Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

DS-TTS: Zero-Shot Speaker Style Adaptation from Voice Clips via Dynamic Dual-Style Feature Modulation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that DS-TTS, by pairing mel-spectrogram and MFCC style encoders with a length-adaptive variance adapter, achieves better zero-shot voice cloning—higher speaker similarity with competitive word error rates—than existing…

desk verdict A plausible dual-style encoder idea, but the paper's own tables contradict its 'superior performance' claim, the SGF mechanism collapses mathematically, and the key threshold is tuned on the test set. read the letter →

arxiv 2506.01020 v1 pith:WKWS2FP6 submitted 2025-06-01 cs.SD eess.AS

Ming Meng , Ziyi Yang , Jian Yang , Zhenjie Su , Yonggui Zhu , Zhaoxin Fan This is my paper · ORCID
classification cs.SDeess.AS
keywords zero-shotvoicecloningtext-to-speechspeakerstyleadaptationdual-styleencodingMFCCgating-filmdynamicvarianceadaptersimilarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes DS-TTS, a text-to-speech system for zero-shot voice cloning: given one audio clip from a speaker never seen in training and an arbitrary text, it aims to synthesize speech that sounds like that speaker. The central claim is that the model outperforms existing zero-shot cloning systems on this task, improving the balance between word error rate and speaker similarity. To do this, DS-TTS extracts two complementary style vectors from the reference clip—one from mel-spectrograms, one from MFCCs—and injects them into both the phoneme encoder and the mel-spectrogram decoder through a Style Gating-Film mechanism. A dynamic variance adapter then switches between convolutional predictors for short phoneme sequences and linear predictors for long ones, which the paper argues fixes the unnatural short-utterance output of prior models. If this is right, voice cloning from a single clip becomes more reliable for short commands as well as long sentences.

What carries the argument

The central mechanism is the pair formed by DuSEN and DyGN, joined by the Style Gating-Film (SGF) layer. SGF normalizes a hidden feature vector $h$, computes $\gamma = \tanh(f(x))$, $\beta = \tanh(f(x))$, $\eta = \tanh(f(x))$, $\delta = \sigma(f(x))$ from the style vector $f(x)$, and blends them as $\gamma = \gamma\delta + \eta(1-\delta)$ and $\beta = \beta\delta + \eta(1-\delta)$ before applying $\text{SGF}(y) = \gamma y + \beta$. This lets the style vector scale and shift normalized features while a gating term filters out parts of the style that would hurt synthesis. The other load-bearing piece is the Dynamic Variance Adapter, which routes pitch, energy, and duration prediction through conv1d-based predictors when the phoneme sequence is at most 85 tokens and through linear-layer predictors above that, so that network complexity matches input length.

What would settle it

Evaluate the threshold sweep on a held-out validation split or a second corpus: if 85 phonemes no longer gives the best WER and SMCS, or the margin over nearby thresholds disappears, the dynamic variance adapter's claimed contribution is not established.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that speaker identity for zero-shot cloning is better captured by two complementary acoustic views than by one, and that acoustic-feature predictors should be chosen by input length. The Mel-Style encoder reads global prosody from the mel-spectrogram; the MFCC-Style encoder models fine timbral detail with a bidirectional LSTM and multi-head self-attention; their concatenated 256-dimensional vector is fed through Style Gating-Film, a normalization modulation with additional gating parameters, into the generator. The Dynamic Generator Network's variance adapter splits phoneme sequences at 85 phonemes: shorter sequences use Conv1D predictors for pitch, energy, and duration, while longer ones use linear-layer predictors. Reported on 108 unseen VCTK speakers, DS-TTS achieves a word error rate of 0.047 and speaker similarity of 0.865, with similarity-based listening scores (SMOS) of 3.72 on short and 3.99 on long sequences, which the paper reads as the best overall balance among the compared systems.

Load-bearing premise

The dynamic adapter's advantage depends on the 85-phoneme cutoff being a real property of the architecture and data, not a number selected by sweeping on the same VCTK test set used to report the final results.

