Pith. sign in

REVIEW 3 major objections 5 minor 45 references

LinearVC: Linear transformations of self-supervised features through the lens of voice conversion

T0 review · 3 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 A rank-100 projection converts voices as well as top systems.

desk verdict A carefully controlled empirical paper that makes a simple, credible point about linear maps on SSL features for voice conversion, with a real but addressable gap in the evidence for the low-rank content-subspace interpretation. read the letter →

arxiv 2506.01510 v1 pith:TKFKQXGP submitted 2025-06-02 eess.AS cs.CL

Herman Kamper , Benjamin van Niekerk , Julian Zaïdi , Marc-André Carbonneau This is my paper
classification eess.AScs.CL
keywords voiceconversionself-supervisedspeechfeaturesWavLMlinearprojectioncontentfactorizationsingularvaluedecompositionspeakerdisentanglementfeaturegeometry
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

LinearVC claims that converting one voice into another can be reduced to a single learned linear transformation of self-supervised speech features. Working in the sixth layer of WavLM, the paper pairs source and target frames by nearest neighbours, solves a least-squares regression for a projection matrix, and vocodes the projected frames; the result matches state-of-the-art systems on intelligibility and speaker similarity. Constraining the projection to a rotation or reflection barely hurts, which the paper reads as evidence that phonetic content lives in a low-dimensional subspace that speakers share. An explicit singular value decomposition factorization confirms the picture: content stays intelligible at rank 16, and a rank-100 projection gives competitive voice conversion. If true, the result matters because it isolates a concrete, cheap mechanism inside large speech models and explains why simple SSL-based conversion works.

What carries the argument

The carrier of the argument is the linear projection learned by multivariate least-squares regression in Eq. (1), together with its constrained variants. Restricting $W$ to the orthogonal group turns the regression into the orthogonal Procrustes problem, which isolates rotation and reflection; comparing that to translation-only, unconstrained, and bias-augmented versions maps out which part of the linear map is doing the work. The final mechanism is the SVD factorization of a concatenated block matrix $X = [X_1 \cdots X_K]$, approximated as $U \Sigma [S_1 \cdots S_K]$, where $U\Sigma$ is the shared content $C$ and each $S_k$ is a speaker-specific linear transformation; using $S_{\mathrm{src}}^+ S_{\mathrm{tgt}}$ as the projection gives the low-rank conversion whose rank can be varied explicitly.

What would settle it

Replace the nearest-neighbour pairing with random pairing of source and target frames while keeping everything else fixed; if intelligibility and speaker similarity stayed high, the content-alignment story would be wrong. A second check: train the rank-100 factorization on one set of frames and test whether a linear phoneme classifier trained on the content subspace $C$ at ranks 16 and 100 performs no worse than on the full features, which would confirm that the low-dimensional content is genuinely phonetic.

Watch

Extended reading notes

Core claim

The paper's central discovery is that WavLM's sixth-layer features are organized so that speaker identity and phonetic content occupy different subspaces, and a linear map between them is enough to change the voice without changing the words. LinearVC learns $W$ by minimizing $\|Y - XW\|_F^2$ over nearest-neighbour paired frames; at inference it projects source frames with $X_{\mathrm{tgt}} = X_{\mathrm{src}}W$ and feeds them to HiFi-GAN. When $W$ is constrained to be orthogonal, intelligibility is unchanged and speaker similarity is nearly as high as with an unconstrained map, so rotation around the origin is the dominant effect. Finally, factorizing paired features from many speakers as $X_k \approx C S_k$ via SVD shows that the common content $C$ alone keeps a character error rate below 4% at rank 16, and that $r=100$ gives objective and subjective scores on par with kNN-VC and SoundStorm. The conclusion is that phonetic content is embedded in a shared low-dimensional subspace that each speaker reaches through a speaker-specific linear map.

Load-bearing premise

The load-bearing premise is that each source frame's single nearest neighbour in target speech really is the same phonetic content; if these pairings are often wrong, the learned maps could transfer content-dependent speaker cues, making the reported disentanglement look better than it is.

