4.1; only 8 LibriTTS speakers (filtered by utterance length and token content) and 6 in-house YoBind speakers (recording conditions undisclosed) are used, so external validity is limited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZeSTA: Zero-Shot TTS Augmentation with Domain-Conditioned Training for Data-Efficient Personalized Speech Synthesis." pith.science (2026). https://pith.science/paper/ISEWEFLJ

@misc{pith2026260304219,
  author       = {Pith},
  title        = {Pith review of: ZeSTA: Zero-Shot TTS Augmentation with Domain-Conditioned Training for Data-Efficient Personalized Speech Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ISEWEFLJ}},
  note         = {Machine review of arXiv:2603.04219}
}
read the original abstract

We investigate the use of zero-shot text-to-speech (ZS-TTS) as a data augmentation source for low-resource personalized speech synthesis. While synthetic augmentation can provide linguistically rich and phonetically diverse speech, naively mixing large amounts of synthetic speech with limited real recordings often leads to speaker similarity degradation during fine-tuning. To address this issue, we propose ZeSTA, a simple domain-conditioned training framework that distinguishes real and synthetic speech via a lightweight domain embedding, combined with real-data oversampling to stabilize adaptation under extremely limited target data, without modifying the base architecture. Experiments on LibriTTS and an in-house dataset with two ZS-TTS sources demonstrate that our approach improves speaker similarity over naive synthetic augmentation while preserving intelligibility and perceptual quality. Audio samples are available on our web page.

Figures

Figures reproduced from arXiv: 2603.04219 by the authors.

Figure 1
Figure 1. ZeSTA framework: training pipeline (a) and inference pipeline (b). by the speaker embedding cosine similarity scores in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of latent representations for an example target speaker under different synthetic augmentation settings [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 linked inside Pith

  1. [1]

    With these advances, personalized TTS, which adapts a model to a specific target speaker, has gained increasing attention with the growing demand for custom voices [5, 6]

    Introduction Recent neural text-to-speech (TTS) models [1–4] have achieved near human-level naturalness under sufficient training data, in- cluding lightweight architectures suitable for practical deploy- ment. With these advances, personalized TTS, which adapts a model to a specific target speaker, has gained increasing attention with the growing demand ...

  2. [2]

    V oice conver- sion (VC) methods [16–18] have been used to create additional target-speaker utterances without parallel data

    Related Works Previous studies have explored generating synthetic speech samples to expand training data for TTS systems. V oice conver- sion (VC) methods [16–18] have been used to create additional target-speaker utterances without parallel data. However, these approaches typically require training or adapting a VC model using recordings from the target ...

  3. [3]

    Method 3.1. Zero-Shot Speech Synthesis for Data Augmentation To mitigate data scarcity in personalized TTS, we leverage publicly available ZS-TTS models as external data genera- tors to synthesize additional speech for adaptation. Given a limited set of reference utterances from a target speaker, the source ZS-TTS models generate speech conditioned on the...

    arXiv 2026

  4. [4]

    Experiments 4.1. Experimental Setup Models.We employ two publicly available ZS-TTS systems with different architectures as source models to examine the Table 1:Summary statistics of the LibriTTS and YoBind datasets. Both datasets are gender-balanced. Ranges denote the minimum and maximum number of utterances per speaker. Corpus #Spk (M/F) Train Valid Test...

  5. [5]

    Conclusion We investigate the use of ZS-TTS as a data augmentation source for low-resource personalized speech synthesis and show that naively incorporating synthetic speech often degrades speaker similarity during fine-tuning. To address this challenge, we propose ZeSTA, a simple domain-conditioned training frame- work with real-data oversampling that mi...

  6. [6]

    Generative AI Use Disclosure During the preparation of this manuscript, the authors used OpenAI’s ChatGPT for editing and language polishing, and the Cursor IDE for implementation-level assistance such as code refinement and debugging. These tools were used solely to im- prove readability, grammar, expression, and code clarity, and were not involved in th...

