Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Private kNN-VC: Interpretable Anonymization of Converted Speech

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 Phone duration and per-phone phonetic variation encode speaker identity in converted speech; anonymizing them lifts the attack's equal-error rate from 10.4% to 49.9%.

desk verdict Useful ablation and a real protocol-confound finding, but the causal 'proving' claim overshoots the evidence. read the letter →

arxiv 2505.17584 v1 pith:GNCRONDB submitted 2025-05-23 eess.AS cs.SD

Carlos Franzreb , Arnab Das , Tim Polzehl , Sebastian Möller This is my paper
classification eess.AScs.SD
keywords speakeranonymizationvoiceconversioninterpretabilityprosodyphonedurationphoneticvariationk-meansquantizationprivacy-utilitytrade-off
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 asks why kNN-VC, a voice conversion model that preserves spoken content, performs poorly at hiding who is speaking. It hypothesizes that prosodic characteristics, specifically the durations of phones and the variation among the acoustic features representing each phone, leak identity. To test this, the authors add two interpretable controls to kNN-VC: a predictor that rewrites phone durations and a per-phone clustering step that limits phonetic variation. With both controls active, the equal-error rate of the challenge's speaker-recognition attack rises from 10.4% to 49.9% (near chance), while intelligibility and emotion scores stay close to the challenge baselines. The paper concludes that these prosodic factors encode speaker identity and are exploited by the attack, and it further shows that how target speakers are chosen changes measured privacy by more than 20 EER points.

What carries the argument

The machinery has three parts. The base is kNN-VC, a voice converter that replaces each source speech feature with the average of its four nearest target-speaker features. The first addition is a phone predictor together with a duration predictor: the duration predictor estimates how long each phone should be from the phonetic transcript, and the source feature sequence is resampled to match that estimate, so the duration weight $w$ blends predicted and true durations. The second addition is per-phone k-means quantization of the target features: each phone's target features are clustered, and the cluster centers, rather than raw features, are used in the nearest-neighbor search, so the number of clusters directly controls how much per-phone variation survives. These two knobs allow the paper to remove phone-duration and phonetic-variation information independently and in combination, with privacy and utility read out through the challenge's equal-error-rate, word-error-rate, and emotion-recognition protocol.

What would settle it

A decisive check is to repeat the same duration and variation ablations under fully random target selection while keeping intelligibility and emotion scores matched; if equal-error rates no longer climb toward chance as the duration weight and cluster count change, the earlier gains came from target-selection confounding rather than from prosody. A complementary check is to give the attack enrollment and trial utterances from the same target speaker; if the near-chance scores collapse, the privacy was target-set confusion, not removal of source identity.

Watch

Extended reading notes

Core claim

The central claim is that two quantifiable, independently controllable properties of converted speech, the durations of phones and the variation of the feature representations assigned to each phone, carry source-speaker identity. The authors demonstrate this by ablating each property with one parameter, the duration weight $w$ and the number of clusters $k$, and observing monotonic increases in the attack's equal-error rate, up to 49.9% for the (3-16) configuration. Variation control is the stronger lever: 8 clusters alone reach 49.6% EER, and combining it with duration rewriting preserves utility while keeping near-chance privacy. The paper also finds that the attack's success depends heavily on target selection: same-gender or disjoint target sets give EERs near 50%, but fully random target selection caps EER at 29.4%, which the authors interpret as evidence that the attacker learns properties of the target set and that residual personally identifiable information remains in the anonymized speech.

Load-bearing premise

The privacy gains are caused by removing the intended prosodic dimensions rather than by generic speech degradation or by the attacker being confused by the way target speakers were chosen.

Editorial extensions

If this is right

  • Duration and variation should be treated as first-class privacy axes in anonymization, not incidental side effects of voice conversion.
  • The configurable knobs allow a gradual privacy-utility trade-off: for example, the (3-16) configuration gives near-chance EER with the same intelligibility and emotion scores as the uncombined (0-16).
  • Privacy scores are sensitive to target selection: gender-constrained or disjoint target sets can raise EER by more than 20 points, so evaluations should control or randomize target selection before interpreting EER as a measure of source anonymity.
  • Because the two techniques are complementary, stronger duration anonymization can buy back phonetic variation (more clusters) without sacrificing privacy.