Editorial extensions

If this is right

  • A voice converter can be built by learning one matrix per speaker pair: paired frames from a single 2.7-minute sample, a least-squares solve, and a pretrained vocoder.
  • The same layer of an SSL model can hold content and speaker information in separate subspaces, so probing or editing one without touching the other is feasible.
  • Content can be compressed to about 100 dimensions without losing conversion quality, and to 16 dimensions without losing intelligibility, giving a cheap phonetic bottleneck.
  • Rotation-based conversion means the geometry of SSL features, not the complexity of the decoder, is what carries voice identity.

Reading between the lines

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

  • Beyond the paper: the same factorization could serve as a speaker anonymizer, since replacing $S_k$ with a different or averaged speaker map while keeping $C$ should preserve content while changing identity; this is directly testable.
  • Beyond the paper: if the subspace claim generalizes, the rotation-only result should replicate with other SSL speech models and layers; a negative result would locate the effect in WavLM's training objective rather than in SSL in general.
  • Beyond the paper: the rank-16 content subspace is a candidate compact phonetic representation; one could test whether a linear classifier on $C$ at rank 16 predicts phoneme labels as well as the full 1024-dimensional features.
  • Beyond the paper: because the factorization is learned from nearest-neighbour pairs, its quality upper-bounds the degree of true phonetic alignment in SSL features; training on cross-lingual pairs would test whether the shared content subspace is language-independent.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. LinearVC proposes a voice conversion method that learns a linear transformation between source and target speaker frames extracted from the sixth layer of WavLM-Large. Source and target frames are paired by nearest-neighbor matching, and a linear projection is learned via least squares. The paper first shows that this simple linear model achieves conversion quality comparable to kNN-VC, FreeVC, and SoundStorm. It then constrains the transformation to translations, rotations, and other linear maps, finding that rotation alone preserves intelligibility and yields substantial speaker similarity. Finally, it factorizes the feature matrix for multiple speakers into a shared low-rank content component and speaker-specific transforms via SVD, reporting competitive results at rank 100. The paper's central claim is that these results reveal a low-dimensional, speaker-independent phonetic content subspace in WavLM features that can be linearly manipulated for voice conversion.

Significance. If the geometric interpretation is correct, the paper offers a practical, highly transparent voice conversion method and a concrete characterization of how SSL speech representations organize content and speaker information. The empirical backbone is solid: the authors use held-out evaluation utterances, multiple objective metrics, MUSHRA-style listening tests with statistical testing, and they release code and samples. The ablations across transformation types and factorization ranks are informative and directly support the method's simplicity. The main caveats are that the central philosophical claim about a shared content subspace is inferred from nearest-neighbor pairing that is not itself validated, and the objective evidence for the rotation claim lacks confidence intervals. These issues are addressable with additional controls, and the practical contribution would survive even if the subspace interpretation were weakened.

