Pith. sign in

REVIEW 4 major objections 6 minor 48 references

Phoneme-Level Analysis for Person-of-Interest Speech Deepfake Detection

T0 review · 4 major / 6 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 Per-phoneme voice profiles match whole-utterance deepfake checks

desk verdict New pipeline combining phoneme-level analysis with POI profiles; the robustness numbers are plausible but interpretation is muddied by the changing phoneme set under distortion. read the letter →

arxiv 2507.08626 v1 pith:5SHEZIUS submitted 2025-07-11 cs.SD eess.AS

Davide Salvi , Viola Negroni , Sara Mandelli , Paolo Bestagini , Stefano Tubaro This is my paper
classification cs.SDeess.AS
keywords speechdeepfakedetectionperson-of-interestphoneme-levelanalysisWav2Vec2.0interpretabilitymultimediaforensicsvoicespoofingone-classlearning
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 proposes a person-of-interest (POI) speech deepfake detector that works at the level of individual phonemes rather than whole utterances. It claims that by building a speaker profile from per-phoneme embeddings of clean reference audio, and comparing each phoneme of a test clip against that profile, the detector matches the accuracy of a standard utterance-level POI baseline while being substantially more robust to noise, compression, and quantization. The method also pinpoints exactly which phonemes deviate, turning a binary fake/real decision into an interpretable, localizable forensic signal. A sympathetic reader would care because this offers a path toward explainable, speaker-centric deepfake detection that can survive real-world post-processing.

What carries the argument

The machinery is a two-stage Wav2Vec 2.0 pipeline: a fine-tuned phoneme segmenter aligns each frame to an IPA phoneme, while a second Wav2Vec 2.0 instance produces 768-dimensional frame embeddings; embeddings of frames belonging to the same phoneme are averaged into one per-phoneme vector. A speaker profile is a dictionary mapping each phoneme to the set of such vectors from 100 clean reference utterances. At inference, each test phoneme vector is scored by its minimum cosine distance to the matching profile entry, and the track score is the mean of these per-phoneme minimum distances; a threshold on this score classifies the track as real or fake.

What would settle it

Run the proposed pipeline under conditions where the phoneme aligner is known to fail — for example, heavily compressed or adversarial synthetic audio that breaks the ASR alignment, or a non-English language outside the aligner's training — and compare detection EER against the baseline; if EER degrades to or beyond the baseline while the amount of processed signal remains small, the claimed phoneme-level robustness is an artifact of alignment accuracy rather than of phoneme-level modeling.

Watch

Extended reading notes

Core claim

The central discovery claimed is that decomposing a POI deepfake detector down to the phoneme level — averaging Wav2Vec 2.0 frame embeddings per phoneme, storing them in a speaker-specific dictionary, and scoring a test track by the mean minimum cosine distance per phoneme — yields detection accuracy comparable to a whole-utterance baseline on four datasets (ASVspoof 2019, In-the-Wild, Purdue, TIMIT-TTS), with better robustness to post-processing and with phoneme-level interpretability. The paper also finds that vowels and plosives are the most discriminative phoneme categories, that no single category beats the full phoneme set, and that the phoneme-level pipeline processes roughly 65% less of the input signal than the baseline.

Load-bearing premise

The pipeline assumes an external phoneme segmenter produces accurate alignments on both real and synthetic audio; if synthetic audio (or noise) misaligns phonemes, the averaged per-phoneme embeddings and the min-cosine-distance scores inherit the errors, which could destroy the claimed robustness and interpretability.

Editorial extensions

If this is right

  • Phoneme-level POI detection can serve as an interpretability layer for deepfake forensics, identifying the specific phonemes and time spans where a synthetic clip deviates from the protected speaker.
  • Because the method processes only phoneme-bearing regions, it cuts the analyzed signal by roughly two-thirds, enabling cheaper inference at comparable accuracy.
  • The discriminative power of vowels and plosives suggests deepfake generators can be audited by focusing on specific phoneme classes, yet no single class is sufficient on its own.
  • The method inherits one-class training on real data only, so it can generalize to unseen spoofing methods without retraining on synthetic examples.
  • Post-processing robustness to noise, MP3 compression, and quantization on test clips is a direct corollary; the paper shows EER increases below 5% for the phoneme method in most conditions while the baseline exceeds 10% on two datasets.

