Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Collecting, Curating, and Annotating Good Quality Speech deepfake dataset for Famous Figures: Process and Challenges

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read For ten public figures, this paper builds a voice-deepfake dataset whose synthetic samples fool human listeners 61.9% of the time, the highest rate among the datasets it compares.

desk verdict A genuinely useful dataset for targeted voice deepfakes, but the headline quality claims are statistically unsupported and the reported numbers don't reconcile. read the letter →

arxiv 2507.00324 v1 pith:JGKVZ6SQ submitted 2025-06-30 eess.AS

Hashim Ali , Surya Subramani , Raksha Varahamurthy , Nithin Adupa , Lekha Bollinani , Hafiz Malik This is my paper
classification eess.AS
keywords speechdeepfakedatasettext-to-speechsynthesisvoicespoofingdetectionpublicfigurestranscription-basedsegmentationNISQA-TTSnaturalnesshumanmisclassificationratezero-shotTTS
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 aims to show that a reproducible pipeline can produce a large dataset of real and synthetic speech for ten high-profile political figures, and that synthetic samples in this dataset are harder for people to distinguish from genuine recordings than samples in existing anti-spoofing datasets. It reports a naturalness score of 3.69 from the NISQA-TTS model, a deep-learning scorer for synthetic speech, and a human misclassification rate of 61.9 percent, both higher than the comparison datasets. The key ingredient, according to the authors, is transcription-based sentence segmentation: audio is cut into coherent, naturally timed segments rather than at fixed intervals or silence boundaries, which makes the synthesized speech sound less artifact-laden. The paper also documents how synthesis quality improved as the authors moved from training single-speaker models to few-shot fine-tuning and zero-shot inference with large pre-trained text-to-speech models. If the reported numbers are right, the dataset offers a more demanding evaluation ground for systems that try to protect public figures from voice impersonation.

What carries the argument

The load-bearing mechanism is the transcription-based segmentation step in the automated collection pipeline. Word-level transcripts with timestamps are grouped into sentence segments: words accumulate until a segment nears the target duration, the next punctuation mark within a window closes the segment, fragments too short to keep are discarded, and a small silence is appended when no punctuation appears. This produces coherent, naturally timed clips with an average duration of eight seconds, which the paper identifies as the main reason synthesized speech sounds better than from datasets that cut at fixed intervals or at 500 millisecond silences. On the synthesis side, the progression from speaker-specific training on eleven to thirteen thousand samples per speaker to few-shot fine-tuning and zero-shot inference with pre-trained multi-speaker models is what brings prosody and expressiveness close to real speech.

What would settle it

Run the same web listening test with at least 300 participants and ten or more samples per dataset per participant, compute confidence intervals for each dataset's misclassification rate, and check whether the Famous Figures rate still sits above the rates for SpoofCeleb and In-The-Wild; if it overlaps with or falls below them, the paper's strongest quality claim is not supported. A secondary check is to re-score the dataset's synthetic samples with the NISQA-TTS model and confirm the 3.69 naturalness value and its ranking against the other datasets.

Watch

Extended reading notes

Core claim

The central discovery is that a carefully engineered collection pipeline, combined with modern zero-shot and few-shot TTS models, can generate deepfake speech for targeted public figures that is more natural-sounding and more deceptive to human listeners than the speech in established identity-specific datasets. The authors attribute the improvement to segmenting real audio by sentence boundaries using word-level transcripts with timestamps, yielding segments with an average duration of eight seconds that preserve prosody and avoid the extended silences and abrupt cuts seen in other datasets. The resulting dataset contains 26,500 bonafide and 265,000 synthetic utterances from ten figures in political speeches, interviews, and public statements. In a listening test with 32 participants, listeners misclassified synthetic samples from this dataset as real speech 61.9 percent of the time, and the NISQA-TTS model scored the synthetic speech at 3.69, the highest naturalness among the datasets compared.

Load-bearing premise

The claim that this dataset has the highest human misclassification rate rests on a listening test with only 32 participants and two samples per dataset per participant, so the 61.9 percent figure has a wide confidence interval and the ordering against other datasets could change with a larger or differently selected sample.