major comments (3)
  1. [Section 5.1 / Figure 4] The interpretation of the SVD factorization as extracting a shared phonetic content subspace rests entirely on the nearest-neighbor pairing at the top of Figure 1 and in Section 5.1: the matrices X_k are assumed to contain 'the same content spoken by different speakers.' The paper provides no validation of these pairs against phonetic labels or forced alignment, and no null experiment such as training the factorization on randomly permuted frame pairs. Without such a control, the low-rank reconstruction and the intelligibility preservation in Figure 4 are also compatible with a weaker explanation: any sufficiently smooth low-rank linear map between the source and target frame distributions could preserve enough information for the downstream vocoder and ASR. I request a random-pairing baseline and, if possible, a frame-level phonetic agreement check, to make the low-dimensional shared-content claim load-bearing rather than suggestive.
  2. [Section 4 / Table 2] The claim that 'rotations and reflections around the origin are the main contributors' to LinearVC's performance is supported only by EER values in Table 2, which are reported without confidence intervals or significance tests. The gap between the orthogonal condition (27.7) and the unconstrained condition (31.8) is small and could be within noise, especially given the paper's own statement in Section 3.2 that EER gives only a coarse correspondence with perceived similarity. In addition, the subjective evaluation is only reported for the unconstrained LinearVC in Table 1, not for the constrained variants. To make the rotation conclusion robust, I ask for confidence intervals or repeated sampling for the objective metrics and, if feasible, a smaller subjective test comparing the orthogonal and unconstrained conditions.
  3. [Section 3.2 / Table 1 / Section 5.2] The paper's abstract and conclusion generalize the findings to 'self-supervised speech representations' broadly, but all experiments use a single model (WavLM-Large) and a single layer (layer 6). The authors acknowledge this only in the final sentence of the conclusion. This is a scope limitation rather than an internal inconsistency, but because the title and abstract speak generally, the claims should either be explicitly scoped to WavLM layer 6 or supported by at least one additional SSL model. I would also note that the 'rank of just 100' in the abstract is a hyperparameter tuned on dev-clean (Figure 4), not a discovered model-free constant; the manuscript should state this more explicitly when presenting the rank-100 result.
minor comments (5)
  1. [Figure 1 and Section 2] The notation is inconsistent: Section 2 defines the projection as Xtgt = Xsrc W with row vectors, but the Figure 1 caption shows 'W⊤xt' and labels the output as 'W⊤xt', which suggests column-vector convention. Please align the caption with the equations.
  2. [Section 5.1, footnote 1] The footnote says 'at least r non-zero eigenvalues', but the SVD decomposes a matrix using singular values, not eigenvalues. Please rephrase to 'non-zero singular values' or clarify.
  3. [Section 5.1] The constraint rank(CS)≤r is automatically satisfied because C has r columns, so spelling out the constraint as a separate condition is confusing. It may help to state that C is constructed from the top r singular vectors.
  4. [Table 1] The row 'LinearVC factor.' is introduced in Table 1 before the factorization method is described in Section 5. Consider moving the last row to the later table or adding a forward reference in the table caption.
  5. [Section 3.1] The comparison to SoundStorm uses a model trained on 60k hours of LibriLight, while LinearVC and kNN-VC use 2.7 minutes of reference audio per speaker. This is a reasonable demonstration of data efficiency, but the difference should be acknowledged in the discussion, as it affects the interpretation of 'comparable performance.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical demonstrations on held-out data, not derivations that reduce to fitted inputs.

full rationale

The paper's derivation chain is self-contained against external benchmarks. LinearVC fits a projection W in Eq. (1) to nearest-neighbour paired frames, but then evaluates the projection on utterances that are explicitly disjoint from the reference data ('The evaluation utterances do not overlap with the reference data'), so the reported intelligibility and similarity are genuine held-out results. The constrained-transformation experiments (Section 4) compare translation, orthogonal, and unconstrained maps; the conclusion that rotations and reflections are the main contributors is an empirical observation about the resulting EER values, not an identity forced by the optimization. The factorized method in Section 5.1 uses an SVD low-rank approximation of the paired feature matrix; the 'content' subspace is fitted to training data, but the claims about rank are validated through downstream CER and EER on dev/test splits (Figure 4, Table 1), not through reconstruction error on the training targets. The rank r=100 is selected on dev-clean and then applied to test-clean, which is standard model selection rather than circular prediction. Self-citations ([14], [21]) are used for architectural inspiration, layer choice, and data augmentation details, but none of these citations supplies a load-bearing theorem or uniqueness argument; the method's effectiveness is demonstrated by direct comparison with kNN-VC, FreeVC, and SoundStorm. No equation in the paper is equivalent by construction to the conclusions drawn from it, and no fitted parameter is renamed as a prediction. The unvalidated nearest-neighbour pairing assumption is a potential confound for the geometric interpretation, but it is an empirical validity concern, not a circularity that reduces the derivation to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or particles are introduced. The 'shared content subspace' is a representational construct derived from SVD, not an independently posited entity.

