Pith. sign in

REVIEW 4 major objections 5 minor 32 references

ReFlow-VC: Zero-shot Voice Conversion Based on Rectified Flow and Speaker Feature Optimization

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 ReFlow-VC claims that rectified flow turns zero-shot voice conversion into an ODE solvable in one sampling step, and that refining the target speaker embedding with the source utterance's content and pitch closes the quality gap to…

desk verdict Solid rectified-flow VC system with a load-bearing but fixable flaw: the zero-shot claim isn't tested because the split is by utterance, not by speaker. read the letter →

arxiv 2506.01032 v1 pith:IMK3BVED submitted 2025-06-01 cs.SD eess.AS

Pengyu Ren , Wenhao Guan , Kaidi Wang , Peijie Chen , Qingyang Hong , Lin Li This is my paper
classification cs.SDeess.AS
keywords zero-shotvoiceconversionrectifiedflowODE-basedspeechgenerationspeakerfeaturefusioncross-attentionmel-spectrogramsimilaritysmall-datatraining
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

ReFlow-VC proposes that rectified flow — an ODE that carries Gaussian noise to the target mel-spectrogram distribution along nearly straight trajectories — is sufficient for high-quality zero-shot voice conversion even when trained on a small dataset of 38.13 hours. It further proposes that the target speaker embedding should not be treated as a static condition but dynamically refined by cross-attention and gated fusion with the source utterance's content and pitch features. If these claims hold, voice conversion can run in as few as one sampling step while matching the quality of a 30-step diffusion baseline, and reach the best naturalness and speaker-similarity scores among the compared systems at higher step counts. That would make real-time, personalized voice conversion practical without large training corpora.

What carries the argument

The central object is the rectified flow ODE, which learns a drift $v_\theta(X_t,t,c)$ by regressing $(X_1-X_0)$ so that generation becomes solving $dZ_t=v_\theta(Z_t,t,c)\,dt$ from Gaussian noise to mel-spectrograms. Because the learned paths are nearly straight, the Euler solver with one step already works, while the adaptive RK45 solver refines quality; the recursive 2-ReFlow-VC straightens the trajectories once more. The second machinery is the fusion encoder: the average voice encoder's mel, Hubert-Soft continuous content features, and VQ-VAE pitch representations are aligned with the target speaker embedding through cross-attention, multihead attention, gated fusion, and iterative self-attention, producing a speaker condition that tracks the current utterance's content and pitch instead of a static embedding.

What would settle it

Take a held-out set of unseen speaker pairs, run ReFlow-VC, and measure the speaker-encoder cosine similarity of each converted sample to the target reference and to the source reference; if a substantial fraction of samples are closer to the source than to the target, the disentanglement premise fails and the zero-shot claim collapses. A simpler human version is a listening test in which listeners are asked to identify the source speaker from the converted audio — if they succeed at high rates, the source timbre has leaked through.

Watch

Extended reading notes

Core claim

The central claim is that a rectified-flow ODE model, trained with a single least-squares loss on the velocity field $X_1-X_0$ along the linear interpolation $X_t=tX_1+(1-t)X_0$, outperforms diffusion-based voice conversion while requiring dramatically fewer sampling steps, and that the speaker condition should be optimized per utterance through a fusion module. In the reported experiments on a small LibriTTS subset, ReFlow-VC with the RK45 ODE solver achieves the highest NMOS (3.78), SMOS (2.81), and SECS (0.843) among AutoVC, Free-VC, and Diff-VC at 1, 30, and 1000 steps; at a single Euler step it attains SECS 0.830 with CER and WER lower than Diff-VC at any step count. The paper also shows that the cross-attention and gated fusion module, which adjusts speaker features using content and pitch, is what lifts speaker similarity well above the no-fusion ablation, and that recursively re-running the rectification procedure (2-ReFlow-VC) yields only a marginal further gain at one step.

Load-bearing premise

The load-bearing premise is that the average-mel, Hubert-Soft content, and VQ-VAE pitch features extracted from the source speech are speaker-independent enough that fusing them with the target embedding cannot leak the source speaker's timbre.