Editorial extensions

If this is right

  • If the naturalness and misclassification figures hold, this dataset gives deepfake detector developers a more challenging real-world benchmark than existing identity-specific datasets such as SpoofCeleb or In-The-Wild.
  • A detector trained on this dataset would need to handle varied speaking contexts and acoustic conditions of political speech rather than clean read speech, which is closer to how impersonation attacks actually appear.
  • The transcription-based segmentation recipe can be applied to other speakers or languages without changing the underlying TTS model, making it a reusable component for future dataset construction.
  • Because the pipeline is automated and the selection criteria are explicit, the dataset can be extended to new public figures as fresh source material becomes available.

Reading between the lines

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

  • If the 61.9 percent misclassification rate is replicated with a larger listener panel, it would suggest that humans can no longer reliably tell zero-shot synthetic voices of famous figures apart from authentic recordings, a concrete risk for phone-based voter outreach and public messaging.
  • The reported listening test used only 32 participants and two samples per dataset per participant, so a natural extension is to run the same protocol with hundreds of listeners to see whether the ranking across datasets remains stable.
  • One evaluation the paper does not report is whether a detector trained on this dataset generalizes to a held-out public figure synthesized by a completely different TTS model; that test would show whether the dataset captures spoofing artifacts broadly enough for real-world deployment.
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. This manuscript describes a dataset of bonafide and synthetic speech for ten US political figures, built with an automated pipeline from YouTube videos using yt-dlp, AssemblyAI diarization, Whisper transcription, and a transcription-based segmentation method (§3). Synthetic speech is produced with speaker-specific StyleTTS2, few-shot/fine-tuned XTTSv2, and multiple zero-shot models (§4). The authors claim superior naturalness, reporting a NISQA-TTS score of 3.69 and a human fake-miss rate of 61.9%, the highest among the compared datasets (§5, Table 2). The paper also documents challenges involving SNR, audiobook prosody, and segmentation.

Significance. The dataset itself would fill a real gap: identity-specific, real-world recordings of high-profile political voices with a broad set of modern TTS attacks, which is directly relevant to targeted voice-spoofing detection, election-related misinformation, and public-figure protection. The manuscript is honest about the iterative nature of the pipeline and covers a wide range of synthesis systems. Its main quantitative evaluation, however, is not statistically grounded, so the 'superior quality' and 'highest misclassification rate' claims are not yet supported. The paper does not ship code or machine-checked proofs; its strength is the documented automated pipeline and the breadth of synthesis approaches.

major comments (4)
  1. [§5, Table 2] The claim that Famous Figures has the highest human fake-misclassification rate is not statistically supported. With 32 participants and one fake clip per participant per dataset, the 61.9% value corresponds to about 20 of 32 binary judgments, giving a 95% Wilson confidence interval of roughly [0.45, 0.77]; the 57.5% for CodecFake yields an interval of roughly [0.40, 0.73]. These intervals overlap, and no significance test is reported, so the observed ordering across datasets may be chance. The paper should report per-participant uncertainty, confidence intervals, and a statistical comparison, and should soften the abstract's 'highest' claim until then.
  2. [Table 1, §4.1.1, Figure 3] The dataset statistics are internally inconsistent. Table 1 reports 26,500 bonafide utterances and ~590 hours for 10 speakers, whereas §4.1.1 states that single-speaker training uses 11,000–13,000 samples per speaker (which would make 110,000–130,000 total), and Figure 3 says '12000 Speeches' per speaker. With the stated average segment length of about 8 s, 26,500 utterances would be roughly 59 hours, not 590 hours. The authors need to reconcile these counts and durations before the dataset description can be trusted.
  3. [§5, Table 2] The NISQA-TTS naturalness score of 3.69 is reported as a single mean with no sample count, variance, or inter-file distribution. A difference of 0.16 over the next-closest score (MLAAD, 3.53) cannot be evaluated without these quantities or a proper comparison test; the 'superior quality' statement in the abstract should be supported by error bars or replaced by a descriptive comparison.
  4. [§3.2, §4.2.2, Abstract] The abstract states that transcription-based segmentation 'significantly improves' synthetic speech quality, but no controlled experiment is reported. The paper describes the evolution from fixed-length to silence-based to transcription-based segmentation and asserts improvements (§4.2.2), yet there is no ablation in which the same TTS method is run with different segmentation strategies and the outputs are compared by the same metric. This should either be added or the claim softened to an anecdotal observation.
minor comments (5)
  1. [§3.2] The segmentation parameters D, T, and U are named but their numerical values are never given; without them the pipeline cannot be reproduced.
  2. [§5] Table 2's 'Fake Miss-Rates' is not defined, and no explanation is given for the missing SpoofCeleb entry; define the metric and report standard errors.
  3. [§5] The listening-test protocol should specify how participants were recruited, whether they were allowed to replay clips, and how the real/fake clip pairs were selected (for example, whether the fake clip used the same transcript as the real clip).
  4. [§4.1] The manuscript does not state which synthesis model produced which synthetic files in the released dataset, nor the per-speaker train/test split; this information is needed for reproducible use of the dataset.
  5. [General] The paper lacks an ethics/dual-use statement; since the dataset contains deepfakes of living political figures, a discussion of intended use, access restrictions, and consent/permissions for the original videos would be expected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: dataset quality claims rest on external NISQA predictions and a human listening test, not on fitted inputs.