free parameters (3)
  • Factorization rank r = 100 (also 16, 24, 200 in analyses)
    Chosen from dev-clean results in Figure 4 as a compromise between CER and EER; central to the low-rank claim.
  • WavLM layer number = 6
    Chosen 'based on [14]' rather than from a sweep in this paper; the geometric findings may depend on this layer.
  • Number of nearest neighbours = 1
    Selected after development experiments in Section 3.1; affects how paired training frames are built.
assumptions (5)
  • domain assumption WavLM's sixth-layer features are a valid representation of speech content and speaker identity for this task.
    The entire method operates on these features; no independent verification that this layer is the best for disentanglement.
  • domain assumption Nearest-neighbour frame pairing produces same-content pairs.
    Section 2, Figure 1 top; the regression in Eq. (1) depends on this alignment.
  • domain assumption The low-rank SVD of the concatenated matrix yields a shared content subspace across speakers.
    Section 5.1; the interpretation of U Sigma as content relies on the pairing and on the singular vectors separating speaker-specific directions.
  • domain assumption HiFi-GAN trained on WavLM features can synthesize intelligible speech from projected frames.
    Used at inference; if the vocoder fails for certain speakers, metrics would be affected.
  • standard math Orthogonal Procrustes solver [38] returns the optimal orthogonal W.
    Section 4; this is a known closed-form solution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LinearVC: Linear transformations of self-supervised features through the lens of voice conversion." pith.science (2026). https://pith.science/paper/TKFKQXGP

@misc{pith2026250601510,
  author       = {Pith},
  title        = {Pith review of: LinearVC: Linear transformations of self-supervised features through the lens of voice conversion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKFKQXGP}},
  note         = {Machine review of arXiv:2506.01510}
}
read the original abstract

We introduce LinearVC, a simple voice conversion method that sheds light on the structure of self-supervised representations. First, we show that simple linear transformations of self-supervised features effectively convert voices. Next, we probe the geometry of the feature space by constraining the set of allowed transformations. We find that just rotating the features is sufficient for high-quality voice conversion. This suggests that content information is embedded in a low-dimensional subspace which can be linearly transformed to produce a target voice. To validate this hypothesis, we finally propose a method that explicitly factorizes content and speaker information using singular value decomposition; the resulting linear projection with a rank of just 100 gives competitive conversion results. Our work has implications for both practical voice conversion and a broader understanding of self-supervised speech representations. Samples and code: https://www.kamperh.com/linearvc/.

Figures

Figures reproduced from arXiv: 2506.01510 by the authors.

Figure 1
Figure 1. LinearVC: At learning time (top), source speech frames are paired with target frames by finding their single nearest neighbours. A projection matrix is learned by finding the least squares solution for mapping from source to target frames. At inference time (bottom), frames from an unseen source utterance are linearly projected and then vocoded to get speech in the target speaker’s voice. The intuition behind this a… view at source ↗
Figure 2
Figure 2. A cartoon illustration of the effects of different linear transformations on the representation space. The dark and light blobs respectively represent where features from two distinct speakers are situated in the SSL space [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Intelligibility (CER) and speaker similarity (EER) on dev-clean as a function of rank when using content factorization. apply the target speaker transformation to convert to the desired speaker’s voice: Xtgt = XsrcS + srcStgt The S + srcStgt term functions like the projection matrix W of the non-factorized LinearVC. The difference here is that we can explicitly set the transformation’s rank r. 5.2. Voice conversion … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 36 canonical work pages

  1. [1]

    It has applications in entertainment [3], language tutoring [4], accessible speech processing [5–7], and anonymization [8]

    Introduction V oice conversion aims to alter input speech to mimic a target speaker’s voice [1, 2]. It has applications in entertainment [3], language tutoring [4], accessible speech processing [5–7], and anonymization [8]. This wide range of applications has naturally led to diverse solutions. Some systems leverage large spoken language models, prompting...

    work page

  2. [2]