Editorial extensions

If this is right

  • A single clip from an unseen speaker can drive synthesis at speaker similarity close to a large-codec model (0.865 vs 0.868 SMCS) while cutting word error rate from 0.228 to 0.047.
  • Short-utterance cloning—where prior models sound unnatural—improves: on sequences of 85 or fewer phonemes, DS-TTS reports SMOS 3.72, above StyleSpeech (3.58), YourTTS (3.57), VALL-E-X (3.48), and StyleTTS2 (3.46).
  • MFCC style features are not redundant with mel features: removing them raises WER from 0.047 to 0.060 and drops SMCS from 0.865 to 0.832.
  • Both branches of the dynamic variance adapter are needed: dropping the short or long predictor raises WER to 0.066 or 0.067 and lowers SMCS.
  • The 85-phoneme threshold is presented as the optimal balance: thresholds 75, 80, 90, and 95 all give worse WER-SMCS trade-offs in the paper's sweep.

Reading between the lines

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

  • Beyond the paper: the same dual-style encoding could likely be attached to other TTS backbones without the dynamic adapter; the paper only tests it inside its own FastSpeech-style generator, so the generality is untested.
  • Beyond the paper: the 85-phoneme cutoff is likely sensitive to language and speaking rate, so deploying DS-TTS on non-English text would probably require re-tuning the threshold rather than inheriting it.
  • Beyond the paper: if the short-utterance gain is real, the most immediate practical payoff is in voice assistants, navigation prompts, and other short-command uses where single-clip cloning currently sounds choppy; the paper does not run such an evaluation.
  • Beyond the paper: the SGF gating idea could be applied to other conditioning signals such as emotion or dialect, but the paper evaluates only speaker style.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DS-TTS, a zero-shot voice cloning model combining a Dual-Style Encoding Network (mel-spectrogram and MFCC encoders) with a Dynamic Generator Network that uses a Style Gating-Film (SGF) mechanism and a length-dependent variance adapter. The adapter switches between convolutional and linear predictors according to a phoneme-length threshold (set to 85) to handle short and long sentences. The authors claim superior overall performance over StyleSpeech, YourTTS, VALL-E-X, StyleTTS2, and XTTS v2 on VCTK, based on WER, SMCS, MOS, and SMOS scores.

Significance. The paper addresses a relevant problem and proposes an interesting combination of ideas: fusing MFCC and mel-spectrogram style representations, a gating-style modulation layer, and a length-adaptive variance predictor. The empirical comparison includes several strong baselines with official checkpoints. However, the central claims are undermined by a mathematical degeneracy in the SGF equations and by a test-set tuning protocol for the threshold, which inflates the reported gains. The paper also explicitly acknowledges a higher WER than two baselines while the abstract claims improvements in WER. The proposed architecture and evaluation therefore do not currently establish the stated advantages, though the underlying questions remain worthwhile.