full rationale

The paper does not derive predictions from fitted parameters. Its central claims are that the curated bonafide samples are clean, that the synthetic samples score 3.69 on NISQA-TTS, and that human listeners misclassified Famous Figures fakes 61.9% of the time. The NISQA-TTS score is produced by an externally trained model applied to the generated audio; the paper does not fit the model to its own data or define naturalness in terms of its own samples. The misclassification rate is a raw measurement from a web-based listening test; while the test is statistically weak (32 participants, two samples per dataset per participant, no error bars), statistical underpowering is a validity concern, not circularity. The only self-citation is reference [15], used to describe ASVSpoofLD (a laundering-attack dataset); that citation is descriptive and does not support the dataset-quality claims. No equation in the paper equates an output to an input, no parameter is fitted and then relabeled a prediction, and no uniqueness claim is imported from prior author work. The transcription-based segmentation is a preprocessing method, not a renamed result, and the comparison to SpoofCeleb and In-the-Wild is an external benchmark comparison rather than a self-definition. Therefore no circular step can be exhibited, and the honest finding is no significant circularity.

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

The central claims of dataset quality rest on the assumption that the chosen public figures, the YouTube source material, and the automated transcription are representative and accurate. The segmentation algorithm introduces several hand-chosen parameters whose values are not disclosed, and the NISQA-TTS and listening-test metrics are treated as valid without independent validation.

free parameters (4)
  • Target segment duration D = Not specified
    Used in the transcription-based segmentation algorithm to determine when to search for punctuation and cut segments; value is user-defined but not reported.
  • Segmentation threshold T = Not specified
    Tolerance around D that controls segment extension and the decision to discard trailing words; value is not reported.
  • Appended silence duration = 0.25 s
    Silence appended when no punctuation is found in the [D-T, D+1] window; a fixed value chosen by the authors.
  • Quality-control SNR threshold = Not specified
    The pipeline filters segments by SNR range, but the threshold value is not given.
assumptions (3)
  • domain assumption Publicly available YouTube videos of the selected figures are representative of real-world speech used in spoofing attacks.
    The design assumes that YouTube speeches and interviews capture the speaking contexts most relevant for targeted impersonation.
  • domain assumption NISQA-TTS naturalness scores are a valid proxy for perceptual quality and can be compared across datasets.
    The paper uses NISQA-TTS as the main objective quality metric and ranks datasets by it.
  • domain assumption Whisper Large Turbo word-level timestamps are accurate enough to produce semantically coherent sentence segments.
    The segmentation pipeline depends on Whisper's timestamps; systematic errors would create poorly aligned segments and degrade synthetic quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Collecting, Curating, and Annotating Good Quality Speech deepfake dataset for Famous Figures: Process and Challenges." pith.science (2026). https://pith.science/paper/JGKVZ6SQ

@misc{pith2026250700324,
  author       = {Pith},
  title        = {Pith review of: Collecting, Curating, and Annotating Good Quality Speech deepfake dataset for Famous Figures: Process and Challenges},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGKVZ6SQ}},
  note         = {Machine review of arXiv:2507.00324}
}
read the original abstract