Reading between the lines

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

  • Editorial inference: the same duration and variation ablations applied to non-kNN anonymizers would test whether these prosodic dimensions are general identity carriers or specific to the kNN-VC pipeline.
  • Editorial inference: the target-selection sensitivity implies EER conflates two distinct properties, source anonymization and target-set anonymization; an open-set or many-target evaluation would separate them.
  • Editorial inference: the phone-level controls could be extended to pitch and energy contours within the same framework, giving a systematic map of which prosodic features the attacker actually uses.
  • Editorial inference: making the number of clusters per phone speaker-dependent could test the paper's open question of whether some target speakers anonymize better than others.
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 Private kNN-VC, an interpretable extension of the kNN-VC voice conversion model for speaker anonymization. Two components are added: a duration predictor that rewrites phone durations from a phonetic transcript, and a per-phone k-means quantization of target features that limits phonetic variation. The system is evaluated with the VoicePrivacy 2024 objective protocol. The authors report that reducing phonetic variation raises the equal error rate (EER) from 10.4% (kNN-VC) to as high as 49.9% for configuration (3-16), and that duration anonymization alone raises EER to about 22.5%. They also report that changing target selection from same-gender to random decreases EER by more than 20 points for the strongest configurations, and they analyze this effect with gender-balanced disjoint target sets and varying numbers of target speakers. The paper concludes that phone duration and variation encode speaker identity and are exploited by the VPC 2024 attacker.

Significance. The paper addresses an important question in speaker anonymization: which speech properties are exploited by a privacy attack. The proposed system is genuinely interpretable and configurable, and the code and weights are released, which is a strength. The target-selection experiments are valuable because they show that the VPC 2024 EER is sensitive to the choice of target speakers, a protocol-level concern that affects the field's evaluation practice. If the causal claim were supported, the paper would be a useful contribution to interpretable anonymization. However, the central causal conclusion is not currently supported by the experiments, because the privacy gain largely disappears under random target selection and no control for generic speech degradation is provided. The evidence is suggestive rather than conclusive.

major comments (4)
  1. [Abstract and Section 7] The claim that the experiments 'prove' that phone duration and variation encode speaker identity is not supported by the reported evidence. Table 1 shows that under completely random target selection the best EER is 29.4% (10-8)r, versus 49.9% under same-gender selection for (3-16). Section 6 explicitly concedes that 'the attacker fails to exploit it for certain target selections' and that selection algorithms 'may confound the attacker, leading to an overestimation of privacy.' In this situation, the large EER difference between kNN-VC and the proposed configurations may reflect target-set confounding or generic degradation rather than the removal of the intended prosodic attributes. The causal statement should be softened to a hypothesis, or additional experiments must isolate the effects of duration and variation from the target-selection artifact.
  2. [Section 4.1, Figures 2-3, Table 1] All privacy and utility results are reported as single point estimates with no confidence intervals, significance tests, or repeated runs. The comparison between (0-8) at 49.6% EER and (3-16) at 49.9% EER, and the claim that these configurations are equivalent, cannot be evaluated without knowing the variability of the metric across random target selections, training runs of the attack model, or evaluation subsets. At least a small number of repetitions with reported standard deviations is needed for load-bearing comparisons.
  3. [Section 4.1 and Figure 2] The utility metrics (WER and UAR) decrease monotonically as privacy increases, so the observed EER gains may be due to generic degradation of the speech signal rather than to the removal of phone duration and variation as causally relevant identity cues. For example, reduced intelligibility and emotion recognition could indicate that the anonymized speech is simply lower quality or noisier, which would confuse the speaker recognition model. The paper should include a control condition that applies comparable degradation unrelated to the studied prosodic dimensions, such as random feature perturbation, additive noise, or vocoder-only artifacts, and measure the resulting EER. Without such a control, the causal interpretation is confounded.
  4. [Section 5.1 and Figure 4] The disjoint-target experiment is informative, but the interpretation that the attacker 'learns particularities about each target' is an inference from aggregate EER values, not directly tested. The alternative explanation is that with two disjoint target sets, each target speaker appears in fewer utterances, reducing the amount of training data per target for the attack model, which would also lower EER without implying that the attacker has learned target-specific details. The effect of number of targets in Figure 4 supports this alternative reading. A cleaner test would control for the number of utterances per target speaker while varying whether the target sets overlap across source genders.
minor comments (5)
  1. [Throughout] There are typographical artifacts in the text, including 'Tim P olzehl', 'V ariation', 'T erms', and 'V oicePrivacy'; these should be corrected.
  2. [Section 2.2] It is unclear whether the k-means clustering for phonetic variation is fitted per target speaker, per utterance, or pooled across all target speakers. Please specify, since this affects the interpretation of what 'phonetic variation' is being constrained.
  3. [Section 4.2] The statement that 'privacy can be improved by reducing phonetic variation' should be qualified as holding for the same-gender target selection protocol; under random target selection, the corresponding configurations in Table 1 show much smaller gains.
  4. [Figure 4] The figure caption and text do not report error bars or the range of results across repeated runs. The claim that privacy 'saturates at an EER of 27%' should be accompanied by uncertainty information.
  5. [Section 6] The limitation paragraph appropriately acknowledges residual PII and possible overestimation of privacy, but the conclusion in Section 7 reverts to unqualified causal language. The conclusion should reflect the limitations stated in Section 6.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ablations are evaluated against the external VPC 2024 attacker, and the model components are trained on independent phonetic and prosodic labels, not on the privacy metric.