major comments (4)
  1. [Section III-D, Style Gating-Film] The SGF equations define gamma, beta, and eta all as tanh(f(x)) and delta as sigma(f(x)); after the weighted combination gamma = gamma*delta + eta*(1-delta) and beta = beta*delta + eta*(1-delta), both reduce algebraically to tanh(f(x)) for any input. Thus gamma and beta are always identical, so the mechanism cannot perform independent scaling and shifting of the normalized features as claimed in the text. This degeneracy invalidates the described novelty of the SGF mechanism and calls into question any performance benefit attributed to it.
  2. [Section V-B-1 and Tables I, II, IV] The phoneme-length threshold of 85 is selected by sweeping values from 75 to 95 on the same set of 108 VCTK samples that is subsequently used as the test set for Tables I and II. The reported WER of 0.047 and SMCS of 0.865 are therefore the best values chosen on the evaluation set rather than unbiased estimates. The baselines were not given an analogous threshold sweep, making the comparison asymmetric and the claimed superiority over StyleTTS2 and XTTS v2 unsupported.
  3. [Abstract, Section IV-C, Table I, Section V-A] The abstract and introduction claim that DS-TTS shows 'notable improvements in both word error rate and speaker similarity,' but Table I reports DS-TTS WER 0.047, which is worse than StyleTTS2 (0.039) and XTTS v2 (0.032). The paper's own text in Section V-A acknowledges the WER is 'slightly higher' than these baselines. The claim of improvements in both metrics is directly contradicted by the reported numbers, and the 'superior overall performance' conclusion must be revised or supported with a different evaluation protocol.
  4. [Section III-D, Dynamic Variance Adaptor] The paper states that 'Empirical analysis shows that Convolutional Layers outperform Linear Layers for short phoneme sequences,' but no such analysis is presented. The only empirical evidence is the threshold sweep in Table IV, which varies the cutoff while keeping the predictor architecture fixed; it does not compare convolutional versus linear predictors for short sequences. The claimed architectural motivation for the length-dependent switch is therefore unsupported.
minor comments (5)
  1. [Section III-D, Figure 3] The text refers to short predictors as shown in Figure 3(c) and long predictors in Figure 3(d), but the figure caption lists (c) as the long predictor and (d) as the short predictor; the cross-references are reversed.
  2. [Section IV-A] There is a typo: 'sutilized the librosa library' should be 'utilized the librosa library.'
  3. [Section IV-D and Table II] Subjective MOS and SMOS scores are reported without confidence intervals, number of raters, or statistical significance tests; this limits the interpretability of the subjective comparisons, especially when differences are small (e.g., SMOS 3.72 vs 3.59).
  4. [Section IV-A and V-A] The paper mentions extensive evaluations on LibriTTS and VCTK, but the reported experiments only evaluate on VCTK; LibriTTS is used for training. This should be stated more precisely to avoid overstating the evaluation scope.
  5. [References] References [17] and [23] are duplicates (both are the YourTTS paper by Casanova et al.); please merge or cite one instance.

Circularity Check

2 steps flagged · score 6.0 of 10

The phoneme-length threshold 85 is tuned on the VCTK test set itself (Table IV), so the reported WER/SMCS superiority is a selected maximum rather than an independent prediction; the SGF equations also define gamma=eta, making the mixing step cancel by construction.

  1. fitted input called prediction [Section V-A test set definition and Section V-B-1 / Table IV threshold sweep]
    "We randomly select 108 audio samples from different speakers from the VCTK dataset as the test set... To assess this, ablation studies are conducted by evaluating the model across a range of thresholds from 75 to 95, as shown in Table IV... The results indicate that the model with a threshold of 85 outperformed other configurations, achieving the lowest WER of 0.047 and the highest SMCS score of 0.865."

    The dynamic variance adapter's threshold is a hyperparameter selected by sweeping over 75-95 on the same 108-utterance VCTK test set that later produces Tables I and II. The reported WER of 0.047 and SMCS of 0.865 are therefore the best values found by searching on the evaluation set, not measurements of a fixed model's independent performance. The comparison is asymmetric: StyleTTS2, XTTS v2, and other baselines were evaluated with official checkpoints and no analogous test-set tuning. Thus the claimed 'superior overall performance' is partly a product of the selection procedure, so the central empirical claim reduces to a fitted maximum rather than a prediction.

  2. self definitional [Section III-D, Style Gating-Film (SGF) equations]
    "Subsequently, the style vector modulates the normalized feature information using the following parameters: γ = tanh(f (x)), β = tanh(f (x)), η = tanh(f (x)), δ = σ(f (x)) ... γ = γ · δ + η · (1 − δ) β = β · δ + η · (1 − δ)"

    By the paper's own definitions, η is identical to γ and also to β, since all three are tanh(f(x)). Consequently, the mixing expression γ·δ + η·(1−δ) algebraically reduces to γ, and β·δ + η·(1−δ) reduces to β. The claimed additional flexibility from η and δ is vacuous: SGF(y) collapses to a single scale tanh(f(x)) and shift tanh(f(x)) with no second modulation parameter. The architectural contribution of the 'Style Gating-Film' mechanism is therefore equivalent to a plain FiLM layer by construction, and the stated finer control over mean and variance is not present in the equations.