Reading between the lines

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

  • A testable extension is to use the per-phoneme distances as features for a small classifier or to perform phoneme-level calibration per speaker, which could recover the accuracy gap on short-utterance datasets like ASVspoof 2019.
  • The same phoneme-profile machinery could be turned into a localization system: instead of a binary score, the method could output a phoneme-level anomaly map that an investigator overlays on the transcript, and the paper's case study already gestures at this.
  • Because the authors find affricates least informative and vowels and plosives most, a future detector might weight phoneme classes by discriminability or by occurrence frequency, which the paper lists as future work but does not implement.
  • The dependency on an external aligner suggests that alignment uncertainty could be propagated into the score rather than assumed away; evaluating with alignment confidence weighting is a natural next step not explored in the paper.
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 / 6 minor

Summary. The paper proposes a Person-of-Interest (POI) speech deepfake detector that operates at the phoneme level. Reference and test tracks are segmented by a fine-tuned Wav2Vec2 phoneme extractor, each phoneme instance is embedded with a base Wav2Vec2 model, and the decision score is the average over test phonemes of the minimum cosine distance between each phoneme embedding and the corresponding entry in a speaker profile built from I=100 reference utterances (Eqs. 1--3). The method is compared against an utterance-level POI baseline that uses the same feature extractor and the minimum cosine distance between whole-utterance embeddings. Experiments on ASVspoof 2019, In-the-Wild, Purdue, and TIMIT-TTS report AUC/EER, EER variations under noise/MP3/quantization, per-phoneme-category EERs, and a two-speaker case study. The main findings are that the proposed method is substantially worse on ASVspoof 2019 (EER 25.82 vs 6.84), better on In-the-Wild and Purdue, and comparable on TIMIT-TTS, while appearing more robust to post-processing and providing per-phoneme distances for interpretation.

Significance. If the robustness and interpretability claims can be substantiated, the paper describes a useful and genuinely interpretable POI detector: it is training-free on synthetic speech, builds on two self-supervised pre-trained models, and outputs per-phoneme deviation scores that could support forensic explanations. The paper has clear strengths: the decision score in Eqs. (2)--(3) contains no fitted parameters and is not calibrated to the test labels; the evaluation spans four public datasets; and Section 5.5 candidly acknowledges the dependency on external phoneme segmentation and the restriction to English. The main weaknesses are that the evidence for the headline claims is mixed: the accuracy claim is contradicted by the ASVspoof 2019 results and lacks statistical support, the robustness comparison is confounded by segmentation instability, and the interpretability analysis is largely anecdotal. Because all three headline claims are presented as the paper's central contributions, these issues are load-bearing for the current version.