Recent advances in speech synthesis have introduced unprecedented challenges in maintaining voice authenticity, particularly concerning public figures who are frequent targets of impersonation attacks. This paper presents a comprehensive methodology for collecting, curating, and generating synthetic speech data for political figures and a detailed analysis of challenges encountered. We introduce a systematic approach incorporating an automated pipeline for collecting high-quality bonafide speech samples, featuring transcription-based segmentation that significantly improves synthetic speech quality. We experimented with various synthesis approaches; from single-speaker to zero-shot synthesis, and documented the evolution of our methodology. The resulting dataset comprises bonafide and synthetic speech samples from ten public figures, demonstrating superior quality with a NISQA-TTS naturalness score of 3.69 and the highest human misclassification rate of 61.9\%.

Figures

Figures reproduced from arXiv: 2507.00324 by the authors.

Figure 1
Figure 1. Speech clips from Spoofceleb (Top Left) and In the Wild (Top Right) with extended silence (Red tint) , less duration and abrupt cut at the end (purple tint) . The bottom two plots are from Famous figures dataset with an average duration of 8 seconds and Transcription based segments developing generic audio spoof detection systems. The second category, Identity-Specific Datasets, addresses the challenges of protectin… view at source ↗
Figure 3
Figure 3. Schematic diagram of speech production. abrupt cutting at fixed intervals of n seconds or minutes. For example, a 30-minute utterance would be divided into 300 segments of 6 seconds each. Through continuous experi￾mentation, we transitioned to silence- or speaker-pause-based segmentation, as described in SpoofCeleb [20]. In this ap￾proach, if silence extends beyond 500 ms, the utterance is segmented at that point. W… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 33 canonical work pages

  1. [1]

    This high quality of synthesized speech and the ability to distribute it through so- cial media platforms are giving rise to manipulated information in the digital ecosystem

    Introduction The last few years have seen an exceptional increase in the re- alism of synthesized speech [1, 2, 3, 4, 5]. This high quality of synthesized speech and the ability to distribute it through so- cial media platforms are giving rise to manipulated information in the digital ecosystem. According to a Global Risk Report by the World Economic Foru...

    work page 2022

  2. [2]

    We can broadly clas- sify these datasets into two categories, based on their speaker characteristics and intended applications

    Existing Audio Anti-Spoofing Datasets The research community has developed various datasets to ad- vance the field of Audio Spoof Detection. We can broadly clas- sify these datasets into two categories, based on their speaker characteristics and intended applications. The first category, General Purpose Speaker Datasets, comprises audio data of anonymous ...

    arXiv 2025

  3. [3]

    Dataset Design and Methodology Our dataset design process is guided by three primary objec- tives: (1) ensuring high-quality authentic speech samples across various speaking contexts, (2) maintaining speaker diversity while capturing sufficient data per individual to represent their unique vocal characteristics, and (3) establishing a reproducible pipelin...

    work page 2018

  4. [4]

    Audio Segmentation: We utilize transcription-based seg- mentation that predicts word-level transcripts along with their timestamps in the utterance, as illustrated in Figure 2. This step processes word-level transcriptions with timestamps and groups them into sentence segments based on the utterance duration U, user-defined duration D, and the threshold d...

    work page

  5. [5]

    It uses ffmpeg to download the best audio available in the W A V format, and resample it 16 kHz

    Audio Extraction: We utilize yt-dlp 2 to directly download audio from YouTube links. It uses ffmpeg to download the best audio available in the W A V format, and resample it 16 kHz. Along with the YouTube video link, we specify the target speaker’s speech starting time to trim the audio to begin from the specified timestamp, ensuring that the first speake...

    work page

  6. [6]

    This step eliminates cross-talk and background speakers

    Speaker Diarization: We employ Assembly AI 3 to isolate segments that contain only the target speaker’s voice. This step eliminates cross-talk and background speakers

    work page

  7. [7]

    We also experimented with the Google speech recognition api package and other commer- cial tools; however, they generated text with less accuracy and without proper punctuation

