Pith. sign in

REVIEW 4 major objections 7 minor 55 references

VocalCrypt: Novel Active Defense Against Deepfake Voice Based on Masking Effect

T0 review · 4 major / 7 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 By burying pseudo-timbre below the auditory masking threshold, VocalCrypt aims to make AI voice cloning produce speech that fails speaker verification while leaving the audio perceptually unchanged.

desk verdict Plausible masking-based defense idea, but missing no-defense baseline and inverted robustness metric leave the central claim unmeasured. read the letter →

arxiv 2502.10329 v1 pith:IRJHCHSB submitted 2025-02-14 cs.SD cs.CRcs.MMeess.AS

Qingyuan Fei , Wenjie Hou , Xuan Hai , Xin Liu This is my paper · ORCID
classification cs.SDcs.CRcs.MMeess.AS
keywords voicecloningdefenseactivemaskingeffectpseudo-timbrequantizationindexmodulationspeakerverificationconversionaudiorobustness
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

VocalCrypt is an active defense against AI voice cloning that hides 'pseudo-timbre' signals in parts of an audio recording the human ear cannot perceive, using the psychoacoustic masking effect. The paper claims this makes voice-conversion and text-to-speech cloning models learn the wrong voice identity, lowering automatic speaker verification similarity between original and cloned speech while leaving perceived audio quality intact. Tests across Chinese and English speech and five commercial and open-source cloning systems report post-protection similarity scores below the same-speaker threshold, and the paper argues the method survives denoising and resampling attacks that break GAN-based adversarial defenses. It also reports generation speed about five times higher than those defenses, which the authors say makes real-time, preemptive protection practical.

What carries the argument

The load-bearing mechanism is the adaptive masking-threshold calculation: each of 25 Bark-scale critical bands gets an absolute masking threshold, corrected by the spectral flatness measure and a tonal coefficient, and the embedding strength is set so the quantization noise stays at least 5 dB below that threshold. Pseudo-timbre is then embedded by quantization index modulation on DCT coefficients obtained from a three-layer Haar wavelet decomposition. The calculation's job is to put jamming energy exactly where human perception will not notice it but where voice-conversion and TTS feature extractors will.

What would settle it

Re-run the same five cloning systems on the same Chinese and English test sentences without VocalCrypt protection and measure ASV similarity between genuine and cloned audio. If the unprotected baseline already averages at or below the reported 0.4 to 0.5 range, the observed scores do not demonstrate any defense effect.

Watch

Extended reading notes

Core claim

The central claim is that the auditory masking threshold is a usable hiding place for defensive jamming information. VocalCrypt computes a per-critical-band masking threshold from spectral flatness and a tonal coefficient, then embeds pseudo-timbre with quantization index modulation in the low-frequency Bark bands where speech energy concentrates and the ear is least sensitive. The embedded signal is inaudible to listeners but is picked up by cloning systems, which then synthesize speech resembling the pseudo-timbre rather than the target speaker. The paper reports protected-audio quality near 0.94 while post-protection speaker similarity averages roughly 0.43 to 0.52 across five cloning models, and argues this makes the defense both effective and perceptually transparent.

Load-bearing premise

The evaluation assumes the drop in speaker-similarity scores is caused by VocalCrypt, but it reports only scores for protected audio and never compares against cloning the same speakers' unprotected recordings.

Editorial extensions

If this is right

  • If the central claim holds, a speaker can pre-emptively protect a recording before it is distributed, rather than relying on post-hoc forgery detection.
  • Protected audio remains useful for listening: quality scores near 0.94 indicate the defense can be applied without perceptible degradation.
  • Because the interference survives denoising and resampling to 8 kHz, attackers cannot easily strip the defense with common preprocessing.
  • The reported speedup makes real-time embedding feasible, opening the door to protection at the microphone or sound-card level.