  7. [7]

    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,” inICLR, 2021

    2021

  8. [8]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” inICML, 2021

    2021

Show all 44 references
  1. [9]

    Jets: Jointly training fastspeech2 and hifi-gan for end to end text to speech,

    D. Lim, S. Jung, and E. Kim, “Jets: Jointly training fastspeech2 and hifi-gan for end to end text to speech,” inInterspeech, 2022

    2022

  2. [10]

    Glow-tts: A genera- tive flow for text-to-speech via monotonic alignment search,

    J. Kim, S. Kim, J. Kong, and S. Yoon, “Glow-tts: A genera- tive flow for text-to-speech via monotonic alignment search,” in NeurIPS, 2020

    2020

  3. [11]

    Adaspeech: Adaptive text to speech for custom voice,

    M. Chen, X. Tan, B. Li, Y . Liu, T. Qin, T.-Y . Liuet al., “Adaspeech: Adaptive text to speech for custom voice,” inICLR, 2021

    2021

  4. [12]

    Meta-stylespeech: Multi-speaker adaptive text-to-speech generation,

    D. Min, D. B. Lee, E. Yang, and S. J. Hwang, “Meta-stylespeech: Multi-speaker adaptive text-to-speech generation,” inICML, 2021

    2021

  5. [13]

    Leveraging low-rank adapta- tion for parameter-efficient fine-tuning in multi-speaker adaptive text-to-speech synthesis,

    C. Hong, J. H. Lee, and H. K. Kim, “Leveraging low-rank adapta- tion for parameter-efficient fine-tuning in multi-speaker adaptive text-to-speech synthesis,”IEEE Access, 2024

    2024

  6. [14]

    V oice- box: text-guided multilingual universal speech generation at scale,

    M. Le, A. Vyas, B. Shi, B. Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokaret al., “V oice- box: text-guided multilingual universal speech generation at scale,” inNeurIPS, 2023

    2023

  7. [15]

    Audiolm: a language modeling approach to audio genera- tion,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi et al., “Audiolm: a language modeling approach to audio genera- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 25...

    2023

  8. [16]

    Speak, read and prompt: High-fidelity text-to-speech with min- imal supervision,

    E. Kharitonov, D. Vincent, Z. Borsos, R. Marinier, S. Girgin, O. Pietquin, M. Sharifi, M. Tagliasacchi, and N. Zeghidour, “Speak, read and prompt: High-fidelity text-to-speech with min- imal supervision,”Transactions of the Association for Computa- tional Linguistics, vol. 11,...

    2023

  9. [17]

    Sc-glowtts: An efficient zero-shot multi-speaker text-to-speech model,

    E. Casanova, C. Shulby, E. G ¨olge, N. M. M¨uller, F. S. de Oliveira, A. Candido Jr, A. da Silva Soares, S. M. Aluisio, and M. A. Ponti, “Sc-glowtts: An efficient zero-shot multi-speaker text-to-speech model,” inINTERSPEECH, 2021

    2021

  10. [18]

    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,” inICML, 2022

    2022

  11. [19]

    Hierspeech: Bridging the gap between text and speech by hierarchical variational inference using self-supervised represen- tations for speech synthesis,

    S.-H. Lee, S.-B. Kim, J.-H. Lee, E. Song, M.-J. Hwang, and S.-W. Lee, “Hierspeech: Bridging the gap between text and speech by hierarchical variational inference using self-supervised represen- tations for speech synthesis,” inNeurIPS, 2022

    2022

  12. [20]

    Adaptermix: Exploring the efficacy of mixture of adapters for low-resource tts adaptation,

    A. Mehrish, A. R. Kashyap, Y . Li, N. Majumder, and S. Poria, “Adaptermix: Exploring the efficacy of mixture of adapters for low-resource tts adaptation,” inINTERSPEECH, 2023

    2023

  13. [21]

    Meta-tts: Meta-learning for few-shot speaker adaptive text-to- speech,

    S.-F. Huang, C.-J. Lin, D.-R. Liu, Y .-C. Chen, and H.-y. Lee, “Meta-tts: Meta-learning for few-shot speaker adaptive text-to- speech,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 30, pp. 1558–1571, 2022

    2022

  14. [22]

    Low-resource expressive text-to-speech us- ing data augmentation,

    G. Huybrechts, T. Merritt, G. Comini, B. Perz, R. Shah, and J. Lorenzo-Trueba, “Low-resource expressive text-to-speech us- ing data augmentation,” inICASSP, 2021

    2021

  15. [23]

    Low-data? no problem: low- resource, language-agnostic conversational text-to-speech via f0- conditioned data augmentation,

    G. Comini, G. Huybrechts, M. S. Ribeiro, A. Gabry ´s, and J. Lorenzo-Trueba, “Low-data? no problem: low- resource, language-agnostic conversational text-to-speech via f0- conditioned data augmentation,” inINTERSPEECH, 2022

    2022

  16. [24]

    Cross-speaker style transfer for text-to- speech using data augmentation,

    M. S. Ribeiro, J. Roth, G. Comini, G. Huybrechts, A. Gabry ´s, and J. Lorenzo-Trueba, “Cross-speaker style transfer for text-to- speech using data augmentation,” inICASSP, 2022