    Transcription Generation: We integrate OpenAI Whisper Large Turbo4 for transcription, which gave word-level tran- scripts with timestamps. We also experimented with the Google speech recognition api package and other commer- cial tools; however, they generated text with less accuracy and without proper punctuation

    work page

  8. [8]

    The Biden Deepfake Robocall Is Only the Beginning,

    V . Elliott, “The Biden Deepfake Robocall Is Only the Beginning,” Wired, Jan. 2024, section: tags. [Online]. Available: https://www.wired.com/story/biden-robocall-deepfake-danger/

    work page 2024

Show all 51 references
  1. [9]

    Quality Control: Our initial segmentation strategy involved 2https://github.com/yt-dlp/yt-dlp 3https://www.assemblyai.com/ 4https://github.com/openai/whisper/ Famous Speaker - 1 ( 12000 Speeches * 10 seconds ) Clip - 1 Clip - 2 Clip - 3 . . . . . Clip - 11,999 Clip - 12,000 Si...

  2. [10]

    Al- though early models produced robotic-sounding speech despite extensive training data, SSL-based methods significantly im- proved quality through large-scale pre-training

    Synthetic Speech Generation The evolution of synthetic speech has progressed from su- pervised neural models to self-supervised learning (SSL). Al- though early models produced robotic-sounding speech despite extensive training data, SSL-based methods significantly im- proved ...

  3. [11]

    We train StyleTTS2 [22] only using this approach

    Training is performed using high-performance GPUs, such as three NVIDIA A100 GPUs, and typically takes five days per speaker. We train StyleTTS2 [22] only using this approach. The synthetic speech generated through this approach exhibited limitations in prosody and naturalness...

  4. [12]

    Audiobook Data Approach: We first attempted using high- quality audiobook data (5-10 hours per speaker) from Ama- zon Audible. While this improved signal quality and reduced computational artifacts, particularly with HiFi-GAN vocoder, the synthetic speech exhibited notable mon...

  5. [13]

    This approach preserved linguistic coherence and improved phoneme alignment, leading to reduced noise and more ac- curate spectrogram generation

    Enhanced Segmentation: We implemented transcription- based sentence-level segmentation using Whisper Large Turbo’s word-level timestamps as described in Section 3.2. This approach preserved linguistic coherence and improved phoneme alignment, leading to reduced noise and more ...

  6. [14]

    Advanced Model Architecture: Finally, we transitioned to Few-Shot and Zero-Shot TTS models, which demonstrated superior performance compared to single-speaker training approaches, effectively addressing our remaining challenges. This iterative progression from traditional appr...

  7. [15]

    For subjective evaluation, we implemented a web-based listening test with 32 unique participants

    Dataset Statistics and Analysis To evaluate the perceptual quality and detection difficulty of synthetic speech in datasets, we conducted subjective and ob- jective assessments. For subjective evaluation, we implemented a web-based listening test with 32 unique participants. E...

  8. [16]

    Both datasets are derived from the VCTK corpus, which comprises high-quality speech recordings collected in a controlled laboratory environment

    and CODECFake dataset [17] represent contemporary datasets specifically designed to evaluate detection systems against recent neural TTS architectures and codec-based neural speech synthesis methods. Both datasets are derived from the VCTK corpus, which comprises high-quality ...

  9. [17]

    A survey on neural speech synthesis,

    X. Tan, T. Qin, F. Soong, and T.-Y . Liu, “A survey on neural speech synthesis,” arXiv preprint arXiv:2106.15561, 2021

    2021 arXiv

  10. [18]

    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,” arXiv preprint arXiv:2006.04558, 2020

    2006 arXiv

  11. [19]

    Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,” in International Con- ference on Machine Learning. PMLR, 2022, pp. 2709–2720

    2022

  12. [20]

    Neural codec language models are zero-shot text to speech synthesizers, 2023,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Liet al., “Neural codec language models are zero-shot text to speech synthesizers, 2023,” URL: https://arxiv. org/abs/2301.02111. doi: doi, vol. 10

    2023 arXiv

  13. [21]

    Zse-vits: A zero-shot expressive voice cloning method based on vits,

    J. Li and L. Zhang, “Zse-vits: A zero-shot expressive voice cloning method based on vits,” Electronics, vol. 12, no. 4, p. 820, 2023