Reading between the lines

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

  • The paper reports no baseline in which the same cloning systems clone unprotected recordings; the decisive check is whether unprotected cloning scores are substantially higher than the reported 0.4 to 0.5 range.
  • If the masking-threshold hiding strategy generalizes, the same pseudo-timbre idea could protect other biometric signals, such as video identity embeddings, though the paper does not explore this.
  • The 500% speed comparison is against GAN-based adversarial defenses on protection generation time; a full deployment comparison would include cloning and verification time.
  • The defense's success presumably depends on cloning models actually consuming the low-frequency DCT bands where the pseudo-timbre is hidden; a cloning pipeline that strips or ignores those bands could evade it.
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 / 7 minor

Summary. VocalCrypt is an active defense method that embeds pseudo-timbre signals into audio below the psychoacoustic masking threshold, using Bark-scale critical-band decomposition, spectral flatness-based masking thresholds, and quantization index modulation (QIM). The goal is to disrupt AI voice conversion and TTS cloning while preserving the perceptual quality of the protected audio. The paper evaluates the method against five cloning models (ElevenLabs, GPT-SoVITS, XTTSv2, SEED-VC, StyleTTS2) on Chinese and English speech, comparing it with Huang's adversarial-perturbation method and Dong's GAN-based method. It reports ASV similarity, a quality column, robustness to denoising and downsampling, and processing time, and claims stronger defense, better quality, robustness, and a 500% generation-speed improvement.

Significance. The idea of using psychoacoustic masking thresholds to embed jamming information is reasonable and distinct from prior adversarial-noise defenses, and the manuscript correctly identifies limitations of post-hoc watermarking. The reliance on standard masking models and QIM is a strength in principle, and the robustness tests against denoising and downsampling address a practically relevant threat model. However, the experimental evaluation as presented does not demonstrate that VocalCrypt changes cloning outcomes, and the internal comparison in Table II contradicts the claimed superiority. If properly re-evaluated with baselines and a consistent metric direction, the approach could be of interest to the audio security community.

major comments (4)
  1. [Section V-B, Table II] The reported ASV similarity scores are all post-protection; no row reports the score obtained by cloning the original, unprotected audio. Without such a baseline, the paper cannot attribute the observed values (e.g., Ours average 0.474) to the defense, and the central claim that VocalCrypt 'actively lowers' ASV similarity is unmeasurable. This is load-bearing for every efficacy statement in Sections V-B and V-C.
  2. [Section V-B, Table II and Section V-C, Table III] The data contradict the paper's superiority claim. Under the paper's own rule that lower ASV indicates better defense, Dong's method has a lower mean ASV (0.460) than Ours (0.474) in Table II, yet the text states that Ours shows the 'strongest protective effect.' Table III repeats the direction error: the header 'ASV Similarity↑' and the phrase 'highest average score' treat higher values as better, which is exactly opposite to the rule stated in Section V-B. These inconsistencies undermine the reported comparisons.
  3. [Section IV-C and IV-D] Several equations use undefined or inconsistent quantities: Eq. (3) uses SFi while the text refers to SFj; Eq. (6) uses coefficient 'a' although Eq. (5) defines α; Eq. (10) introduces Ij without definition and lj is never used; Eq. (14) defines Ij as a 'scaling factor' but gives no procedure for setting it. Because these equations determine the embedding strength that is the core of the method, the algorithm is not reproducible as written.
  4. [Section V-A3, Table II] The 'Quality' column in Table II is not tied to any described metric. Section V-A3 defines only the SVA/ASV similarity metric; no subjective listening test, objective speech quality metric (e.g., PESQ, STOI), or MOS protocol is described. Consequently, the claims of 'near-perfect speech quality' and perceptual preservation are unsupported.
minor comments (7)
  1. [Abstract vs. Section IV-C] The abstract uses 'SFS' while Section IV-C uses 'SFM' and 'spectral flatness measure'; please standardize the terminology.
  2. [Section IV-A and IV-B] Section IV-A says critical bands 1 through 7 are used for embedding, whereas Section IV-B says bands 1 through 9 are selected; please reconcile the inconsistency.
  3. [Abstract/Introduction vs. Section V-A1] The dataset names are inconsistent across the abstract and introduction (Zhvoice, VCTK) and Section V-A1 (CSTR VCTK and zhvoice); please unify.
  4. [Section V-A3 vs. V-B] Section V-B states that ASV values above 0.9 indicate the same speaker, while Section V-A3 sets the threshold at 0.8; please correct the inconsistency.
  5. [Section V-A2] Section V-A2 mentions 'Huang's and Wang's models' but Table II compares Huang's, Dong's, and Ours; no Wang method is presented, suggesting a typo.
  6. [Equation (7)] Equation (7) is ambiguous: '10lg Cj −Oj /10' could be read as 10 log(Cj) - Oj/10 or 10 log((Cj-Oj)/10); please add parentheses.
  7. [Fig. 4] Fig. 4 lacks axis labels and units, and the 500% speed-up claim is not supported by a table of raw processing times.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the masking-threshold embedding is derived from standard psychoacoustic models, and the ASV evaluation uses an external x-vector system; the main weakness is a missing unprotected baseline, which is an experimental control problem, not a circularity.