    During train- ing (top), a linear transformation from a source to a target speaker is learned

    LinearVC The LinearVC framework is illustrated in Figure 1. During train- ing (top), a linear transformation from a source to a target speaker is learned. Utterances are first encoded into D-dimensional fea- ture frames using a large SSL speech model, like WavLM [16]. Then, for each of the N source frames, we find the closest neigh- bour from the set of M...

    arXiv 2025

  3. [3]

    V oice conversion experiments We compare LinearVC to three state-of-the-art voice conversion methods, demonstrating its effectiveness despite its simplicity. 3.1. Experimental setup Data.Experiments are performed on the English Libri- Speech corpus [25]. We perform development experiments on the dev-clean subset and report final scores on the test-clean s...

    work page

  4. [4]

    Figure 2 gives a cartoon illustration of the different configura- tions

    Further analysis: Constrained linear transformations We now constrain the type of transformation in LinearVC to gain insights into the organization of the SSL representation space. Figure 2 gives a cartoon illustration of the different configura- tions. First, we exclusively use a bias vector, allowing only for translation. Then, W is constrained to be or...

    work page

  5. [5]

    In this section, we validate this interpretation by explicitly disentangling content and speaker information

    Factorizing out a shared content subspace The experiments in the previous section support the hypothe- sis that content information is embedded in a low-dimensional subspace, which can be linearly transformed to produce a target voice. In this section, we validate this interpretation by explicitly disentangling content and speaker information. Specificall...

    work page

  6. [6]

    In addition to providing a practical solution to voice conversion without requiring complex model training, it offers valuable insights into the organization of SSL features

    Conclusion We introduced LinearVC, a simple yet effective method for voice conversion that uses a linear projection of SSL speech representations. In addition to providing a practical solution to voice conversion without requiring complex model training, it offers valuable insights into the organization of SSL features. Our experiments revealed that phone...

    work page

  7. [7]

    An overview of voice conversion systems,

    S. H. Mohammadi and A. Kain, “An overview of voice conversion systems,”Speech Communication, 2017

    work page 2017

  8. [8]

    An overview of voice conversion and its challenges: From statistical modeling to deep learning,

    B. Sisman, J. Yamagishi, S. King, and H. Li, “An overview of voice conversion and its challenges: From statistical modeling to deep learning,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2020

    work page 2020

Show all 45 references
  1. [9]

    Making of Season 2 finale,

    Disney, “Making of Season 2 finale,”Disney Gallery: The Man- dalorian, 2021

    2021

  2. [10]

    A pilot study of applying sequence-to-sequence voice conversion to evaluate the intelligibil- ity of L2 speech using a native speaker’s shadowings,

    H. Geng, D. Saito, and N. Minematsu, “A pilot study of applying sequence-to-sequence voice conversion to evaluate the intelligibil- ity of L2 speech using a native speaker’s shadowings,” inAPSIPA ASC, 2024

    2024

  3. [11]

    Respeecher gives voice to Michael York in healthcare initiative,

    G. Singer, “Respeecher gives voice to Michael York in healthcare initiative,”Respeecher Blog, 2021

    2021

  4. [12]

    V oice conversion for stuttered speech, instruments, unseen languages and textually described voices,

    M. Baas and H. Kamper, “V oice conversion for stuttered speech, instruments, unseen languages and textually described voices,” Communications in Computer and Information Science, 2023

    2023

  5. [13]

    Unsupervised rhythm and voice conversion of dysarthric to healthy speech for ASR,

    K. E. Hajal, E. Hermann, A. Kulkarni, and M. Magimai.-Doss, “Unsupervised rhythm and voice conversion of dysarthric to healthy speech for ASR,” inSPADE, 2025

    2025

  6. [14]

    The V oicePrivacy 2022 Challenge: Progress and perspectives in voice anonymisation,

    M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Champion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamagishi, “The V oicePrivacy 2022 Challenge: Progress and perspectives in voice anonymisation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2022

  7. [15]

