Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Noro: Noise-Robust One-shot Voice Conversion with Hidden Speaker Representation Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Noise-agnostic reference embeddings keep one-shot voice conversion intelligible even when the target speaker's clip is noisy.

desk verdict A competent, readable paper on a real problem whose central attribution is undercut by a missing ablation: the noise-robustness gain may come from simple reference-path augmentation, not the dual-branch encoder and contrastive loss. read the letter →

arxiv 2411.19770 v2 pith:UBAVURW2 submitted 2024-11-29 cs.SD cs.CLeess.AS

Haorui He , Yuchen Song , Yuancheng Wang , Haoyang Li , Xueyao Zhang , Li Wang , Gongping Huang , Eng Siong Chng
show 1 more author
Zhizheng Wu
This is my paper · ORCID
classification cs.SDcs.CLeess.AS
keywords one-shotvoiceconversionnoiserobustnessdual-branchreferenceencodernoise-agnosticcontrastivespeakerlossrepresentationlearningself-superviseddiffusionmodelverification
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 argues that the standard approach of feeding a speaker-reference embedding from a noisy clip to a diffusion-based voice converter is the main failure point in real-world use, and that the failure can be fixed by making the reference embedding itself noise-invariant. Noro does this with a weight-shared dual-branch reference encoder that sees both a clean reference and a stochastically noise-corrupted copy, together with a contrastive speaker loss that forces the two branches to converge. At 0–5 dB SNR, Noro's character error rate stays near the clean level (4.66 versus 7.26 for the baseline) and speaker similarity improves substantially, while the clean-condition numbers are essentially unchanged. The same reference encoder, repurposed directly as a speaker encoder without fine-tuning, reaches a 5.32% equal error rate on a standard speaker-verification task, which the paper presents as evidence that one-shot VC training doubles as self-supervised speaker representation learning.

What carries the argument

The load-bearing mechanism is the dual-branch reference encoding module with the noise-agnostic contrastive speaker loss. Two transformer encoders share weights: one encodes the clean reference mel-spectrogram, the other encodes the same utterance corrupted by one of eight DEMAND noise types at an SNR sampled from a normal distribution with mean 0 dB and standard deviation 20 dB. The two utterance-level embeddings are averaged and then fed to the WaveNet diffusion model as conditioning, and a contrastive loss over the concatenated batch pulls clean and noisy embeddings of the same speaker together while pushing different speakers apart. At inference only one encoder branch is used, so the deployed model has the same structure and cost as the baseline. The paper attributes Noro's robustness to this noise-agnostic reference representation.

What would settle it

A concrete falsifying experiment: evaluate Noro on reference clips corrupted by non-stationary noise types not in DEMAND (e.g., street babble or music) at SNR below 0 dB. If the character error rate jumps back to the baseline's noisy-condition level, the claimed noise-agnostic robustness is limited to the trained noise distribution; likewise, ablating the contrastive loss while keeping the dual-branch averaging would show how much of the gain comes from the loss itself.

Watch

Extended reading notes

Core claim

The central discovery is that noise robustness in one-shot voice conversion can be trained into the reference representation rather than bolted on as enhancement. The paper's Noro system trains a dual-branch reference encoder with shared weights on clean and noisy versions of the same reference speech, feeding the average of the two representations to the diffusion model and supervising the branches with a noise-agnostic contrastive speaker loss. After this training, clean and noisy reference speech produce reference embeddings that are mixed in t-SNE space instead of separated by noise condition, and at inference the model uses only a single encoder branch. On the VCTK test set with unseen noise types at 0–5 dB SNR, Noro achieves CER 4.66 and SECS 80.09 versus the baseline's 7.26 and 77.28, while in clean conditions the two systems are nearly identical (CER 4.74 vs 4.71, SECS 82.38 vs 82.35). The paper also reports that the baseline reference encoder, evaluated without adaptation as a speaker encoder called VC-SPK2VEC, attains an EER of 5.32% on the SUPERB speaker-verification benchmark, beating several established self-supervised speech models.