full rationale

The most significant circular step is the threshold selection: the paper tunes the 85-phoneme boundary on the exact VCTK test set used for the headline comparisons, so the final WER and SMCS are selected maxima rather than unbiased predictions. This makes the empirical superiority claim partially circular with the evaluation protocol. A second, independent issue is that the SGF modulation equations define γ, β, and η as identical functions of f(x), so the mixing step cancels and the claimed additional gating parameters have no effect by construction. No self-citations are load-bearing, and the baselines are external, but the fitted threshold and the SGF degeneracy together justify a circularity score of 6 rather than a clean bill.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a free parameter (the phoneme-length threshold) tuned to the test set, plus assumptions about feature complementarity and predictor architecture preferences that are not independently established.

free parameters (1)
  • short/long phoneme threshold = 85
    Selected by evaluating thresholds 75 to 95 on the VCTK test set (Table IV); the best WER/SMCS at 85 is used in the final model, making the reported result partly a fit to the test data.
assumptions (3)
  • domain assumption MFCC and mel-spectrogram representations provide complementary, non-redundant speaker identity information
    The DuSEN design (Section III-C) concatenates 128-d mel and 128-d MFCC embeddings; the paper provides no analysis of redundancy beyond an ablation that removes the MFCC branch.
  • ad hoc to paper A hard phoneme-length threshold of 85 is a stable property of optimal predictor architecture
    Justified only by the test-set sweep in Table IV; no cross-dataset or theoretical justification is given.
  • domain assumption Convolutional predictors are better than linear predictors for short phoneme sequences
    Stated in the introduction and used in the DVA design (Section III-D), but the supporting 'empirical analysis' is not shown; it functions as a design assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DS-TTS: Zero-Shot Speaker Style Adaptation from Voice Clips via Dynamic Dual-Style Feature Modulation." pith.science (2026). https://pith.science/paper/WKWS2FP6

@misc{pith2026250601020,
  author       = {Pith},
  title        = {Pith review of: DS-TTS: Zero-Shot Speaker Style Adaptation from Voice Clips via Dynamic Dual-Style Feature Modulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKWS2FP6}},
  note         = {Machine review of arXiv:2506.01020}
}
read the original abstract

Recent advancements in text-to-speech (TTS) technology have increased demand for personalized audio synthesis. Zero-shot voice cloning, a specialized TTS task, aims to synthesize a target speaker's voice using only a single audio sample and arbitrary text, without prior exposure to the speaker during training. This process employs pattern recognition techniques to analyze and replicate the speaker's unique vocal features. Despite progress, challenges remain in adapting to the vocal style of unseen speakers, highlighting difficulties in generalizing TTS systems to handle diverse voices while maintaining naturalness, expressiveness, and speaker fidelity. To address the challenges of unseen speaker style adaptation, we propose DS-TTS, a novel approach aimed at enhancing the synthesis of diverse, previously unheard voices. Central to our method is a Dual-Style Encoding Network (DuSEN), where two distinct style encoders capture complementary aspects of a speaker's vocal identity. These speaker-specific style vectors are seamlessly integrated into the Dynamic Generator Network (DyGN) via a Style Gating-Film (SGF) mechanism, enabling more accurate and expressive reproduction of unseen speakers' unique vocal characteristics. In addition, we introduce a Dynamic Generator Network to tackle synthesis issues that arise with varying sentence lengths. By dynamically adapting to the length of the input, this component ensures robust performance across diverse text inputs and speaker styles, significantly improving the model's ability to generalize to unseen speakers in a more natural and expressive manner. Experimental evaluations on the VCTK dataset suggest that DS-TTS demonstrates superior overall performance in voice cloning tasks compared to existing state-of-the-art models, showing notable improvements in both word error rate and speaker similarity.

Figures

Figures reproduced from arXiv: 2506.01020 by the authors.