    V ALL-E R: Robust and efficient zero- shot text-to-speech synthesis via monotonic alignment,

    B. Han, L. Zhou, S. Liu, S. Chen, L. Meng, Y . Qian, Y . Liu, S. Zhao, J. Li, and F. Wei, “V ALL-E R: Robust and efficient zero- shot text-to-speech synthesis via monotonic alignment,”arXiv preprint arXiv:2406.07855, 2024

    2024 arXiv

  8. [16]

    SoundStorm: Efficient parallel audio generation,

    Z. Borsos, M. Sharifi, D. Vincent, E. Kharitonov, N. Zeghidour, and M. Tagliasacchi, “SoundStorm: Efficient parallel audio generation,” arXiv preprint arXiv:2305.09636, 2023

    2023 arXiv

  9. [17]

    QuickVC: Any-to- many voice conversion using inverse short-time fourier transform for faster conversion,

    H. Guo, C. Liu, C. T. Ishi, and H. Ishiguro, “QuickVC: Any-to- many voice conversion using inverse short-time fourier transform for faster conversion,”arXiv preprint arXiv:2302.08296, 2023

    2023 arXiv

  10. [18]

    StreamVC: Real-time low-latency voice conversion,

    Y . Yang, Y . Kartynnik, Y . Li, J. Tang, X. Li, G. Sung, and M. Grund- mann, “StreamVC: Real-time low-latency voice conversion,” in ICASSP, 2024

    2024

  11. [19]

    Frag- mentVC: Any-to-any voice conversion by end-to-end extracting and fusing fine-grained voice fragments with attention,

    Y . Y . Lin, C.-M. Chien, J.-H. Lin, H.-y. Lee, and L.-s. Lee, “Frag- mentVC: Any-to-any voice conversion by end-to-end extracting and fusing fine-grained voice fragments with attention,” inICASSP, 2021

    2021

  12. [20]

    V oice conversion with just nearest neighbors,

    M. Baas, B. van Niekerk, and H. Kamper, “V oice conversion with just nearest neighbors,” inInterspeech, 2023

    2023

  13. [21]

    Optimal transport maps are good voice converters,

    A. Asadulaev, R. Korst, V . Shutov, A. Korotin, Y . Grebnyak, V . Egiazarian, and E. Burnaev, “Optimal transport maps are good voice converters,”arXiv preprint arXiv:2411.02402, 2024

    2024 arXiv

  14. [22]

    WavLM: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, , and F. Wei, “WavLM: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selec...

    2022

  15. [23]

    Self-supervised predic- tive coding models encode speaker and phonetic information in orthogonal subspaces,

    O. D. Liu, H. Tang, and S. Goldwater, “Self-supervised predic- tive coding models encode speaker and phonetic information in orthogonal subspaces,” inInterspeech, 2023

    2023

  16. [24]

    Orthogo- nality and isotropy of speaker and phonetic information in self- supervised speech representations,

    M. Mohamed, O. D. Liu, H. Tang, and S. Goldwater, “Orthogo- nality and isotropy of speaker and phonetic information in self- supervised speech representations,” inInterspeech, 2024

    2024

  17. [25]

    Layer-wise analysis of a self-supervised speech representation model,

    A. Pasad, J.-C. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” inASRU, 2021

    2021

  18. [26]

    How should we extract discrete audio tokens from self-supervised models?

    P. Mousavi, J. Duret, S. Zaiem, L. D. Libera, A. Ploujnikov, C. Sub- akan, and M. Ravanelli, “How should we extract discrete audio tokens from self-supervised models?” inInterspeech, 2024

    2024

  19. [27]

    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

  20. [28]

    Phoneme hallucinator: One-shot voice conversion via set expansion,

    S. Shan, Y . Li, A. Banerjee, and J. B. Oliva, “Phoneme hallucinator: One-shot voice conversion via set expansion,” inAAAI, 2024

    2024

  21. [29]

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

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversar- ial networks for efficient and high fidelity speech synthesis,” in NeurIPS, 2020