major comments (4)
  1. [Section 5.2, Eq. (3), Table 2] The robustness claim is confounded by the fact that the phoneme segmentation itself changes under post-processing. In Eq. (3), D(x, F_S) averages over the N phonemes returned by the external extractor P applied to the distorted test signal, so a clean-vs-distorted comparison changes not only the feature values but also the set of analyzed phonemes and N. The paper's own explanation of the negative ΔEER values in Section 5.2 -- that post-processing leaves only the clearest phonemes and filters out ambiguous content -- confirms that Table 2 compares different analyzed subsets, not a fixed phoneme-level decision rule. To support the claim of superior robustness, the authors should isolate the effect of segmentation instability, for example by fixing the phoneme alignment from the clean signal and recomputing distances on distorted features, or by reporting results on the matched set of phonemes that are detected in both clean and distorted conditions. As written, the robustness advantage may be a segmentation-driven sample-selection effect rather than an intrinsic property of the phoneme-level embeddings, which directly affects the central claim.
  2. [Section 5.1, Table 1] The claim of 'comparable accuracy' is not supported by the reported numbers. On ASVspoof 2019 the baseline EER is 6.84% while the proposed method is 25.82%, and the AUC drops from 98.12% to 79.45%; this is a large degradation, not a comparable outcome. The qualitative conclusion in the abstract and Section 5.1 is based on two wins, one tie, and one substantial loss, but no confidence intervals, error bars, per-speaker variability, or significance tests are reported, and the comparison uses only a single baseline. The authors should report per-speaker AUC/EER distributions with bootstrap or paired statistical tests, and ideally include at least one additional POI baseline from the literature, before claiming that accuracy is comparable overall.
  3. [Sections 5.3--5.4, Table 4, Figure 4] The interpretability claim is presented as a key contribution, but the evidence is mostly anecdotal. In Table 4, evaluating 'only one category at a time' changes both the phoneme type and the number of phonemes N in Eq. (3), so differences across rows may reflect phoneme frequency and segment count as much as discriminative power. No error bars or significance tests are given, so the ordering of categories (vowels and plosives as most discriminative, affricates as least) is not established. The case study in Figure 4 covers only two speakers and two utterances, and it does not compare the identified high-deviation phonemes against any baseline or chance level. To substantiate the interpretability claim, the authors should report a quantitative phoneme-level discrimination measure (e.g., per-phoneme or per-category AUC over many utterances) and, for the case study, a comparison of the phoneme-distance ranking in fake versus real utterances.
  4. [Sections 4.4 and 5.1] The evaluation protocol is under-specified. The paper does not state how many speakers are used per dataset, how reference and test utterances are split per speaker, how many test utterances remain after withholding I=100 reference tracks, and whether any real test clips share source material with the reference tracks (a concern for datasets such as In-the-Wild, where real clips and reference clips may come from the same public videos). Without this information, the reported AUC/EER values cannot be reproduced, and the minimum-cosine-distance score in Eq. (2) could be inflated by near-duplicate reference and test utterances. The authors should specify the exact split procedure and check for utterance-level overlap.
minor comments (6)
  1. [Table 1] There is a formatting error in the TIMIT-TTS row: '84.5421.33' should be '84.54 21.33'.
  2. [Section 4.2] The choice of I=100 is described as 'based on empirical validation'; please report the range of values explored and a sensitivity analysis, since I affects both phoneme coverage and the number of available test utterances.
  3. [Figure 4] The x-axis phoneme labels are difficult to read and appear garbled (e.g., 'w ni t c h e n ei b r s'); the figure should use proper IPA typesetting and a clear time-aligned axis.
  4. [Section 3.2, Eq. (2)] The phrase 'minimum element-wise cosine distance' is ambiguous; what is meant is the minimum cosine distance over the reference set, and this should be stated in words or by a clearer notation.
  5. [Abstract and Section 2] The phrase 'traditional approaches' overstates the comparison, since only one utterance-level baseline is evaluated; consider wording such as 'the considered utterance-level baseline'.
  6. [Section 4.1] Please clarify whether the frame-level feature vectors for a phoneme instance are averaged before computing the cosine distance in Eq. (2), and whether the same averaging is applied to the reference profile entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the phoneme-level decision score is a parameter-free cosine-distance aggregation built from external pretrained models, with no fitted quantity or self-citation chain reproducing the test labels.