Figure 1
Figure 1. Comparison of mel-spectrograms of different models in voice cloning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of DS-TTS, which consists of the Dual-Style Encoding Network (DuSEN) and the Dynamic Generator Network (DyGN). The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Details of some components. (a)FFT Block. (b) Variance Adaptor. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Marco-Voice Technical Report

    cs.CL 2025-08 reject novelty 4.0 of 10

    Marco-Voice is a TTS system combining voice cloning and emotional speech generation via speaker-emotion disentanglement, contrastive learning, and a new Mandarin emotional dataset, with claimed quality gains over Cosy...

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Fgp-gan: Fine- grained perception integrated generative adversarial network for expres- sive mandarin singing voice synthesis,

    X. Liu, W. Zhang, Z. Zheng, M. Pan, and R. Wang, “Fgp-gan: Fine- grained perception integrated generative adversarial network for expres- sive mandarin singing voice synthesis,” IEEE Transactions on Consumer Electronics, vol. 70, no. 3, pp. 6054–6063, 2024

    work page 2024

  2. [2]

    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,” in International Conference on Learning Representations , 2020

    work page 2020

  3. [3]

    Multilingual speech-to-speech translation system for mobile consumer devices,

    S. Yun, Y .-J. Lee, and S.-H. Kim, “Multilingual speech-to-speech translation system for mobile consumer devices,” IEEE Transactions on Consumer Electronics , vol. 60, no. 3, pp. 508–516, 2014

    work page 2014

  4. [4]

    Multi-speaker and multi-dialectal catalan tts models for video gaming,

    A. Peir ´o-Lilja, J. Giraldo, M. Llopart-Font, C. Armentano-Oller, B. K ¨ulebi, and M. Farr ´us, “Multi-speaker and multi-dialectal catalan tts models for video gaming,” in Proc. Interspeech 2024 , 2024, pp. 999– 1000. 10

    work page 2024

  5. [5]

    Neural voice cloning with a few samples,

    S. Arik, J. Chen, K. Peng, W. Ping, and Y . Zhou, “Neural voice cloning with a few samples,”Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  6. [6]

    Adaspeech: Adaptive text to speech for custom voice,

    M. Chen, X. Tan, B. Li, Y . Liu, T. Qin, S. Zhao, and T.-Y . Liu, “Adaspeech: Adaptive text to speech for custom voice,” arXiv preprint arXiv:2103.00993, 2021

    arXiv 2021

  7. [7]

    Rapid speaker adaptation in low resource text to speech systems using synthetic data and transfer learning,

    R. Joshi and N. Garera, “Rapid speaker adaptation in low resource text to speech systems using synthetic data and transfer learning,” in Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation , 2023, pp. 267–273

    work page 2023

  8. [8]

    Quantum target recognition enhancement algorithm for uav consumer applications,

    F. Yousef Alghayadh, J. Venkata Naga Ramesh, I. Keshta, M. Soni, R. Rivera, K. D. V . Prasad, A. Muhammad Soomar, and M. Tiwari, “Quantum target recognition enhancement algorithm for uav consumer applications,” IEEE Transactions on Consumer Electronics , vol. 70, no. 3, pp. 5553–5560, 2024

    work page 2024

Show all 44 references
  1. [9]

    Dual channel based speech enhancement using novelty filter for robust speech recognition in automobile environ- ment,

    J. Beh, R. Baran, and H. Ko, “Dual channel based speech enhancement using novelty filter for robust speech recognition in automobile environ- ment,” IEEE Transactions on Consumer Electronics , vol. 52, no. 2, pp. 583–589, 2006

    2006

  2. [10]

    Transfer learning from speaker verification to multispeaker text-to-speech synthesis,

    Y . Jia, Y . Zhang, R. Weiss, Q. Wang, J. Shen, F. Ren, P. Nguyen, R. Pang, I. Lopez Moreno, Y . Wu et al. , “Transfer learning from speaker verification to multispeaker text-to-speech synthesis,” Advances in neural information processing systems , vol. 31, 2018

    2018

  3. [11]

    In- vestigating on incorporating pretrained and learnable speaker represen- tations for multi-speaker multi-style text-to-speech,

    C.-M. Chien, J.-H. Lin, C.-y. Huang, P.-c. Hsu, and H.-y. Lee, “In- vestigating on incorporating pretrained and learnable speaker represen- tations for multi-speaker multi-style text-to-speech,” in ICASSP 2021- 2021 IEEE International Conference on Acoustics, Speech and Signal...