full rationale

The paper's central claim is an ablation claim: adding phone-duration anonymization and phonetic-variation quantization to kNN-VC raises the EER of the VPC 2024 attack. The components are not fitted to EER or to the attack model. The phone predictor is trained on LibriSpeech aligned phones with cross-entropy and CTC losses (Section 2.1); the duration predictor is trained on LJSpeech durations with MSE (Section 2.3). Neither loss includes privacy or identity information. Privacy is measured with the official VPC 2024 protocol, an externally defined speaker recognizer trained on anonymized LibriSpeech, and utility is measured with independent ASR and emotion models (Section 3). No equation in the paper defines a privacy outcome in terms of the fitted predictors; Equation 1 only mixes predicted and true durations. The conclusion that phone duration and variation encode speaker identity is an interpretation of ablation outcomes, not a construction that presupposes the conclusion. The self-citations [6], [9], and [19] are not load-bearing: [6] and [9] are listed alongside external references [7] and [8] as prior observations of kNN-VC's low privacy, and [19] merely describes the SpAnE implementation framework. The target-selection findings in Section 5 and the limitations in Section 6 weaken the causal 'proving' claim by showing that disjoint target sets can confound the attacker and overestimate privacy, but that is a validity concern, not circularity. There is no fitted parameter renamed as a prediction and no imported uniqueness theorem. Overall, the evaluation is self-contained against an external benchmark, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities are proposed. The central claim depends on the VPC 2024 EER as a privacy proxy, on the phone and duration predictors behaving accurately enough to isolate prosody, and on the clustering uniformly controlling phonetic variation. The trained predictor weights are not listed because they are not tuned to privacy outcomes; the only hand-chosen values are the ablation parameters w and the cluster count.