    2022

  17. [25]

    Strawnet: Self-training wavenet for tts in low-data regimes

    M. Sharma, T. Kenter, and R. Clark, “Strawnet: Self-training wavenet for tts in low-data regimes.” inINTERSPEECH, 2020

    2020

  18. [26]

    Tts-by- tts: Tts-driven data augmentation for fast and high-quality speech synthesis,

    M.-J. Hwang, R. Yamamoto, E. Song, and J.-M. Kim, “Tts-by- tts: Tts-driven data augmentation for fast and high-quality speech synthesis,” inICASSP, 2021

    2021

  19. [27]

    Tts-by-tts 2: Data- selective augmentation for neural speech synthesis using ranking support vector machine with variational autoencoder,

    E. Song, R. Yamamoto, O. Kwon, C.-H. Song, M.-J. Hwang, S. Oh, H.-W. Yoon, J.-S. Kim, and J.-M. Kim, “Tts-by-tts 2: Data- selective augmentation for neural speech synthesis using ranking support vector machine with variational autoencoder,” inINTER- SPEECH, 2022

    2022

  20. [28]

    Zero shot text to speech augmentation for automatic speech recognition on low-resource accented speech corpora,

    F. Nespoli, D. Barreda, and P. A. Naylor, “Zero shot text to speech augmentation for automatic speech recognition on low-resource accented speech corpora,” inAsilomar Conference on Signals, Systems, and Computers, 2023

    2023

  21. [29]

    Generative data augmentation challenge: Zero- shot speech synthesis for personalized speech enhancement,

    J.-S. Bae, A. Kuznetsova, D. Manocha, J. Hershey, T. Kristjans- son, and M. Kim, “Generative data augmentation challenge: Zero- shot speech synthesis for personalized speech enhancement,” in ICASSP Workshops, 2025

    2025

  22. [30]

    Data augmentation using speech synthesis for speaker-independent dysarthria sever- ity classification,

    M. Kim, M. Han, S. Hong, and M.-w. Koo, “Data augmentation using speech synthesis for speaker-independent dysarthria sever- ity classification,” inINTERSPEECH, 2025

    2025

  23. [31]

    Multispeech: Multi-speaker text to speech with transformer,

    M. Chen, X. Tan, Y . Ren, J. Xu, H. Sun, S. Zhao, and T. Qin, “Multispeech: Multi-speaker text to speech with transformer,” in INTERSPEECH, 2020

    2020

  24. [32]

    Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,

    S. Liao, Y . Wang, T. Li, Y . Cheng, R. Zhang, R. Zhou, and Y . Xing, “Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,”arXiv preprint arXiv:2411.01156, 2024

    2024 arXiv

  25. [33]

    Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    2024 arXiv

  26. [34]

    CSTR VCTK Cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK Cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” 2019

    2019

  27. [35]

    Libritts: A corpus derived from librispeech for text- to-speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text- to-speech,” inINTERSPEECH, 2019

    2019

  28. [36]

    Zero-shot tts with enhanced audio prompts: Bsc submission for the 2026 wildspoof challenge tts track,

    J. Giraldo, A. Peir ´o-Lilja, R. Zevallos, and C. Espa ˜na-Bonet, “Zero-shot tts with enhanced audio prompts: Bsc submission for the 2026 wildspoof challenge tts track,”arXiv preprint arXiv:2602.05770, 2026, accepted to ICASSP 2026

    2026

  29. [37]

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

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inINTERSPEECH, 2020

    2020

  30. [38]

    V oxceleb: A large- scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A large- scale speaker identification dataset,” inINTERSPEECH, 2017

    2017

  31. [39]

    V oxceleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,” inINTERSPEECH, 2018

    2018

  32. [40]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inICML, 2023

    2023

  33. [41]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” inICASSP, 2015

    2015

  34. [42]

    Evaluating and reducing the distance between synthetic and real speech distributions,

    C. Minixhofer, O. Klejch, and P. Bell, “Evaluating and reducing the distance between synthetic and real speech distributions,” in INTERSPEECH, 2023

    2023

  35. [43]

    Advanc- ing zero-shot text-to-speech intelligibility across diverse domains via preference alignment,

    X. Zhang, Y . Wang, C. Wang, Z. Li, Z. Chen, and Z. Wu, “Advanc- ing zero-shot text-to-speech intelligibility across diverse domains via preference alignment,” inACL, 2025

    2025

  36. [44]

    Visualizing data using t-sne,

    L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of machine learning research, vol. 9, no. Nov, pp. 2579– 2605, 2008

    2008

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.