full rationale

The derivation chain in VocalCrypt is self-contained rather than circular. The pseudo-timbre embedding strength in Eq. (14) is computed from the masking threshold Tz(j), which in turn is obtained from the standard Bark-scale psychoacoustic model and spectral flatness measure (Eqs. (1)-(10)), citing established sources such as Zwicker [39] and Johnston [42]; the paper does not fit any parameter to the ASV similarity scores it later reports. The evaluation metric is an external automatic speaker verification system (x-vectors, [53]), and the claimed defense effect is a measured outcome rather than a quantity constructed from the method's own equations. There are no load-bearing self-citations: the references to the authors' own prior work are absent, and the uniqueness of the approach is asserted from the design itself rather than imported from a cited theorem by the same authors. The robustness and speed claims, while weakly documented, are empirical measurements, not identities. The serious weakness is that Table II reports ASV similarity only after protection, with no row for cloning the original unprotected audio, so a drop in similarity cannot be demonstrated from the reported numbers; Table III also contains a direction inconsistency in its header ('ASV Similarity↑') relative to Section V-B's lower-is-better rule. These are experimental-control and reporting flaws, not circular reasoning, because the measured quantity is not defined in terms of the method's inputs and no fitted value is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method relies on psychoacoustic masking theory from prior literature (Zwicker, Johnston) as an external benchmark, so it is not internally circular. However, it assumes without validation that sub-threshold pseudo-timbre is both inaudible and disruptive to AI cloning models, and the experiment lacks a no-defense control that would let the reader attribute ASV changes to the defense.

free parameters (4)
  • Frame length L = 504 samples
    Set to 504 in Section IV-B to obtain 7.94 Hz DCT resolution; no optimization or justification is given.
  • Absolute energy spectral density S = Not specified
    Introduced in Section IV-C Eq. (10) for converting dB thresholds to DCT energy; the value is never given.
  • Band scaling factor I_j = Not specified
    Used in Eqs. (10) and (14) as a per-band scaling factor but never defined or calibrated.
  • Tonal coefficient a = Undefined
    Appears in Eq. (6) for the masking threshold correction but is never defined or assigned a value.
assumptions (4)
  • domain assumption The Bark-scale critical band model and Johnston/Zwicker masking thresholds describe human auditory perception accurately for this embedding.
    Section IV-B and IV-C rely on Zwicker and Johnston to define inaudibility; if the thresholds are wrong, the embedded pseudo-timbre may be audible or too weak to matter.
  • domain assumption Pseudo-timbre placed below the masking threshold remains imperceptible to humans.
    Assumed in Section IV-D and used to claim preserving naturalness; no listening test or subjective evaluation is reported.
  • domain assumption ASV similarity measured by x-vectors is the correct proxy for voice-cloning success.
    Section V-A3 uses an ASV score threshold of 0.8 to define same-speaker; defense effectiveness is defined only as lowering this score.
  • ad hoc to paper The reported post-protection ASV scores can be attributed to VocalCrypt without an unprotected baseline.
    Table II reports only post-protection scores; without a no-defense control the experiment cannot isolate the defense effect.
invented entities (1)
  • Pseudo-timbre (pseudo-tone) signal
    purpose: Embedded into inaudible regions of audio to mislead voice conversion and TTS models into cloning a fake speaker identity.
    The paper defines pseudo-timbre only by its effect on ASV scores in its own experiments; no separate falsifiable prediction or external measurement is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VocalCrypt: Novel Active Defense Against Deepfake Voice Based on Masking Effect." pith.science (2026). https://pith.science/paper/IRJHCHSB