Editorial extensions

If this is right

  • With the RK45 solver, a single ReFlow-VC model reaches the highest naturalness, similarity, and SECS among all compared systems, so the quality ceiling of diffusion-based VC does not require thousands of denoising steps.
  • At one Euler step, ReFlow-VC matches Diff-VC at 30 steps on naturalness and similarity, which implies near-real-time conversion on modest hardware such as a single consumer GPU.
  • The fusion module improves SECS from 0.751 (no fusion, 1 step) to 0.830, so using content and pitch to refine the speaker embedding is the mechanism that drives speaker similarity.
  • The 2-ReFlow-VC result shows the recursive rectification procedure transfers to voice conversion and slightly improves one-step sampling, confirming the conditioning does not break the straightening property.
  • Training on 38 hours of LibriTTS suffices for zero-shot conversion to unseen speakers in these evaluations, indicating small-data viability for practical deployment.

Reading between the lines

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

  • A natural next experiment the paper does not run is to swap the Hubert-Soft content extractor for discrete HuBERT units or a different pitch representation; if SECS stays at 0.83, the fusion module is representation-agnostic, while a drop would show the method inherits its ceiling from the frozen feature extractors.
  • Because CER and WER remain higher than AutoVC despite superior similarity, an untested extension would add an explicit content-preservation term (for example an ASR loss on the converted mel) to see whether similarity can be retained while closing the content gap.
  • The one-step quality claim suggests the same rectified-flow decoder could be applied to other conditional audio generation tasks such as singing voice conversion or emotion conversion, where multi-step diffusion has been the bottleneck; this is an extrapolation, not a paper claim.
  • All experiments come from LibriTTS read speech; testing on noisy, reverberant, or conversational recordings would tell whether 'zero-shot' generalizes beyond clean audiobook conditions, which remains an open empirical question.
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 ReFlow-VC, a zero-shot voice conversion method based on a rectified flow ODE. The model is trained with the standard rectified flow objective in Eq. (5), converting Gaussian noise into mel-spectrograms conditioned on a target speaker embedding and on a set of source speech features (average-mel from an average voice encoder, HuBERT-Soft content features, and VQ-VAE pitch features). A 'feature fusion module' combines these conditions through cross-attention and gating. The decoder is a U-Net with a GradTTS-style architecture. Experiments on a small LibriTTS subset compare ReFlow-VC with Diff-VC, AutoVC, and Free-VC using NMOS, SMOS, CER, WER, SECS, and sampling time. The paper reports that ReFlow-VC outperforms baselines on speaker similarity, is competitive with Diff-VC at vastly fewer sampling steps, and that a 2-ReFlow-VC reflow variant behaves similarly.

Significance. If the claims hold, the paper makes a useful practical contribution: a rectified-flow voice conversion model that produces high-quality converted speech with one-step or few-step sampling on a small training set. The core training objective (Eq. 5) is the well-established rectified flow loss, the decoder builds on the known GradTTS U-Net, and the authors provide objective metrics, MOS evaluations, and a public demo page. The reported speed advantage over Diff-VC is a concrete and potentially valuable result. However, the central 'zero-shot' claim is not actually tested by the current experimental design, because the data split is by utterance rather than by speaker; in addition, the baseline comparison is not fully controlled. These issues must be addressed before the main claims can be accepted.