Load-bearing premise

The load-bearing premise is that noise invariance learned from eight DEMAND noise types at SNRs centered on 0 dB transfers to unseen real-world noises, and that the train-time averaging of clean and noisy reference embeddings is harmless when only a single branch is used at inference; the paper does not analyze the latter mismatch explicitly.

Editorial extensions

If this is right

  • In noisy conditions (0–5 dB SNR), Noro's character error rate of 4.66 is nearly identical to its clean-condition 4.74, meaning intelligibility no longer collapses when the reference clip is degraded.
  • Speaker similarity under noise (SECS 80.09) approaches the clean-condition 82.38, so the converted voice keeps the target identity even from a noisy prompt.
  • Subjective scores (CMOS 2.95 vs 2.09 at 0–5 dB) confirm human listeners perceive Noro as substantially more natural than the baseline in noise.
  • Since the noisy test set uses noise types unseen in training, the improvement indicates the learned noise invariance generalizes beyond the augmentation set.
  • A one-shot VC reference encoder trained without speaker labels can serve as a competitive self-supervised speaker encoder (EER 5.32%), suggesting that large unlabeled speech corpora used for VC can also yield speaker representations.

Reading between the lines

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

  • The paper leaves the train-inference mismatch implicit: during training the diffusion model conditions on the average of clean and noisy reference embeddings, while at inference it sees only a single branch's output. Testing stochastic branch selection or clean-only conditioning during training would reveal whether this mismatch costs robustness.
  • The noise-agnostic contrastive loss is a general recipe for condition-invariant reference embeddings, so it could transfer to other reference-conditioned generative models (e.g., text-to-speech) or to robust speaker verification in noisy conditions.
  • The reported connection between VC quality and speaker-embedding quality suggests a mutual-training loop: improving the reference encoder for generation could improve downstream speaker tasks, and vice versa, but the paper does not test this synergy explicitly.
  • An obvious stress test beyond the paper's scope is SNR below 0 dB or non-stationary noise like cafeteria babble; the current normal(0,20) dB augmentation schedule covers moderate conditions and the paper's robustness claim would be stronger if it held there.
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 introduces Noro, a noise-robust one-shot voice conversion system built on a diffusion-based baseline. Noro replaces the baseline's reference encoder with a dual-branch module that shares weights, processes clean and noisy reference speeches, and feeds their averaged representation to the acoustic model during training; it also adds a noise-agnostic contrastive speaker loss. In a second contribution, the authors repurpose the baseline reference encoder as an SSL speaker encoder (VC-SPK2VEC) and evaluate it on the SUPERB speaker verification task. Experiments compare Noro against the baseline and three one-shot VC systems under clean and noisy conditions, reporting objective (CER, SECS) and subjective (CMOS, SMOS) metrics, plus a t-SNE visualization of reference representations.

Significance. If the central claims hold, Noro offers a simple training strategy that substantially improves one-shot VC robustness to noisy reference speech, and the hidden speaker representation result is a noteworthy finding that connects VC pretraining to SSL speaker encoders. The paper is well-motivated and builds on public code and datasets (Amphion, LibriLight, VCTK, DEMAND), which is a strength. However, the experimental design currently does not isolate the contribution of the proposed dual-branch module and contrastive loss from the effect of noisy-reference data augmentation alone, and it omits comparisons with the noise-robust VC methods cited in the introduction. These gaps weaken the attribution of the reported gains and the claimed state-of-the-art relevance.