free parameters (2)
  • duration weight w = 0, 0.3, 0.7, 1.0
    Hand-chosen ablation values controlling how much predicted phone duration is used. It is an independent variable swept in experiments, not fitted to privacy data, so it does not add a hidden degree of freedom to the causal claim.
  • number of k-means clusters per phone = 8, 16, 32, or none
    Hand-chosen values controlling phonetic variation. Again a controlled experimental condition rather than a parameter fitted to optimize privacy.
assumptions (4)
  • domain assumption VPC 2024 equal error rate on a speaker verifier trained on anonymized speech is a valid proxy for speaker privacy.
    The whole causal story is read off EER changes. Section 6 admits that target selection may confound the attacker, so this axiom is load-bearing and partially challenged by the paper itself.
  • domain assumption Phone predictor outputs (ACC=92%, PER=2%) and the transcript from duplicate removal are accurate enough that duration and variation edits isolate prosody rather than introduce unrelated errors.
    Section 2.1 reports accuracy on a held-out test set, but the downstream causal attribution assumes residual phone errors do not drive the EER changes.
  • domain assumption k-means clustering of per-phone target features reduces phonetic variation in a uniform way across target speakers.
    Section 2.2 controls variation via cluster count, but Section 6 states: 'We cannot ensure that the clustering algorithm constrains phonetic variation in the same way for all target speakers.'
  • domain assumption FastSpeech2-style variance prediction and WavLM representations behave as described in their source papers.
    The method inherits prior models without re-validating them. This is routine in the field and not especially risky, but it is still an unproved background premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Private kNN-VC: Interpretable Anonymization of Converted Speech." pith.science (2026). https://pith.science/paper/GNCRONDB

@misc{pith2026250517584,
  author       = {Pith},
  title        = {Pith review of: Private kNN-VC: Interpretable Anonymization of Converted Speech},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GNCRONDB}},
  note         = {Machine review of arXiv:2505.17584}
}
read the original abstract

Speaker anonymization seeks to conceal a speaker's identity while preserving the utility of their speech. The achieved privacy is commonly evaluated with a speaker recognition model trained on anonymized speech. Although this represents a strong attack, it is unclear which aspects of speech are exploited to identify the speakers. Our research sets out to unveil these aspects. It starts with kNN-VC, a powerful voice conversion model that performs poorly as an anonymization system, presumably because of prosody leakage. To test this hypothesis, we extend kNN-VC with two interpretable components that anonymize the duration and variation of phones. These components increase privacy significantly, proving that the studied prosodic factors encode speaker identity and are exploited by the privacy attack. Additionally, we show that changes in the target selection algorithm considerably influence the outcome of the privacy attack.

Figures

Figures reproduced from arXiv: 2505.17584 by the authors.

Figure 1
Figure 1. Private kNN-VC: purple components are the same as those of kNN-VC. S is the source speech, T is the target speaker’s speech, and A is the anonymized speech. age of its four nearest target features according to cosine similarity. This works because the representation space of speech language models like WavLM is heavily influenced by phones [21]. The converted features are synthesized with a HiFi-GAN vocoder [22] tra… view at source ↗
Figure 2
Figure 2. Privacy vs. intelligibility [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [1]

    The evaluation simulates an attack on the anonymized speech, leveraging a speaker recognition model trained to discriminate anonymized speakers by learning their characteristics

    Introduction Speaker anonymization has become an active field of research since the first edition of the V oicePrivacy Challenge (VPC) [1], which provides an evaluation protocol to assess the strength of anonymization systems. The evaluation simulates an attack on the anonymized speech, leveraging a speaker recognition model trained to discriminate anonym...

    work page 2024

  2. [2]

    Private kNN-VC: Interpretable Anonymization of Converted Speech

    Anonymization system kNN-VC [5] is an any-to-any VC model. It first extracts features from both source and target speech using WavLM [20], a speech language model. Then, each source feature is replaced with the aver- 1https://github.com/carlosfranzreb/private_ knnvc arXiv:2505.17584v1 [eess.AS] 23 May 2025 Figure 1:Private kNN-VC: purple components are th...

    work page Pith review arXiv 2025

  3. [3]

    As target speakers, we use 100 speakers from the LibriTTS [30]train-other-500dataset, evenly distributed across genders

    Evaluation We evaluate our anonymization system with the official imple- mentation of the VPC 2024 [3]. As target speakers, we use 100 speakers from the LibriTTS [30]train-other-500dataset, evenly distributed across genders. The target selection algorithm randomly selects a target speaker while preserving the source speaker’s gender. It is applied indepen...

    work page 2024

  4. [4]