    2023

  14. [22]

    Global Risks Report 2024

    M. McLennan et al. , “Global Risks Report 2024.” World Economic Forum, 2024. [Online]. Available: https://www. weforum.org/publications/global-risks-report-2024/digest/

    2024

  15. [23]

    Russian War Report: Hacked news program and deep- fake video spread false Zelenskyy claims,

    ebaker, “Russian War Report: Hacked news program and deep- fake video spread false Zelenskyy claims,” Mar. 2022

    2022

  16. [24]

    Sadiq Khan says fake AI audio of him nearly led to serious disorder,

    M. Spring, “Sadiq Khan says fake AI audio of him nearly led to serious disorder,” Feb. 2024. [Online]. Available: https://www.bbc.com/news/uk-68146053

    2024

  17. [25]

    Asvspoof 2019: A large-scale public database of synthe- sized, converted and replayed speech,

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . Vestman, T. Kinnunen, K. A. Lee et al., “Asvspoof 2019: A large-scale public database of synthe- sized, converted and replayed speech,” Computer Speech & Lan- guage, vol. 64, p. 101114, 2020

    2019

  18. [26]

    ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild,

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kinnunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautsch, and K. A. Lee, “ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol....

    2021

  19. [27]

    Asvspoof 2021: accelerating progress in spoofed and deep- fake speech detection,

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, X. Liu, K. A. Lee, T. Kinnunen, N. Evans et al. , “Asvspoof 2021: accelerating progress in spoofed and deep- fake speech detection,” in ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoofing C...

    2021

  20. [28]

    Asvspoof 5: crowd- sourced speech data, deepfakes, and adversarial attacks at scale,

    X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. H. Kinnunen, N. Evans, K. A. Lee, and J. Yamagishi, “Asvspoof 5: crowd- sourced speech data, deepfakes, and adversarial attacks at scale,” in The Automatic Speaker Verific...

    2024

  21. [29]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” arXiv preprint arXiv:2012.03411, 2020

    2012 arXiv

  22. [30]

    Is audio spoof detection robust to laundering attacks?

    H. Ali, S. Subramani, S. Sudhir, R. Varahamurthy, and H. Malik, “Is audio spoof detection robust to laundering attacks?” in Pro- ceedings of the 2024 ACM Workshop on Information Hiding and Multimedia Security, 2024, pp. 283–288

    2024

  23. [31]

    Dfadd: The diffusion and flow- matching based audio deepfake dataset,

    J. Du, I.-M. Lin, I.-H. Chiu, X. Chen, H. Wu, W. Ren, Y . Tsao, H.-y. Lee, and J.-S. R. Jang, “Dfadd: The diffusion and flow- matching based audio deepfake dataset,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 921– 928

    2024

  24. [32]

    The codecfake dataset and counter- measures for the universally detection of deepfake audio,

    Y . Xie, Y . Lu, R. Fu, Z. Wen, Z. Wang, J. Tao, X. Qi, X. Wang, Y . Liu, H. Cheng et al. , “The codecfake dataset and counter- measures for the universally detection of deepfake audio,” arXiv preprint arXiv:2405.04880, 2024

    2024 arXiv

  25. [33]

    Mlaad: The multi- language audio anti-spoofing dataset,

    N. M. M ¨uller, P. Kawa, W. H. Choong, E. Casanova, E. G ¨olge, T. M¨uller, P. Syga, P. Sperl, and K. B¨ottinger, “Mlaad: The multi- language audio anti-spoofing dataset,”International Joint Confer- ence on Neural Networks (IJCNN), 2024

    2024

  26. [34]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?” In- terspeech, 2022

    2022

  27. [35]

    Spoofceleb: Speech deepfake detection and sasv in the wild,

    J.-w. Jung, Y . Wu, X. Wang, J.-H. Kim, S. Maiti, Y . Matsunaga, H.-j. Shim, J. Tian, N. Evans, J. S. Chung et al. , “Spoofceleb: Speech deepfake detection and sasv in the wild,”IEEE Open Jour- nal of Signal Processing, 2025

    2025

  28. [36]