major comments (4)
  1. [Section II-B.1 and Table I] The experimental protocol cannot uniquely credit the proposed components with the observed robustness improvement. Noro is initialized from the pre-trained baseline and then trained with noisy reference speeches generated by mixing DEMAND noise at SNRs drawn from a normal distribution, while the baseline was trained only on clean references. Therefore the improvement at 0-5 dB (CER 4.66 vs 7.26, SECS 80.09 vs 77.28) could result from ordinary noisy-reference data augmentation alone, independent of the dual-branch averaging and Lref. The paper reports no control ablation with a single-branch encoder trained on the same augmented references and the same diffusion loss, nor an ablation that removes the contrastive loss. Without such controls, the central attribution in the abstract and Section II-B is not established.
  2. [Section II-B.1 and Fig. 2] There is a train-inference mismatch: during training the diffusion model is conditioned on the average of href and h'ref, but at inference it is conditioned on a single branch output. The t-SNE visualization shows that the clean and noisy branch outputs are similar for Noro, but it does not establish that the score network behaves identically when fed the average versus a single branch output. The authors should either provide a theoretical or empirical argument that the mismatch is benign, or modify the training procedure (e.g., randomly dropping one branch during training) to align conditioning between training and inference.
  3. [Section III-A.1 and Tables I-II] The introduction cites several existing noise-robust VC methods, including data augmentation approaches [7], domain adversarial training [12], and conditional denoising [13], but the evaluation does not compare against any of them. Since the paper claims that Noro improves robustness in challenging real-world conditions, a comparison with at least the most related prior work under the same protocol is necessary to assess the contribution. The current comparison only includes FaCodec-VC, FreeVC, DiffVC, and the author's own baseline, none of which were specifically designed for noisy reference robustness.
  4. [Section III-A and Tables I-II] The evaluation lacks statistical rigor: the subjective tests use only 10 pairs and 12 listeners, and the objective metrics are reported as point estimates without error bars or significance tests. In clean conditions, Noro and the baseline are nearly identical (CER 4.74 vs 4.71), and the claimed subjective advantage in noisy conditions (CMOS 2.95 vs 2.09) could be sensitive to listener variability. The authors should report confidence intervals, per-utterance standard deviations, or significance tests, and ideally increase the number of subjective evaluation items.
minor comments (5)
  1. [References] Reference [23] is identical to reference [10]; one of the two duplicates should be removed or replaced with the intended source.
  2. [Section III-A.2 and Table I] The test set is described as using 'unseen types' of noise, but the specific noise types and their proportions are not listed; reporting per-condition results would help the reader judge the generality of the robustness transfer.
  3. [Section II-A.3] The diffusion loss equation lacks explicit norms and a clear dependence of the score network on the segment length; rewriting it with proper L1 norm notation would improve clarity.
  4. [Section III-B.2 and Table III] The comparison in Table III mixes models trained on different amounts of data and with different architectures; a sentence noting that this is not a strictly controlled comparison would be useful for interpreting the EER differences.
  5. [Section II-B.2] The definition of the contrastive loss uses CrossEntropy with a mask Mi,j; it should be clarified whether the diagonal and the positive pairs are excluded from the denominator, as is standard in supervised contrastive learning.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central noise-robustness and speaker-representation results are externally measured on held-out, unseen noise types; the self-citations (Amphion, Emilia) are reproducibility/context only, not load-bearing.

full rationale

Derivation-chain walk: Noro's headline claim ('Noro significantly enhances the robustness of our baseline system in diverse noisy environments') is supported by CER, SECS, CMOS, and SMOS measured with external systems (HuBERT ASR, WavLM speaker model, human listeners) on a test set whose noise types are explicitly unseen from training ('we introduce noise of unseen types from the training stage'). The dual-branch module and contrastive loss Lref are training objectives, not fitted parameters renamed as predictions: no reported number is computed from the training loss by construction. The speaker-representation claim (VC-SPK2VEC, EER 5.32%) is a transfer evaluation: the baseline reference encoder was trained for VC without speaker labels, and its SV performance is benchmarked on VoxCeleb1 under the SUPERB protocol against independent results taken from [16] and [20]. This is an externally falsifiable evaluation, not a quantity defined by its training objective. Self-citations are not load-bearing: Amphion [3] is cited for code/hyperparameters (reproducibility support, which counts as real evidence), and Emilia [4], [5] only as large-scale-dataset context; the VC experiments actually train on the external LibriLight corpus and compare against external baselines. Two validity caveats remain, but they are experimental-design/attribution risks, not circular reductions: (1) there is no ablated control (a single-branch baseline retrained with the same DEMAND augmentation and diffusion loss, with and without Lref), so the improvement could plausibly come from ordinary noisy-reference augmentation rather than the proposed module; (2) there is a train-inference mismatch (training conditions on the average of href and h'ref while inference uses one branch). Neither makes a prediction equal to its input by construction, because the reported metrics are still externally measured on held-out data. Finally, the t-SNE 'well-mixed' claim is partly tautological since Lref explicitly maximizes clean/noisy similarity for the same speaker, so the visualization cannot independently confirm noise-invariance; but the central 0-5 dB result rests on external CER/SECS/CMOS, not on the t-SNE. Verdict: no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the dual-branch training scheme, the contrastive loss, and the transfer of noise invariance to unseen noise types. These are implemented with hand-chosen hyperparameters and a domain assumption about noise representativeness. No new physical or conceptual entities are introduced.