    For each parameter, we have selected four values

    Experiments We experiment with different configurations of our model to investigate how each of the prosodic aspects (duration and variation of phones) affect the privacy attack. For each parameter, we have selected four values. Phonetic variation is anonymized with 32, 16 or 8 clusters, or not anonymized, which we define as zero in our configuration iden...

    work page

  5. [5]

    Randomness ensuresunlinkability, i.e

    Effect of target selection on privacy Our model picks target speakers randomly, but preserves the gender of the source speaker. Randomness ensuresunlinkability, i.e. that source speakers cannot be identified by attacking solely the target selection algorithm [12]. Preserving gender is a reasonable requirement for many use cases, as it facilitates fairer e...

    work page

  6. [6]

    Each experiment takes roughly 5 hours on a RTXA6000 GPU

    Zero means that the phone durations are not modified, and one means that the predicted phone durations are used without summing them with the true durations at all. Each experiment takes roughly 5 hours on a RTXA6000 GPU. We compare the proposed system with the original kNN-VC and the baselines from the VPC 2024 [3]. 4.1. Phone duration and variation Figu...

    work page 2024

  7. [7]

    Each anonymization technique is specific and con- figurable, allowing us to research its effect on utility and privacy

    Conclusion We propose an interpretable extension of kNN-VC for speaker anonymization. Each anonymization technique is specific and con- figurable, allowing us to research its effect on utility and privacy. Our experiments show that the variation and duration of phones encode speaker identity, which the VPC 2024 attacker exploits to identify speakers. Cons...

    work page 2024

  8. [8]

    Limitations We identify three areas of improvement for the proposed system: privacy, interpretability and emotion preservation. The large difference in EER when the target selection algorithm is changed leads us to believe that there is still personally identifiable information (PII) in the anonymized speech produced by our system, although the attacker f...

    work page

Show all 45 references
  1. [9]

    Comparing speech anonymization efficacy by voice conversion using knn and disentangled speaker feature representations,

    A. Das, C. Franzreb, T. Herzig, P . Pirlet, and T. Polzehl, “Comparing speech anonymization efficacy by voice conversion using knn and disentangled speaker feature representations,” in4th Symposium on Security and Privacy in Speech Communication. ISCA, Sep. 2024, pp. 121–126

    2024

  2. [10]

    Acknowledgements Funded by Federal Ministry of Education and Research, Germany (BMBF 16KIS2048)

  3. [11]

    Introducing the V oicePrivacy Initiative,

    N. Tomashenko, B. M. L. Srivastava, X. Wang, E. Vincent, A. Nautsch, J. Y amagishi, N. Evans, J. Patino, J.-F. Bonastre, P .-G. No´e, and M. Todisco, “Introducing the V oicePrivacy Initiative,” inInterspeech

  4. [12]

    Anonymizing Speech: Evaluating and Designing Speaker Anonymization Techniques,

    P . Champion, “Anonymizing Speech: Evaluating and Designing Speaker Anonymization Techniques,” Ph.D. dissertation, University of Lorraine, 2023

    2023

  5. [13]

    Safe- guarding Speech Content Style: Enhancing Privacy Beyond Speaker Identity,

    Y . Sinha, M. Raivakhovskyi, M. Schubert, and I. Siegert, “Safe- guarding Speech Content Style: Enhancing Privacy Beyond Speaker Identity,” in4th Symposium on Security and Privacy in Speech Communication. ISCA, Sep. 2024, pp. 92–101

    2024

  6. [14]

    The voiceprivacy 2024 challenge evaluation plan,

    N. Tomashenko, X. Miao, P . Champion, S. Meyer, X. Wang, E. Vincent, M. Panariello, N. Evans, J. Y amagishi, and M. Todisco, “The voiceprivacy 2024 challenge evaluation plan,”arXiv preprint arXiv:2404.02677, 2024

    2024 arXiv

  7. [15]

    An Overview of V oice Conversion and Its Challenges: From Statistical Modeling to Deep Learning,

    B. Sisman, J. Y amagishi, S. King, and H. Li, “An Overview of V oice Conversion and Its Challenges: From Statistical Modeling to Deep Learning,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 132–157, 2021