    V oxceleb: Large-scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxceleb: Large-scale speaker verification in the wild,”Computer Speech & Language, vol. 60, p. 101027, 2020

    2020

  29. [37]

    Styletts 2: Towards human-level text-to-speech through style dif- fusion and adversarial training with large speech language mod- els,

    Y . A. Li, C. Han, V . Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to-speech through style dif- fusion and adversarial training with large speech language mod- els,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  30. [38]

    Xtts: a massively multilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi et al. , “Xtts: a massively multilingual zero-shot text-to-speech model,” arXiv preprint arXiv:2406.04904, 2024

    2024 arXiv

  31. [39]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  32. [40]

    E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tan et al., “E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 682– 689

    2024

  33. [41]

    Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,

    S. Liao, Y . Wang, T. Li, Y . Cheng, R. Zhang, R. Zhou, and Y . Xing, “Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,” arXiv preprint arXiv:2411.01156, 2024

    2024 arXiv

  34. [42]

    Ssr-speech: Towards stable, safe and robust zero-shot text-based speech editing and synthesis,

    H. Wang, M. Yu, J. Hai, C. Chen, Y . Hu, R. Chen, N. De- hak, and D. Yu, “Ssr-speech: Towards stable, safe and robust zero-shot text-based speech editing and synthesis,”arXiv preprint arXiv:2409.07556, 2024

    2024 arXiv

  35. [43]

    Maskgct: Zero-shot text-to- speech with masked generative codec transformer,

    Y . Wang, H. Zhan, L. Liu, R. Zeng, H. Guo, J. Zheng, Q. Zhang, X. Zhang, S. Zhang, and Z. Wu, “Maskgct: Zero-shot text-to- speech with masked generative codec transformer,”arXiv preprint arXiv:2409.00750, 2024

    2024 arXiv

  36. [44]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wang, F. Yu, H. Liu, Z. Sheng, Y . Gu, C. Deng, W. Wang, S. Zhang, Z. Yan, and J. Zhou, “Cosyvoice 2: Scalable streaming speech synthesis with large language models,”

  37. [46]

    Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis,

    Z. Ye, X. Zhu, C.-M. Chan, X. Wang, X. Tan, J. Lei, Y . Peng, H. Liu, Y . Jin, Z. DAI, H. Lin, J. Chen, X. Du, L. Xue, Y . Chen, Z. Li, L. Xie, Q. Kong, Y . Guo, and W. Xue, “Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis,” 2025. [Online]...

    2025 arXiv

  38. [47]

    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, S. Bengio et al. , “Tacotron: Towards end-to-end speech synthesis,” arXiv preprint arXiv:1703.10135, 2017

    2017 arXiv

  39. [48]

    Glow-tts: A genera- tive flow for text-to-speech via monotonic alignment search,

    J. Kim, S. Kim, J. Kong, and S. Yoon, “Glow-tts: A genera- tive flow for text-to-speech via monotonic alignment search,”Ad- vances in Neural Information Processing Systems , vol. 33, pp. 8067–8077, 2020

    2020

  40. [49]

    Hifi- codec: Group-residual vector quantization for high fidelity audio codec,

    D. Yang, S. Liu, R. Huang, J. Tian, C. Weng, and Y . Zou, “Hifi- codec: Group-residual vector quantization for high fidelity audio codec,” arXiv preprint arXiv:2305.02765, 2023

    2023 arXiv

  41. [50]

    Univnet: A neural vocoder with multi-resolution spectrogram discrimi- nators for high-fidelity waveform generation,

    W. Jang, D. Lim, J. Yoon, B. Kim, and J. Kim, “Univnet: A neural vocoder with multi-resolution spectrogram discrimi- nators for high-fidelity waveform generation,” arXiv preprint arXiv:2106.07889, 2021

    2021 arXiv

  42. [51]

    Deep learning based assessment of synthetic speech naturalness,

    G. Mittag and S. M ¨oller, “Deep learning based assessment of synthetic speech naturalness,” arXiv preprint arXiv:2104.11673 , 2021

    2021 arXiv

  43. [2024]

    Available: https://arxiv.org/abs/2412.10117

    [Online]. Available: https://arxiv.org/abs/2412.10117

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.