    2020

  22. [30]

    Self-supervised models of speech infer universal articulatory kine- matics,

    C. J. Cho, A. Mohamed, A. W. Black, and G. K. Anumanchipalli, “Self-supervised models of speech infer universal articulatory kine- matics,” inICASSP, 2024

    2024

  23. [31]

    LibriSpeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “LibriSpeech: An ASR corpus based on public domain audio books,” inICASSP, 2015

    2015

  24. [32]

    Comparative layer-wise analysis of self-supervised speech models,

    A. Pasad, B. Shi, and K. Livescu, “Comparative layer-wise analysis of self-supervised speech models,” inICASSP, 2023

    2023

  25. [33]

    DASB – Discrete audio and speech benchmark,

    P. Mousavi, L. Della Libera, J. Duret, A. Ploujnikov, C. Subakan, and M. Ravanelli, “DASB – Discrete audio and speech benchmark,” arXiv preprint arXiv:2406.14294, 2024

    2024 arXiv

  26. [34]

    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

  27. [35]

    Libri-Light: A benchmark for ASR with limited or no supervision,

    J. Kahn, M. Rivi`ere, W. Zheng, E. Kharitonov, Q. Xu, P. E. Mazar´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen, T. Likhoma- nenko, G. Synnaeve, A. Joulin, A. Mohamed, and E. Dupoux, “Libri-Light: A benchmark for ASR with limited or no supervision,” inICASSP, 2020

    2020

  28. [36]

    V oice Conversion Challenge 2020: Intra-lingual semi-parallel and cross-lingual voice conver- sion,

    Z. Yi, W.-C. Huang, X. Tian, J. Yamagishi, R. K. Das, T. Kin- nunen, Z.-H. Ling, and T. Toda, “V oice Conversion Challenge 2020: Intra-lingual semi-parallel and cross-lingual voice conver- sion,” inVCCBC, 2020

    2020

  29. [37]

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

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

    2023

  30. [38]

    Predictions of subjective ratings and spoofing assessments of Voice Conversion Challenge 2020 submissions,

    R. K. Das, T. Kinnunen, W.-C. Huang, Z.-H. Ling, J. Yamagishi, Z. Yi, X. Tian, and T. Toda, “Predictions of subjective ratings and spoofing assessments of Voice Conversion Challenge 2020 submissions,” inVCCBC, 2020

    2020

  31. [39]

    X-vectors: Robust DNN embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X-vectors: Robust DNN embeddings for speaker recognition,” in ICASSP, 2018

    2018

  32. [40]

    Method for the subjec- tive assessment of intermediate quality level of audio systems,

    International Telecommunication Union, “Method for the subjec- tive assessment of intermediate quality level of audio systems,” 2015

    2015

  33. [41]

    Statistical tests with MUSHRA data,

    C. Mendon c ¸a and S. Delikaris-Manias, “Statistical tests with MUSHRA data,” inAudio Engineering Society International Con- vention, 2018

    2018

  34. [42]

    The use of ranks to avoid the assumption of normal- ity implicit in the analysis of variance,

    M. Friedman, “The use of ranks to avoid the assumption of normal- ity implicit in the analysis of variance,”Journal of the American Statistical Association, 1937

    1937

  35. [43]

    Individual comparisons by ranking methods,

    F. Wilcoxon, “Individual comparisons by ranking methods,”Bio- metrics Bulletin, 1945

    1945

  36. [44]

    A generalized solution of the orthogonal Pro- crustes problem,

    P. H. Sch¨onemann, “A generalized solution of the orthogonal Pro- crustes problem,”Psychometrika, 1966

    1966

  37. [45]

    Eta- WavLM: Efficient speaker identity removal in self-supervised speech representations using a simple linear equation,

    G. Ruggiero, M. Testa, J. Van de Walle, and L. Di Caro, “Eta- WavLM: Efficient speaker identity removal in self-supervised speech representations using a simple linear equation,” inACL, 2025

    2025

Pith tools

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