    2021

  8. [16]

    V oice Conversion With Just Nearest Neighbors,

    M. Baas, B. van Niekerk, and H. Kamper, “V oice Conversion With Just Nearest Neighbors,” inInterspeech 2023. ISCA, Aug. 2023, pp. 2053–2057

    2023

  9. [17]

    Towards Audiovisual Anonymization for Remote Psychotherapy: a Subjective Evaluation,

    C. Franzreb, A. Das, H. Gieseler, E. C. Jahn, T. Polzehl, and S. M¨oller, “Towards Audiovisual Anonymization for Remote Psychotherapy: a Subjective Evaluation,” in4th Symposium on Security and Privacy in Speech Communication. ISCA, 2024, pp. 102–110

    2024

  10. [18]

    Privacy versus emotion preservation trade-offs in emotion-preserving speaker anonymization,

    Z. Cai, H. Li Xinyuan, A. Grag, L. P . Garc ´ıa-Perera, K. Duh, S. Khudanpur, N. Andrews, and M. Wiesner, “Privacy versus emotion preservation trade-offs in emotion-preserving speaker anonymization,” inIEEE Spoken Language T echnology W orkshop, 2024

    2024

  11. [19]

    Anonymising Elderly and Pathological Speech: V oice Conversion Using DDSP and Query-by-Example,

    S. Ghosh, M. Jouaiti, A. Das, Y . Sinha, T. Polzehl, I. Siegert, and S. Stober, “Anonymising Elderly and Pathological Speech: V oice Conversion Using DDSP and Query-by-Example,” inInterspeech

  12. [20]

    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. Y oshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, M. Zeng, X. Y u, and F. Wei, “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected T op...

    2022

  13. [21]

    Analysing Discrete Self Supervised Speech Representation For Spoken Language Modeling,

    A. Sicherman and Y . Adi, “Analysing Discrete Self Supervised Speech Representation For Spoken Language Modeling,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Jun. 2023, pp. 1–5, iSSN: 2379-190X

    2023

  14. [22]

    Jurafsky and J

    D. Jurafsky and J. H. Martin,Speech and language processing: an introduction to natural language processing, computational linguistics, and speech recognition, ser. Prentice Hall series in artificial intelligence. Upper Saddle River, N.J: Prentice Hall, 2000

    2000

  15. [23]

    It consists of the Librispeech train-clean-100dataset [24], for which the aligned phones were predicted with a g2p model

    for their phone recognition task. It consists of the Librispeech train-clean-100dataset [24], for which the aligned phones were predicted with a g2p model. The alphabet is based on the CMU dictionary2: there are 41 phones and lexical stress is ignored. 80% of the data is used ...

  16. [24]

    Disentangling prosody and timbre embeddings via voice conversion,

    N. Gengembre, O. Le Blouch, and C. Gendrot, “Disentangling prosody and timbre embeddings via voice conversion,” inInterspeech 2024. ISCA, Sep. 2024, pp. 2765–2769

    2024

  17. [25]

    Speaker de- identification via voice transformation,

    Q. Jin, A. R. Toth, T. Schultz, and A. W. Black, “Speaker de- identification via voice transformation,” in2009 IEEE W orkshop on Automatic Speech Recognition & Understanding, Nov. 2009, pp. 529–533

    2009

  18. [26]

    Analysis of Speech Temporal Dynamics in the Context of Speaker V erification and V oice Anonymization,

    N. Tomashenko, E. Vincent, and M. Tommasi, “Analysis of Speech Temporal Dynamics in the Context of Speaker V erification and V oice Anonymization,” in2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  19. [27]

    Explainability for speech models: On the challenges of acoustic feature selection,

    D. Fucci, B. Savoldi, M. Gaido, M. Negri, M. Cettolo, L. Bentivogli et al., “Explainability for speech models: On the challenges of acoustic feature selection,” inProceedings of the T enth Italian Conference on Computational Linguistics (CLiC-it 2024), 2024

    2024

  20. [28]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,”Nature Machine Intelligence, vol. 1, no. 5, pp. 206–215, May 2019