major comments (4)
  1. [§4.1.1] The zero-shot claim is not supported by the experimental setup. The paper states that the LibriTTS subset was 'randomly split into a training set (26,000 samples), a validation set (80 samples), and a test set (500 samples)' at the sample level, but LibriTTS contains multiple utterances per speaker. No speaker counts per split are reported, and no verification is given that the 500 test utterances come from speakers unseen during training. Without a speaker-disjoint split, the high SMOS and SECS values in Table 1 could be achieved by memorizing speaker embeddings already encountered in training, and the conclusion that the model 'generalize[s] effectively to any unseen speaker' is untested. Please retrain/evaluate on a speaker-disjoint split, report the number of speakers in each split, and state explicitly whether any test speaker appears in the training set.
  2. [§4.1.3] The comparison against AutoVC and Free-VC is not controlled. The text explicitly says that Diff-VC was retrained on the same small dataset, but no such statement is made for AutoVC or Free-VC, and the provided GitHub links are to repositories that distribute pretrained models. Consequently, the improvements over AutoVC and Free-VC in Table 1 and the conclusions drawn in §4.2 confound the proposed method with the amount and domain of training data. Please either retrain all baselines on the same dataset and feature setup, or clearly label the pretrained baselines and restrict the comparison claims to the systems that were trained under identical conditions.
  3. [§3.1] The feature-fusion module, which is one of the two stated contributions, is described only qualitatively. The text mentions 'two layers of cross-attention', 'gated fusion', 'iterative self-attention', and a multihead-attention layer, but provides no equations, no tensor shapes, and no precise description of how the speaker embedding is modified by the content and pitch features. This makes it impossible to assess formally what 'optimizing speaker features' means and prevents reproduction of the method. Please provide a formal specification (equations and tensor dimensions) or pseudocode for the fusion encoder, and specify which components are ablated in NReFlow-VC.
  4. [§3.1] The disentanglement assumption is load-bearing but not directly validated. The method assumes that the average-mel, HuBERT-Soft content features, and normalized pitch representations are sufficiently speaker-independent that the fusion module can inject only target timbre. The paper does not report any leakage analysis, such as speaker-identification accuracy on the intermediate features or a source-similarity control (e.g., SECS between the converted output and the source speaker). Without such a control, it is unclear whether the high target-similarity scores reflect genuine transfer of target timbre or a failure to fully remove source timbre. Please add a source-similarity or feature-leakage experiment.
minor comments (5)
  1. [§7 (References)] Reference [25] is incorrect: the Whisper ASR model used for WER/CER should be cited as Radford et al., 'Robust Speech Recognition via Large-Scale Weak Supervision' (2022). The currently cited paper, 'Whisper: Tracing the Spatiotemporal Process of Information Diffusion in Real Time', is an unrelated information-diffusion visualization paper.
  2. [Tables 1 and 3] The iteration counts for the RK45 solver are inconsistent: Table 1 lists 512 iterations for ReFlow-VC(RK45) and 146 for NReFlow-VC(RK45), while Table 3 lists 554 and 140, respectively. Please unify these numbers and explain whether the difference is due to stopping criteria or reporting error.
  3. [§3.1] There are typos in the feature names: 'Hubert-Soft' should be 'HuBERT-Soft', 'VQ-V AE' should be 'VQ-VAE', and 'YAPPT' should likely be 'YAAPT'. Please correct these throughout.
  4. [§4.2] The paper claims that one-step ReFlow-VC is 'on par with' 30-step Diff-VC, but this is based on overlapping MOS confidence intervals and single-run SECS values. Please report statistical significance tests or multiple seeds for at least SECS and the one-step comparison.
  5. [§4.1.1] Please specify the exact pretrained HiFi-GAN checkpoint used as the vocoder, since vocoder choice can materially affect NMOS and SMOS and is an uncontrolled variable in the comparison with pretrained baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the rectified-flow objective and the speaker-feature fusion are standard, independently evaluated components, and no claimed prediction reduces to a fitted input or to an author-supplied uniqueness condition.

full rationale

The core derivation is self-contained. Equations (4)-(5) are the standard rectified-flow least-squares objective from Liu et al. [15], applied with X0 as standard Gaussian noise and X1 as the ground-truth mel-spectrogram; the objective contains no fitted constants, no data-dependent normalization, and no term that presupposes the converted output, so the generative result is not equivalent to the training loss by construction. The 2-ReFlow-VC variant is the standard reflow procedure of retraining on generated samples, not a circular reuse of the target. The speaker-feature optimization module (cross-attention, gated fusion, self-attention) is a learned conditioning network; its outputs are not claimed to be predictions derived from their own loss. The central claims, including the best SMOS/NMOS/SECS scores and one-step performance on par with Diff-VC's 30-step sampling, are evaluated against external metrics in Tables 1-3 and against re-trained baselines, so they do not reduce to fitted parameters renamed as predictions. The only self-citation ([13], ReFlow-TTS, with overlapping authors) is used to motivate rectified flow and Euler one-step sampling, but the paper's own Table 1 demonstrates the one-step result, so the citation is not load-bearing. One limitation, flagged for completeness, is not circularity: Section 4.1.1 states the dataset was 'randomly split into a training set (26,000 samples), a validation set (80 samples), and a test set (500 samples)' by utterance rather than by speaker, so the zero-shot/unseen-speaker claim is not strictly verified by the reported experiment; this is an experimental-design concern, not a reduction of the derivation to its inputs.