@misc{pith2026250210329,
  author       = {Pith},
  title        = {Pith review of: VocalCrypt: Novel Active Defense Against Deepfake Voice Based on Masking Effect},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRJHCHSB}},
  note         = {Machine review of arXiv:2502.10329}
}
read the original abstract

The rapid advancements in AI voice cloning, fueled by machine learning, have significantly impacted text-to-speech (TTS) and voice conversion (VC) fields. While these developments have led to notable progress, they have also raised concerns about the misuse of AI VC technology, causing economic losses and negative public perceptions. To address this challenge, this study focuses on creating active defense mechanisms against AI VC systems. We propose a novel active defense method, VocalCrypt, which embeds pseudo-timbre (jamming information) based on SFS into audio segments that are imperceptible to the human ear, thereby forming systematic fragments to prevent voice cloning. This approach protects the voice without compromising its quality. In comparison to existing methods, such as adversarial noise incorporation, VocalCrypt significantly enhances robustness and real-time performance, achieving a 500\% increase in generation speed while maintaining interference effectiveness. Unlike audio watermarking techniques, which focus on post-detection, our method offers preemptive defense, reducing implementation costs and enhancing feasibility. Extensive experiments using the Zhvoice and VCTK Corpus datasets show that our AI-cloned speech defense system performs excellently in automatic speaker verification (ASV) tests while preserving the integrity of the protected audio.

Figures

Figures reproduced from arXiv: 2502.10329 by the authors.

Figure 1
Figure 1. Architecture of the VocalCrypt on the Bark scale, segmenting the auditory range from 20 Hz to 22.05 kHz into 25 bands (as shown in Table I). The basis for this division lies in the fact that the Bark scale, by defining the ”critical bandwidth,” transforms frequency into a psychoacoustic unit that approximates auditory resolution [45]. Specifically, when the frequency separation between two components is smaller than… view at source ↗
Figure 2
Figure 2. Schematic diagram of human ear sensitivity to audio [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of Mel spectrograms between original audio and protected [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Reciprocal of Processing Time Comparison of Different Models [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 51 canonical work pages

  1. [1]

    Music streaming and digital music consumption in sweden,

    J. W. Morris and D. Powers, “Music streaming and digital music consumption in sweden,” Popular Communication, vol. 15, no. 2, pp. 124–138, 2017

    work page 2017

  2. [2]

    The playlist experience: Personal playlists in music streaming services,

    A. N. Hagen, “The playlist experience: Personal playlists in music streaming services,” Popular Music and Society, vol. 38, no. 5, pp. 625–645, 2015

    work page 2015

  3. [3]

    Spotify streaming: patterns and prediction,

    H. Datta, G. Knox, and B. J. Bronnenberg, “Spotify streaming: patterns and prediction,” ACM Transactions on Intelligent Systems and Technology (TIST) , vol. 10, no. 1, pp. 1–30, 2018

    work page 2018

  4. [4]

    The impact of streaming on music industry revenues–a scenario analysis,

    N. Wl ¨omert and D. Papies, “The impact of streaming on music industry revenues–a scenario analysis,” Marketing Science, vol. 38, no. 6, pp. 930–947, 2019

    work page 2019

  5. [5]

    A com- parison of features for synthetic speech detection,

    M. Sahidullah, T. Kinnunen, and C. Hanilc ¸i, “A com- parison of features for synthetic speech detection,” in INTERSPEECH 2015 , pp. 2087–2091, 2015

    work page 2015

  6. [6]

    The DeepFake Detection Challenge Dataset,

    B. Dolhansky, J. Bitton, B. Pflaum, J. Lu, R. Howes, M. Wang, and C. C. Ferrer, “The DeepFake Detection Challenge Dataset,” arXiv, Oct. 2020

    work page 2020

  7. [7]

    Exposing deepfake videos by detecting face warping artifacts,

    Y . Li and S. Lyu, “Exposing deepfake videos by detecting face warping artifacts,” in The IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops, 2019

    work page 2019

  8. [8]

    Fakecatcher: Detec- tion of synthetic portrait videos using biological signals,

    U. A. Ciftci, I. Demir, and L. Yin, “Fakecatcher: Detec- tion of synthetic portrait videos using biological signals,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–1, 2020

    work page 2020

Show all 55 references
  1. [9]

    Artificial intelligence makes voice cloning easy and ’the monster is already on the loose’

    “Artificial intelligence makes voice cloning easy and ’the monster is already on the loose’.”