free parameters (4)
  • Temperature tau in contrastive loss
    The temperature parameter scales the contrastive logits in Lref. Its value is not stated in the paper and is left to the code link.
  • Loss weights alpha and beta
    The total loss is Ltotal = alpha * Ldiff + beta * Lref, but the paper does not report the chosen values for alpha and beta.
  • SNR sampling distribution = N(0, 20) dB
    The paper specifies that noisy references are created by mixing clean speech with DEMAND noise at SNRs sampled from a normal distribution with mean 0 and standard deviation 20 dB. This is a hand-chosen augmentation distribution.
  • Query embedding count m
    The reference encoder uses m randomly initialized query embeddings to pool the hidden sequence into an utterance-level representation. The value of m is not given in the paper.
assumptions (3)
  • ad hoc to paper The clean and noisy branches can share weights and be trained with an averaged representation, and the resulting encoder remains valid when only one branch is used at inference.
    Section II.B.1 states that during training the average of href and h'ref is fed into the diffusion model, while inference uses a single encoder. No analysis or experiment validates that this train-inference mismatch is harmless.
  • domain assumption Speaker labels are available for the contrastive loss on LibriLight.
    The noise-agnostic contrastive loss Lref requires speaker labels yspk. LibriLight comes from LibriVox audiobooks and has speaker metadata, but the paper does not explain how labels were obtained or whether all speakers are used.
  • domain assumption Noise augmentation with eight DEMAND noise types and SNRs in N(0, 20) dB is representative of real-world reference noise.
    Section II.B.1 describes the augmentation. The test set uses unseen noise types, but the paper never specifies what those are, so the coverage of real-world noise conditions is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Noro: Noise-Robust One-shot Voice Conversion with Hidden Speaker Representation Learning." pith.science (2026). https://pith.science/paper/UBAVURW2

@misc{pith2026241119770,
  author       = {Pith},
  title        = {Pith review of: Noro: Noise-Robust One-shot Voice Conversion with Hidden Speaker Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBAVURW2}},
  note         = {Machine review of arXiv:2411.19770}
}
read the original abstract

The effectiveness of one-shot voice conversion (VC) decreases in real-world scenarios where reference speeches, which are often sourced from the internet, contain various disturbances like background noise. To address this issue, we introduce Noro, a noise-robust one-shot VC system. Noro features innovative components tailored for VC using noisy reference speeches, including a dual-branch reference encoding module and a noise-agnostic contrastive speaker loss. Experimental results demonstrate that Noro outperforms our baseline system in both clean and noisy scenarios, highlighting its efficacy for real-world applications. Additionally, we investigate the hidden speaker representation capabilities of our baseline system by repurposing its reference encoder as a speaker encoder. The results show that it is competitive with several advanced self-supervised learning models for speaker representation under the SUPERB settings, highlighting the potential for advancing speaker representation learning through one-shot VC tasks.

Figures

Figures reproduced from arXiv: 2411.19770 by the authors.

Figure 1
Figure 1. Model architecture of our baseline system and Noro. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of reference representations. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. REF-VC: Robust, Expressive and Fast Zero-Shot Voice Conversion with Diffusion Transformers

    eess.AS 2025-08 unverdicted novelty 5.0 of 10

    REF-VC is a zero-shot voice conversion system that random-erases redundant parts of speech-embedding features to stay robust to noise, and uses shortcut-distilled flow matching to convert speech in only four steps.