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

The paper introduces no new free parameters fitted to data or invented physical/formal entities. It relies on standard rectified flow theory and domain assumptions about speech feature disentanglement. The architecture hyperparameters (channel multiplier 4, 256-dimensional fusion) are design choices, not fitted parameters.

assumptions (3)
  • standard math The least-squares objective in Eq. (5) trains a valid conditional ODE transport between Gaussian and mel-spectrogram distributions, per the rectified flow theory of [15].
    Invoked without proof in Sections 2.1 and 2.2; the paper relies on the correctness of rectified flow training.
  • domain assumption The average voice encoder output, Hubert-Soft content, and normalized pitch are speaker-independent and sufficient to preserve linguistic content during conversion.
    This is assumed in Section 3.1; if any of these features leak speaker identity or lose content, the zero-shot conversion fails.
  • domain assumption The 80-dimensional mel-spectrogram with HiFi-GAN inversion faithfully represents speech for the purpose of voice conversion.
    Used implicitly throughout the experiments; the choice of mel parameters and vocoder affects the upper bound on quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReFlow-VC: Zero-shot Voice Conversion Based on Rectified Flow and Speaker Feature Optimization." pith.science (2026). https://pith.science/paper/IMK3BVED

@misc{pith2026250601032,
  author       = {Pith},
  title        = {Pith review of: ReFlow-VC: Zero-shot Voice Conversion Based on Rectified Flow and Speaker Feature Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMK3BVED}},
  note         = {Machine review of arXiv:2506.01032}
}
read the original abstract

In recent years, diffusion-based generative models have demonstrated remarkable performance in speech conversion, including Denoising Diffusion Probabilistic Models (DDPM) and others. However, the advantages of these models come at the cost of requiring a large number of sampling steps. This limitation hinders their practical application in real-world scenarios. In this paper, we introduce ReFlow-VC, a novel high-fidelity speech conversion method based on rectified flow. Specifically, ReFlow-VC is an Ordinary Differential Equation (ODE) model that transforms a Gaussian distribution to the true Mel-spectrogram distribution along the most direct path. Furthermore, we propose a modeling approach that optimizes speaker features by utilizing both content and pitch information, allowing speaker features to reflect the properties of the current speech more accurately. Experimental results show that ReFlow-VC performs exceptionally well in small datasets and zero-shot scenarios.

Figures

Figures reproduced from arXiv: 2506.01032 by the authors.

Figure 1
Figure 1. The diagram of the Rectified Flow model, with the meanings of the icons at the top.(a) Linear interpolation of data samples (X0, X1).(b) The rectified flow Zt induced by (X0, X1).(c) The linear interpolation of data samples (Z0, Z1) of rectified flow Zt.(d) The rectified flow induced from (Z0, Z1), propagating along straight paths. Gaussian) to π1 (the ground truth distribution) via straight-line paths. Given sample… view at source ↗
Figure 2
Figure 2. An illustration of ReFlow-VC. uralness of voice conversion, ensuring clear and natural speech across different speakers through the use of multiple attention and self-attention mechanisms. These strategies significantly improve the expressiveness of the speaker features, resulting in more precise and natural voice conversion outcomes. Specifically, the fusion encoder consists of multiple mod￾ules designed to enhance… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

  1. [19]

    Diffusion-based voice conversion with fast maximum likelihood sampling scheme,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, M. Kudinov, and J. Wei, “Diffusion-based voice conversion with fast maximum likelihood sampling scheme,”arXiv preprint arXiv:2109.13821, 2021

    arXiv 2021

  2. [1]

    VC achieves this by decom- posing the source speech into different components, including the speaker’s timbre, linguistic content, and speaking style

    Introduction Zero-shot voice conversion (VC) aims to convert speech from any source speaker to the speech of any target speaker without changing the linguistic content. VC achieves this by decom- posing the source speech into different components, including the speaker’s timbre, linguistic content, and speaking style. Its applications span across various ...

    work page

  3. [2]

    Rectified Flow Model 2.1. Rectified Flow The rectified flow model is an Ordinary Differential Equation (ODE) model designed to transform a distributionπ 0 (standard arXiv:2506.01032v1 [cs.SD] 1 Jun 2025 Figure 1:The diagram of the Rectified Flow model, with the meanings of the icons at the top.(a) Linear interpolation of data samples (X0,X 1).(b) The rect...

    work page Pith review arXiv 2025

  4. [3]