  2. [10]

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

    J.-c. Chou, C.-c. Yeh, H.-y. Lee, and L.-s. Lee, “One-shot voice conversion by separating speaker and content rep- resentations with instance normalization,” in Interspeech, 2019

    2019

  3. [11]

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

    K. Qian, Y . Zhang, S. Chang, X. Yang, and M. Hasegawa-Johnson, “Autovc: Zero-shot voice style transfer with only autoencoder loss,” Advances in Neural Information Processing Systems , 2019

    2019

  4. [12]

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerry-Ryan, et al. , “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICAS...

    2018

  5. [13]

    Fastspeech: Fast, robust and controllable text to speech,

    Y . Ren, Y . Ruan, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.- Y . Liu, “Fastspeech: Fast, robust and controllable text to speech,” in Advances in Neural Information Processing Systems, 2019

    2019

  6. [14]

    Speech watermarking with controllable robustness using deep steganography,

    G. Zhang, Y . Xu, C. Zhang, et al., “Speech watermarking with controllable robustness using deep steganography,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 3249–3261, 2023

    2023

  7. [15]

    Robust au- dio watermarking via neural steganography learning,

    Z. Wang, J. Wu, Z. Wu, and D. Wang, “Robust au- dio watermarking via neural steganography learning,” in ICASSP 2023 , pp. 1–5, 2023

    2023

  8. [16]

    Deepmark: Deep learning- based audio watermarking for copyright protection,

    W. Chen, Y . Liu, and X. Sun, “Deepmark: Deep learning- based audio watermarking for copyright protection,” IEEE Transactions on Information F orensics and Secu- rity, 2024

    2024

  9. [17]

    V oid: A fast and light voice liveness detection system,

    M. Ahmed, I.-Y . Kwak, J. Huh, I. Kim, T. Oh, and H. Kim, “V oid: A fast and light voice liveness detection system,” in Proceedings of the 29th USENIX Conference on Security Symposium , pp. 2685–2702, 2020

    2020

  10. [18]

    An audio captcha to distinguish humans from computers,

    H. Gao, H. Liu, D. Yao, X. Liu, and U. Aickelin, “An audio captcha to distinguish humans from computers,” in 2010 Third International Symposium on Electronic Commerce and Security , pp. 265–269, IEEE, 2010

    2010

  11. [19]

    ”hello? who am i talking to?

    A. Lieto, D. Moro, F. Devoti, C. Parera, V . Lipari, P. Bestagini, and S. Tubaro, “”hello? who am i talking to?” a shallow cnn approach for human vs. bot speech classification,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASS...

    2019

  12. [20]

    Deepsonar: Towards effective and robust detection of ai-synthesized fake voices,

    R. Wang, F. Juefei-Xu, Y . Huang, Q. Guo, X. Xie, L. Ma, and Y . Liu, “Deepsonar: Towards effective and robust detection of ai-synthesized fake voices,” in Proceedings of the 28th ACM international conference on multimedia , pp. 1207–1216, 2020

    2020

  13. [21]

    Deep- sonar: Towards effective and robust detection of ai- synthesized fake voices,

    X. Zhang, Y . Yang, N. Chen, Y . Chen, et al. , “Deep- sonar: Towards effective and robust detection of ai- synthesized fake voices,” ACM Multimedia Systems Con- ference, 2024

    2024

  14. [22]

    Detecting ai- generated voice using multi-scale time-frequency fea- tures,

    Y . Wang, M. Liu, Z. Wu, and Y . Qian, “Detecting ai- generated voice using multi-scale time-frequency fea- tures,” in ICASSP 2023 , pp. 1–5, 2023

    2023

  15. [23]

    V oiceprint: A two-branch cnn architecture for distinguishing between human and ai-generated speech,