  2. Overview of the Amphion Toolkit (v0.2)

    cs.SD 2025-01 conditional novelty 4.0 of 10

    Amphion v0.2 is an open-source toolkit for audio, music, and speech generation, adding a 101K-hour multilingual dataset, processing pipelines, and pretrained models.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [7]

    Toward Degradation-Robust V oice Conversion,

    C.-Y . Huang, K.-W. Chang, and H.-Y . Lee, “Toward Degradation-Robust V oice Conversion,” in ICASSP, 2022

    work page 2022

  2. [9]

    Learn- ing Noise-Independent Speech Representation for High- Quality V oice Conversion for Noisy Target Speakers,

    L. Xue, S. Yang, N. Hu, D. Su, and L. Xie, “Learn- ing Noise-Independent Speech Representation for High- Quality V oice Conversion for Noisy Target Speakers,” in INTERSPEECH, 2021

    work page 2021

  3. [12]

    Noise-Robust V oice Conver- sion with Domain Adversarial Training,

    H. Du, L. Xie, and H. Li, “Noise-Robust V oice Conver- sion with Domain Adversarial Training,” Neural Net- works, vol. 148, pp. 74–84, 2022

    work page 2022

  4. [13]

    Noise-Robust V oice Conversion by Conditional Denoising Training Using Latent Variables of Recording Quality and En- vironment,

    T. Igarashi, Y . Saito, K. Seki, et al. , “Noise-Robust V oice Conversion by Conditional Denoising Training Using Latent Variables of Recording Quality and En- vironment,” in INTERSPEECH, 2024

    work page 2024

  5. [1]

    Diffusion-Based V oice Con- version with Fast Maximum Likelihood Sampling Scheme,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, M. S. Kudinov, and J. Wei, “Diffusion-Based V oice Con- version with Fast Maximum Likelihood Sampling Scheme,” in ICLR, 2022

    work page 2022

  6. [2]

    SEF-VC: Speaker Embedding Free Zero-Shot V oice Conversion with Cross Attention,

    J. Li, Y . Guo, X. Chen, and K. Yu, “SEF-VC: Speaker Embedding Free Zero-Shot V oice Conversion with Cross Attention,” in ICASSP, 2024

    work page 2024

  7. [3]

    Amphion: An open- source audio, music and speech generation toolkit,

    X. Zhang, L. Xue, Y . Gu, et al. , “Amphion: An open- source audio, music and speech generation toolkit,” in SLT, 2024

    work page 2024

  8. [4]

    Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large- Scale Speech Generation,

    H. He, Z. Shang, C. Wang, et al., “Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large- Scale Speech Generation,” in SLT, 2024

    work page 2024

Show all 28 references
  1. [5]

    Emilia: A Large- Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation,

    H. He, Z. Shang, C. Wang, et al. , “Emilia: A Large- Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation,” in arXiv:2501.15907, 2025

    2025

  2. [6]

    How Far are We from Robust V oice Conversion: A Survey,

    T.-H. Huang, J.-H. Lin, and H.-Y . Lee, “How Far are We from Robust V oice Conversion: A Survey,” in SLT, 2021

    2021

  3. [8]

    Noisy-to-Noisy V oice Conver- sion Under Variations of Noisy Condition,

    C. Xie and T. Toda, “Noisy-to-Noisy V oice Conver- sion Under Variations of Noisy Condition,” IEEE/ACM Transactions on Audio, Speech, and Language Process- ing, vol. 31, pp. 3871–3882, 2023

    2023

  4. [11]

    Speech Enhancement-Assisted Stargan V oice Conversion in Noisy Environments,

    Y .-J. Chan, C.-J. Peng, S.-S. Wang, H.-M. Wang, Y . Tsao, and T.-S. Chi, “Speech Enhancement-Assisted Stargan V oice Conversion in Noisy Environments,” arXiv preprint arXiv:2110.09923 , 2021

    2021 arXiv

  5. [14]