    2021

  4. [12]

    Mrmi-tts: Multi-reference audios and mutual information driven zero-shot voice cloning,

    Y . T. Chen, W. Li, and B. Tang, “Mrmi-tts: Multi-reference audios and mutual information driven zero-shot voice cloning,” ACM Transactions on Asian and Low-Resource Language Information Processing , vol. 23, no. 5, pp. 1–14, 2024

    2024

  5. [13]

    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,” in International Conference on Machine Learning . PMLR, 2021, pp. 7748–7759

    2021

  6. [14]

    Acfusion: Infrared and visible image fusion based on self-attention and convolution with enhanced information extraction,

    H. Zhu, H. Wu, D. He, R. Lan, Z. Liu, and X. Pan, “Acfusion: Infrared and visible image fusion based on self-attention and convolution with enhanced information extraction,” IEEE Transactions on Consumer Electronics, vol. 70, no. 1, pp. 4155–4167, 2024

    2024

  7. [15]

    Multi-feature fusion-based convolutional neural networks for eeg epileptic seizure prediction in con- sumer internet of things,

    I. Ahmad, M. Zhu, Z. Liu, M. Shabaz, I. Ullah, M. C. F. Tong, A. Sambas, L. Men, Y . Chen, and S. Chen, “Multi-feature fusion-based convolutional neural networks for eeg epileptic seizure prediction in con- sumer internet of things,” IEEE Transactions on Consumer Electronics ,...

    2024

  8. [16]

    Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,

    Y . A. Li, C. Han, V . Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  9. [18]

    Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech,

    R. Huang, Y . Ren, J. Liu, C. Cui, and Z. Zhao, “Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech,” Advances in Neural Information Processing Systems , vol. 35, pp. 10 970–10 983, 2022

    2022

  10. [19]

    Openvoice: Versatile instant voice cloning,

    Z. Qin, W. Zhao, X. Yu, and X. Sun, “Openvoice: Versatile instant voice cloning,” arXiv preprint arXiv:2312.01479 , 2023

    2023 arXiv

  11. [20]

    Towards end-to-end prosody transfer for expressive speech synthesis with tacotron,

    R. Skerry-Ryan, E. Battenberg, Y . Xiao, Y . Wang, D. Stanton, J. Shor, R. Weiss, R. Clark, and R. A. Saurous, “Towards end-to-end prosody transfer for expressive speech synthesis with tacotron,” in international conference on machine learning . PMLR, 2018, pp. 4693–4702

    2018

  12. [21]

    Arbitrary style transfer in real-time with adaptive instance normalization,

    X. Huang and S. Belongie, “Arbitrary style transfer in real-time with adaptive instance normalization,” in Proceedings of the IEEE interna- tional conference on computer vision , 2017, pp. 1501–1510

    2017

  13. [22]

    One-shot voice conversion by separating speaker and content representations with instance normalization,

    J.-c. Chou and H.-Y . Lee, “One-shot voice conversion by separating speaker and content representations with instance normalization,” Inter- speech 2019, 2019

    2019

  14. [23]

    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,” in International Conference on Machine Learning. PMLR, 2022, pp. 2709–2720

    2022

  15. [24]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” in Proceedings of the AAAI conference on artificial intelligence , vol. 32, no. 1, 2018

    2018

  16. [25]

    Dynamic neural networks: A survey,

    Y . Han, G. Huang, S. Song, L. Yang, H. Wang, and Y . Wang, “Dynamic neural networks: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7436–7456, 2021

    2021

  17. [26]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural informa- tion processing systems , vol. 25, 2012

    2012

  18. [27]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    2014 arXiv

  19. [28]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 1–9

    2015

  20. [29]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo...