    W. Chen, Y . Li, and J. Zhang, “V oiceprint: A two-branch cnn architecture for distinguishing between human and ai-generated speech,” IEEE Transactions on Information F orensics and Security, vol. 18, pp. 2514–2529, 2023

    2023

  16. [24]

    Real-time detection of ai-cloned voices using wavelet scattering transform,

    J. Monteiro, J. Alam, and P. Kenny, “Real-time detection of ai-cloned voices using wavelet scattering transform,” in Interspeech 2023, pp. 4210–4214, 2023

    2023

  17. [25]

    A survey on detection of ai-synthesized speech: Recent advances and challenges,

    H. Liu, X. Wang, L. Su, et al., “A survey on detection of ai-synthesized speech: Recent advances and challenges,” IEEE Signal Processing Magazine, vol. 40, no. 6, pp. 32– 49, 2023. 9

    2023

  18. [26]

    Factors affecting the frequency discrimination of complex tones and their implications for models of auditory pattern perception,

    B. C. Moore and D. A. Vickers, “Factors affecting the frequency discrimination of complex tones and their implications for models of auditory pattern perception,” Journal of the Acoustical Society of America , vol. 114, no. 5, pp. 2995–3007, 2003

    2003

  19. [27]

    Iso/iec mpeg-2 ad- vanced audio coding,

    M. Bosi, K. Brandenburg, S. Quackenbush, L. Fielder, K. Akagiri, H. Fuchs, and M. Dietz, “Iso/iec mpeg-2 ad- vanced audio coding,” Journal of the Audio engineering society, vol. 45, no. 10, pp. 789–814, 1997

    1997

  20. [28]

    Perceptual coding of digital audio,

    T. Painter and A. Spanias, “Perceptual coding of digital audio,” Proceedings of the IEEE , vol. 88, no. 4, pp. 451– 515, 2000

    2000

  21. [29]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit

    C. Veaux, J. Yamagishi, K. MacDonald, and others, “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit.”

  22. [30]

    Neural text normalization with subword units,

    C. Mansfield, M. Sun, Y . Liu, A. Gandhe, and B. Hoffmeister, “Neural text normalization with subword units,” Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, vol. 2, pp. 190–196, 2019

    2019

  23. [31]

    Tacotron: Towards end-to-end speech synthesis,

    Y . Wang, R. Skerry-Ryan, D. Stanton, Y . Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y . Xiao, Z. Chen, and S. B. et al., “Tacotron: Towards end-to-end speech synthesis,” arXiv preprint arXiv:1703.10135, 2017

    2017 arXiv

  24. [32]

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, and R. S.-R. et al., “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,” ICASSP 2018 IEEE In- ternational Conference on Acoustics, Speech and Signal Proces...

    2018

  25. [33]

    Deep voice 3: 2000- speaker neural text-to-speech,

    W. Ping, K. Peng, A. Gibiansky, S. O. Arik, A. Kannan, S. Narang, J. Raiman, and J. Miller, “Deep voice 3: 2000- speaker neural text-to-speech,” Proceedings of ICLR , pp. 214–217, 2018

    2000

  26. [34]

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

    Y . Ren, Y . Ruan, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.- Y . Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” arXiv preprint arXiv:2006.04558 , 2020

    2006 arXiv

  27. [35]

    Signal estimation from modi- fied short-time fourier transform,

    D. Griffin and J. Lim, “Signal estimation from modi- fied short-time fourier transform,” IEEE Transactions on Acoustics, Speech, and Signal Processing , vol. 32, no. 2, pp. 236–243, 1984

    1984

  28. [36]

    Parallel wave- gan: A fast waveform generation model based on gener- ative adversarial networks with multi-resolution spectro- gram,

    R. Yamamoto, E. Song, and J.-M. Kim, “Parallel wave- gan: A fast waveform generation model based on gener- ative adversarial networks with multi-resolution spectro- gram,” ICASSP 2020 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 6199–6203, 2020

    2020

  29. [37]

    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. 17–033, 2020

    2020

  30. [38]

    Conditional variational autoencoder with adversarial learning for end-to-end text- to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text- to-speech,” International Conference on Machine Learn- ing, pp. 5530–5540, 2021