    Encoder The encoder consists of an average voice encoder, Hubert-Soft [16], a speaker encoder, VQ-V AE, and a feature fusion mod- ule

    ReFlow-VC Model Architecture 3.1. Encoder The encoder consists of an average voice encoder, Hubert-Soft [16], a speaker encoder, VQ-V AE, and a feature fusion mod- ule. We chose the average phoneme-level MEL features as the speaker-independent speech representation, similar to Diff-VC [12].By using the average voice encoder, the source audio is transforme...

    work page

  5. [4]

    Experimental Setup 4.1.1

    Experiments And Results 4.1. Experimental Setup 4.1.1. Dataset We trained the proposed ReFlow-VC model using a subset of the LibriTTS dataset, which was randomly sampled from the full LibriTTS dataset to facilitate model training on a small dataset [22]. This subset consists of 26,580 speech samples with a total duration of 38.13 hours. The dataset was ra...

    work page

  6. [5]

    ReFlow-VC can use the RK45 ODE solver for sampling, generating speech samples with optimal audio quality

    Conclusions In this paper, we propose a simple yet efficient ReFlow-VC, which effectively completes the voice conversion task by lever- aging rectified flow and feature fusion techniques. ReFlow-VC can use the RK45 ODE solver for sampling, generating speech samples with optimal audio quality. Furthermore, due to the excellent one-step sampling capability ...

    work page

  7. [6]

    Acknowledgements This work was supported in part by the National Natural Science Foundation of China under Grants 62276220 and 62371407 and the Innovation of Policing Science and Technol- ogy, Fujian province (Grant number: 2024Y0068)

    work page

  8. [7]

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

    T. Kaneko and H. Kameoka, “Cyclegan-vc: Non-parallel voice conversion using cycle-consistent adversarial networks,” in 2018 26th European Signal Processing Conference (EUSIPCO). IEEE, 2018, pp. 2100–2104

    work page 2018

Show all 32 references
  1. [8]

    Privacy and utility of x-vector based speaker anonymiza- tion,

    B. M. L. Srivastava, M. Maouche, M. Sahidullah, E. Vincent, A. Bellet, M. Tommasi, N. Tomashenko, X. Wang, and J. Yam- agishi, “Privacy and utility of x-vector based speaker anonymiza- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2383–2395, 2022

    2022

  2. [9]

    Autovc: Zero-shot voice style transfer with only au- toencoder loss,

    K. Qian, Y . Zhang, S. Chang, X. Yang, and M. Hasegawa- Johnson, “Autovc: Zero-shot voice style transfer with only au- toencoder loss,” inInternational Conference on Machine Learn- ing. PMLR, 2019, pp. 5210–5219

    2019

  3. [10]

    Again-vc: A one-shot voice conversion using activation guidance and adap- tive instance normalization,

    Y .-H. Chen, D.-Y . Wu, T.-H. Wu, and H.-y. Lee, “Again-vc: A one-shot voice conversion using activation guidance and adap- tive instance normalization,” inICASSP 2021-2021 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5954–5958

    2021

  4. [11]

    Gemo-clap: Gender-attribute-enhanced contrastive language- audio pretraining for accurate speech emotion recognition,

    Y . Pan, Y . Hu, Y . Yang, W. Fei, J. Yao, H. Lu, L. Ma, and J. Zhao, “Gemo-clap: Gender-attribute-enhanced contrastive language- audio pretraining for accurate speech emotion recognition,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Pro...

    2024

  5. [12]

    Ace- vc: Adaptive and controllable voice conversion using explicitly disentangled self-supervised speech representations,

    S. Hussain, P. Neekhara, J. Huang, J. Li, and B. Ginsburg, “Ace- vc: Adaptive and controllable voice conversion using explicitly disentangled self-supervised speech representations,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (I...