    2019

  21. [30]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  22. [31]

    Skipnet: Learning dynamic routing in convolutional networks,

    X. Wang, F. Yu, Z.-Y . Dou, T. Darrell, and J. E. Gonzalez, “Skipnet: Learning dynamic routing in convolutional networks,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 409– 424

    2018

  23. [32]

    Multi-scale dense networks for resource efficient image classification,

    G. Huang, D. Chen, T. Li, F. Wu, L. van der Maaten, and K. Weinberger, “Multi-scale dense networks for resource efficient image classification,” in International Conference on Learning Representations , 2018

    2018

  24. [33]

    Reso- lution adaptive networks for efficient inference,

    L. Yang, Y . Han, X. Chen, S. Song, J. Dai, and G. Huang, “Reso- lution adaptive networks for efficient inference,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2369–2378

    2020

  25. [34]

    Any-precision deep neural networks,

    H. Yu, H. Li, H. Shi, T. S. Huang, and G. Hua, “Any-precision deep neural networks,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 10 763–10 771

    2021

  26. [35]

    Big/little deep neural network for ultra low power inference,

    E. Park, D. Kim, S. Kim, Y .-D. Kim, G. Kim, S. Yoon, and S. Yoo, “Big/little deep neural network for ultra low power inference,” in 2015 international conference on hardware/software codesign and system synthesis. IEEE, 2015, pp. 124–132

    2015

  27. [36]

    A convolutional neural network cascade for face detection,

    H. Li, Z. Lin, X. Shen, J. Brandt, and G. Hua, “A convolutional neural network cascade for face detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 5325– 5334

    2015

  28. [37]

    Changing model behavior at test- time using reinforcement learning,

    A. Odena, D. Lawson, and C. Olah, “Changing model behavior at test- time using reinforcement learning,” arXiv preprint arXiv:1702.07780 , 2017

    2017 arXiv

  29. [38]

    Dynamic deep neural networks: Optimizing accuracy-efficiency trade-offs by selective execution,

    L. Liu and J. Deng, “Dynamic deep neural networks: Optimizing accuracy-efficiency trade-offs by selective execution,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 32, no. 1, 2018

    2018

  30. [39]

    Speech emotion recognition with co-attention based multi-level acoustic information,

    H. Zou, Y . Si, C. Chen, D. Rajan, and E. S. Chng, “Speech emotion recognition with co-attention based multi-level acoustic information,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 7367–7371

    2022

  31. [40]

    Task-adaptive neural process for user cold-start recommendation,

    X. Lin, J. Wu, C. Zhou, S. Pan, Y . Cao, and B. Wang, “Task-adaptive neural process for user cold-start recommendation,” in Proceedings of the Web Conference 2021 , 2021, pp. 1306–1316

    2021

  32. [41]

    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,” arXiv preprint arXiv:1904.02882 , 2019

    1904 arXiv

  33. [42]

    Melgan: Generative adversarial networks for conditional waveform synthesis,

    K. Kumar, R. Kumar, T. De Boissiere, L. Gestin, W. Z. Teoh, J. Sotelo, A. De Brebisson, Y . Bengio, and A. C. Courville, “Melgan: Generative adversarial networks for conditional waveform synthesis,” Advances in neural information processing systems , vol. 32, 2019

    2019

  34. [43]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

    Z. Zhang, L. Zhou, C. Wang, S. Chen, Y . Wu, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,” arXiv preprint arXiv:2303.03926, 2023

    2023 arXiv

  35. [44]

    Neural codec language models are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al., “Neural codec language models are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111 , 2023

    2023 arXiv

  36. [45]

    Xtts: a massively multilin- gual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Alja- fari, J. Meyer, R. Morais, S. Olayemi et al., “Xtts: a massively multilin- gual zero-shot text-to-speech model,” arXiv preprint arXiv:2406.04904, 2024

    2024 arXiv

Pith tools

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