    2021

  31. [39]

    Zwicker and H

    E. Zwicker and H. Fastl, Psychoacoustics: Facts and models. Springer, 1990

    1990

  32. [40]

    Bosi and R

    M. Bosi and R. E. Goldberg, Introduction to digital audio coding and standards . Springer Science & Business Media, 2002

    2002

  33. [41]

    The ISO/MPEG-1 audio coding stan- dard,

    K. Brandenburg, “The ISO/MPEG-1 audio coding stan- dard,” Journal of the Audio Engineering Society , vol. 47, no. 10, pp. 780–792, 1999

    1999

  34. [42]

    Transform coding of audio signals using perceptual noise criteria,

    J. D. Johnston, “Transform coding of audio signals using perceptual noise criteria,” IEEE Journal on selected areas in communications , vol. 6, no. 2, pp. 314–323, 1988

    1988

  35. [43]

    Defend- ing your voice: Adversarial attack on voice conversion

    C.-y. Huang, Y . Y . Lin, H.-y. Lee, and L.-s. Lee, “Defend- ing your voice: Adversarial attack on voice conversion.”

  36. [44]

    Active defense against voice conversion through generative adversar- ial network,

    S. Dong, B. Chen, K. Ma, and G. Zhao, “Active defense against voice conversion through generative adversar- ial network,” IEEE Signal Processing Letters , vol. 31, pp. 706–710, 2024

    2024

  37. [45]

    Subdivision of the audible frequency range into critical bands,

    E. Zwicker, “Subdivision of the audible frequency range into critical bands,” The Journal of the Acoustical Society of America , vol. 33, no. 2, pp. 248–248, 1961

    1961

  38. [46]

    A dwt-based rational dither modulation scheme for effective blind audio watermark- ing,

    H.-T. Hu and L.-Y . Hsu, “A dwt-based rational dither modulation scheme for effective blind audio watermark- ing,” Circuits, Systems, and Signal Processing , vol. 35, no. 2, pp. 553–572, 2016

    2016

  39. [47]

    Vctk: A speech dataset for voice conversion,

    T. Yamashita and T. Sgishi, “Vctk: A speech dataset for voice conversion,” in Proceedings of the 2019 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2345–2349, 2019

    2019

  40. [48]

    Elevenlabs text-to-speech api

    ElevenLabs, “Elevenlabs text-to-speech api.” https://elev enlabs.io, 2023

    2023

  41. [49]

    GPT-SoVITS: Zero-shot voice style transfer through decoupled style delivery,

    R. Team, “GPT-SoVITS: Zero-shot voice style transfer through decoupled style delivery,” arXiv preprint, 2024

    2024

  42. [50]

    XTTS v2: Zero-shot text-to-speech with multi- voice training

    Coqui.ai, “XTTS v2: Zero-shot text-to-speech with multi- voice training.” https://github.com/coqui-ai/TTS, 2023

    2023

  43. [51]

    Seed-vc: Self-supervised learning for voice conversion,

    Y . Wu, T. Qiu, J. Chen, J. Zhou, S. Chen, Z. Wu, and H. Meng, “Seed-vc: Self-supervised learning for voice conversion,” arXiv preprint arXiv:2304.04370 , 2023

    2023 arXiv

  44. [52]

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

    Y . Liu, K. Chen, X. Liu, X. Wu, Q. Fu, M. Song, Z. Dai, D. Bian, M. Johnson, L. Xue, et al., “Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,” arXiv preprint arXiv:2306.07691 , 2023

    2023 arXiv

  45. [53]

    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,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 5329–5333, 2018

    2018

  46. [54]

    Frcrn: Boost- ing feature representation using frequency recurrence for monaural speech enhancement,

    Y . Li, B. Xu, Y . Zhang, and J. Li, “Frcrn: Boost- ing feature representation using frequency recurrence for monaural speech enhancement,” in Proceedings of the 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 741–745, 2020

    2020

  47. [55]

    Pyinstrument: Call stack profiler for python

    R. Lechte, “Pyinstrument: Call stack profiler for python.” https://github.com/joerick/pyinstrument, 2023

    2023

Pith tools

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