    2023

  6. [13]

    Non-parallel voice conversion with cyclic variational autoen- coder,

    P. L. Tobing, Y .-C. Wu, T. Hayashi, K. Kobayashi, and T. Toda, “Non-parallel voice conversion with cyclic variational autoen- coder,”arXiv preprint arXiv:1907.10185, 2019

    1907 arXiv

  7. [14]

    Dddm-vc: Decoupled denoising diffusion models with disentangled representation and prior mixup for verified robust voice conversion,

    H.-Y . Choi, S.-H. Lee, and S.-W. Lee, “Dddm-vc: Decoupled denoising diffusion models with disentangled representation and prior mixup for verified robust voice conversion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 16, 2024, pp. 17 862–17 870

    2024

  8. [15]

    Generative adver- sarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde- Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adver- sarial networks,”Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    2020

  9. [16]

    Cyclegan- vc2: Improved cyclegan-based non-parallel voice conversion,

    T. Kaneko, H. Kameoka, K. Tanaka, and N. Hojo, “Cyclegan- vc2: Improved cyclegan-based non-parallel voice conversion,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 6820– 6824

    2019

  10. [17]

    Stargan-vc: Non-parallel many-to-many voice conversion using star genera- tive adversarial networks,

    H. Kameoka, T. Kaneko, K. Tanaka, and N. Hojo, “Stargan-vc: Non-parallel many-to-many voice conversion using star genera- tive adversarial networks,” in2018 IEEE Spoken Language Tech- nology Workshop (SLT). IEEE, 2018, pp. 266–273

    2018

  11. [18]

    Freevc: Towards high-quality text-free one-shot voice conversion,

    J. Li, W. Tu, and L. Xiao, “Freevc: Towards high-quality text-free one-shot voice conversion,” inICASSP 2023-2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  12. [20]

    Reflow-tts: A rectified flow model for high-fidelity text-to- speech,

    W. Guan, Q. Su, H. Zhou, S. Miao, X. Xie, L. Li, and Q. Hong, “Reflow-tts: A rectified flow model for high-fidelity text-to- speech,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 501–10 505

    2024

  13. [21]

    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

  14. [22]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  15. [23]

    A comparison of discrete and soft speech units for improved voice conversion,

    B. Van Niekerk, M.-A. Carbonneau, J. Za ¨ıdi, M. Baas, H. Seut ´e, and H. Kamper, “A comparison of discrete and soft speech units for improved voice conversion,” inICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 202...

    2022

  16. [24]

    Speech resynthesis from dis- crete disentangled self-supervised representations,

    A. Polyak, Y . Adi, J. Copet, E. Kharitonov, K. Lakhotia, W.-N. Hsu, A. Mohamed, and E. Dupoux, “Speech resynthesis from dis- crete disentangled self-supervised representations,”arXiv preprint arXiv:2104.00355, 2021

    2021 arXiv

  17. [25]

    Zsvc: Zero-shot style voice conversion with disentangled la- tent diffusion models and adversarial training,

    X. Zhu, L. He, Y . Xiao, X. Wang, X. Tan, S. Zhao, and L. Xie, “Zsvc: Zero-shot style voice conversion with disentangled la- tent diffusion models and adversarial training,”arXiv preprint arXiv:2501.04416, 2025

    2025 arXiv

  18. [26]

    Ctefm-vc: Zero-shot voice conversion based on content-aware timbre ensemble modeling and flow matching,

    Y . Pan, Y . Yang, J. Yao, J. Ye, H. Zhou, L. Ma, and J. Zhao, “Ctefm-vc: Zero-shot voice conversion based on content-aware timbre ensemble modeling and flow matching,”arXiv preprint arXiv:2411.02026, 2024

    2024 arXiv

  19. [27]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Sp...

    2015

  20. [29]

    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

  21. [30]

    Adam: A method for stochastic optimization,

    D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  22. [31]

    Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,”Advances in neural information processing systems, vol. 33, pp. 17 022– 17 033, 2020

    2020

  23. [32]

    Whis- per: Tracing the spatiotemporal process of information diffusion in real time,

    N. Cao, Y .-R. Lin, X. Sun, D. Lazer, S. Liu, and H. Qu, “Whis- per: Tracing the spatiotemporal process of information diffusion in real time,”IEEE transactions on visualization and computer graphics, vol. 18, no. 12, pp. 2649–2658, 2012

    2012

  24. [256]

    We used the pre-trained HiFi-GAN [24] as the neural vocoder, which is responsible for converting the mel spectrograms into raw wave- forms

    ReFlow-VC was trained for 200K iterations on a single NVIDIA 2080Ti GPU using the Adam optimizer [23]. We used the pre-trained HiFi-GAN [24] as the neural vocoder, which is responsible for converting the mel spectrograms into raw wave- forms. 4.1.2. Evaluation Metrics We condu...

Pith tools

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