    Learning Speaker Embedding from Text-to- Speech,

    J. Cho, P. Zelasko, J. Villalba, S. Watanabe, and N. Dehak, “Learning Speaker Embedding from Text-to- Speech,” in INTERSPEECH, 2020

    2020

  6. [15]

    Im- proving Reconstruction Loss Based Speaker Embedding in Unsupervised and Semi-Supervised Scenarios,

    J. Cho, P. ˙Zelasko, J. Villalba, and N. Dehak, “Im- proving Reconstruction Loss Based Speaker Embedding in Unsupervised and Semi-Supervised Scenarios,” in ICASSP, 2021

    2021

  7. [16]

    SUPERB: Speech Processing Universal Performance Benchmark,

    S.-w. Yang, P.-H. Chi, Y .-S. Chuang, et al. , “SUPERB: Speech Processing Universal Performance Benchmark,” in INTERSPEECH, 2021

    2021

  8. [17]

    Wav2vec: Unsupervised Pre-Training for Speech Recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “Wav2vec: Unsupervised Pre-Training for Speech Recognition,” in INTERSPEECH, 2019

    2019

  9. [18]

    Wav2vec 2.0: A Framework for Self-Supervised Learn- ing of Speech Representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A Framework for Self-Supervised Learn- ing of Speech Representations,” in NeurIPS, 2020

    2020

  10. [19]

    Hubert: Self- Supervised Speech Representation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self- Supervised Speech Representation Learning by Masked Prediction of Hidden Units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 3451–3460, 2021

    2021

  11. [20]

    Wavlm: Large- scale Self-supervised Pre-training for Full Stack Speech Processing,

    S. Chen, C. Wang, Z. Chen, et al. , “Wavlm: Large- scale Self-supervised Pre-training for Full Stack Speech Processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  12. [21]

    Textless Speech-to-Speech Translation on Real Data,

    A. Lee, H. Gong, P.-A. Duquenne, et al. , “Textless Speech-to-Speech Translation on Real Data,” inNACCL, 2022

    2022

  13. [22]

    Naturalspeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers,

    K. Shen, Z. Ju, X. Tan, et al., “Naturalspeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers,” in ICLR, 2023

    2023

  14. [23]

    Investigating RNN-Based Speech Enhancement Methods for Noise-Robust Text-to-Speech,

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yam- agishi, “Investigating RNN-Based Speech Enhancement Methods for Noise-Robust Text-to-Speech,” in SSW, 2016

    2016

  15. [24]

    Naturalspeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models,

    Z. Ju, Y . Wang, K. Shen, et al. , “Naturalspeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models,” ICML, 2024

    2024

  16. [25]

    FreeVC: Towards High- Quality Text-Free One-Shot V oice Conversion,

    J. Li, W. Tu, and L. Xiao, “FreeVC: Towards High- Quality Text-Free One-Shot V oice Conversion,” in ICASSP, 2023

    2023

  17. [26]

    Libri-Light: A Benchmark for ASR with Limited or No Supervision,

    J. Kahn, M. Riviere, W. Zheng, et al. , “Libri-Light: A Benchmark for ASR with Limited or No Supervision,” in ICASSP, 2020

    2020

  18. [27]

    Unsupervised Pretraining Transfers Well Across Lan- guages,

    M. Rivi `ere, A. Joulin, P.-E. Mazar ´e, and E. Dupoux, “Unsupervised Pretraining Transfers Well Across Lan- guages,” in ICASSP, 2020

    2020

  19. [28]

    Vq-Wav2vec: Self-Supervised Learning of Discrete Speech Represen- tations,

    A. Baevski, S. Schneider, and M. Auli, “Vq-Wav2vec: Self-Supervised Learning of Discrete Speech Represen- tations,” in ICLR, 2019

    2019

  20. [29]

    V oxCeleb: Large-Scale Speaker Verification in the Wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisser- man, “V oxCeleb: Large-Scale Speaker Verification in the Wild,” Computer Speech & Language , vol. 60, p. 101 027, 2020. 5

    2020

Pith tools

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