    2019

  21. [29]

    Prosody Is Not Identity: A Speaker Anonymization Approach Using Prosody Cloning,

    S. Meyer, F. Lux, J. Koch, P . Denisov, P . Tilli, and N. T. Vu, “Prosody Is Not Identity: A Speaker Anonymization Approach Using Prosody Cloning,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Jun. 2023, pp. 1–5

    2023

  22. [30]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Y . Ren, C. Hu, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y . Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” in International Conference on Learning Representations, 2021

    2021

  23. [31]

    A Comprehensive Evaluation Framework for Speaker Anonymization Systems,

    C. Franzreb, T. Polzehl, and S. M¨oller, “A Comprehensive Evaluation Framework for Speaker Anonymization Systems,” in3rd Symposium on Security and Privacy in Speech Communication.ISCA, 2023, pp. 65–72

    2023

  24. [32]

    The voiceprivacy 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. Y amagishi, “The voiceprivacy 2022 challenge: Progress and perspectives in voice anonymisation,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 32, p. 3477–34...

    2022

  25. [33]

    Two-stage and Self-supervised V oice Conversion for Zero-Shot Dysarthric Speech Reconstruction,

    D. Liu, Y . Lin, H. Bu, and M. Li, “Two-stage and Self-supervised V oice Conversion for Zero-Shot Dysarthric Speech Reconstruction,” International Conference on Asian Language Processing, 2024

    2024

  26. [34]

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

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,”Advances in Neural Information Processing Systems, vol. 33, pp. 17 022–17 033, 2020

    2020

  27. [35]

    SUPERB: Speech Processing Universal PERformance Benchmark,

    S.-w. Y ang, P .-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K.-t. Lee, D.-R. Liu, Z. Huang, S. Dong, S.-W. Li, S. Watanabe, A. Mohamed, and H.-y. Lee, “SUPERB: Speech Processing Universal PER...

    2021

  28. [36]

    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,” inProc. ICASSP 2015 - 40 th Annual IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brisbane, Australia, Apr. 2015, pp. 5206–5210

    2015

  29. [37]

    Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fern´andez, F. Gomez, and J. Schmidhuber, “Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” inProceedings of the 23rd International Conference on Machine Learning, ser. ICML ’06. New Y ork, NY , USA: As...

    2006

  30. [38]

    Zero-Shot Sing V oice Conversion: built upon clustering-based phoneme representations,

    W. Zhou, F. Zhang, Y . Liu, W. Guan, Y . Zhao, and H. Qu, “Zero-Shot Sing V oice Conversion: built upon clustering-based phoneme representations,” Sep. 2024, arXiv:2409.08039 [cs, eess]

    2024 arXiv

  31. [39]

    SKQVC: One-Shot V oice Con- version by K-Means Quantization with Self-Supervised Speech Representations,

    Y . Sim, J. Y oon, and Y .-J. Suh, “SKQVC: One-Shot V oice Con- version by K-Means Quantization with Self-Supervised Speech Representations,” Nov. 2024, arXiv:2411.16147

    2024 arXiv

  32. [40]

    A k-means clustering algorithm,

    J. A. Hartigan, M. A. Wonget al., “A k-means clustering algorithm,” Applied statistics, vol. 28, no. 1, pp. 100–108, 1979

    1979

  33. [41]

    The lj speech dataset,

    K. Ito and L. Johnson, “The lj speech dataset,” https: //keithito.com/LJ-Speech-Dataset/, 2017

    2017

  34. [42]

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

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text-to-speech,” inInterspeech 2019, 2019, pp. 1526–1530

    2019

  35. [43]

    Iemocap: Interactive emotional dyadic motion capture database,

    C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, “Iemocap: Interactive emotional dyadic motion capture database,”Language resources and evaluation, vol. 42, pp. 335–359, 2008

    2008

  36. [2020]

    1693–1697

    ISCA, 2020, pp. 1693–1697

    2020

  37. [2024]

    2024, pp

    ISCA, Sep. 2024, pp. 4438–4442

    2024

Pith tools

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