full rationale

The central claim is that a POI detector comparing per-phoneme Wav2Vec2 embeddings against a reference speaker's phoneme profile reaches accuracy comparable to an utterance-level baseline while adding robustness and interpretability. The decision score in Eqs. (2)-(3) is D(x, F_S) = (1/N) sum_n d_n, where each d_n is the minimum cosine distance between a test phoneme embedding and the phoneme entry in the speaker profile. No parameter is fitted to the detection labels, no threshold is tuned, and no normalization is chosen to reproduce test outcomes. The only hand-set quantity is the profile size I = 100, justified by phoneme coverage and dataset-balance considerations, not by detection performance. The feature extractor F and phoneme extractor P are external pretrained Wav2Vec2 models, so the pipeline is self-contained with respect to external benchmarks. The scoring convention (minimum cosine distance) is attributed to Pianese et al. [30], but adopting a previously published scoring rule is a methodological choice, not a load-bearing self-citation: the paper's results would stand or fall on the computed distances regardless of who proposed the cosine-distance rule. The paper's limitation statement in Section 5.5 explicitly acknowledges dependence on accurate phoneme segmentation and propagation of alignment errors; this is an honest validity caveat about robustness to post-processing and not a circularity. The Table 2 robustness comparison may be complicated by the fact that post-processing can change which phonemes the external segmenter returns, but that is an experimental-design concern about comparing different analyzed phoneme sets, not a case where the conclusion is equivalent to the inputs by definition. I find no step in the derivation chain where a predicted quantity is constructed from, or defined in terms of, the quantity it is claimed to predict. Score 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new particles, forces, dimensions, or other entities are introduced; the method reuses pretrained models and existing distance scoring. The central claim rests on 100 reference utterances per speaker and on the transferability of Wav2Vec2 embeddings, not on any newly postulated mechanism.

free parameters (1)
  • Number of reference utterances I per speaker = 100
    Section 4.2 sets I=100 based on empirical validation; phoneme coverage and dataset balance depend on this choice, so the reported accuracy and robustness results are conditioned on it.
assumptions (5)
  • domain assumption Wav2Vec2 base embeddings capture speaker-specific phoneme characteristics and synthetic artifacts.
    The pipeline uses d=768 Wav2Vec2 embeddings without training, so the method's discriminative power is inherited from the pretrained feature extractor.
  • domain assumption The phoneme extractor and feature extractor outputs are time-aligned because they use the same frame-level processing pipeline.
    Section 4.1: 'the two models share the same architecture and processing pipeline, their outputs are inherently time-aligned'; this alignment is required to associate each phoneme with an embedding.
  • domain assumption Minimum cosine distance to the nearest reference phoneme embedding is a valid authenticity score.
    Eq. (2)-(3) define the decision metric directly; no validation is given that nearest-neighbor cosine distance separates real from fake phonemes better than other metrics.
  • domain assumption Deepfake generators are less faithful at the phoneme level than at the whole-utterance level.
    Section 1 and Section 5.5: the method is grounded on the idea that generators struggle to replicate short, fundamental units of speech; this hypothesis is the reason for the pipeline but is not independently tested.
  • domain assumption Discarding non-phoneme frames removes uninformative content rather than useful evidence.
    Section 5.1 attributes the ASVspoof performance drop to short analyzed content; this assumes silence and unclear frames carry no forensic value, which may not hold across datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phoneme-Level Analysis for Person-of-Interest Speech Deepfake Detection." pith.science (2026). https://pith.science/paper/5SHEZIUS

@misc{pith2026250708626,
  author       = {Pith},
  title        = {Pith review of: Phoneme-Level Analysis for Person-of-Interest Speech Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SHEZIUS}},
  note         = {Machine review of arXiv:2507.08626}
}
read the original abstract

Recent advances in generative AI have made the creation of speech deepfakes widely accessible, posing serious challenges to digital trust. To counter this, various speech deepfake detection strategies have been proposed, including Person-of-Interest (POI) approaches, which focus on identifying impersonations of specific individuals by modeling and analyzing their unique vocal traits. Despite their excellent performance, the existing methods offer limited granularity and lack interpretability. In this work, we propose a POI-based speech deepfake detection method that operates at the phoneme level. Our approach decomposes reference audio into phonemes to construct a detailed speaker profile. In inference, phonemes from a test sample are individually compared against this profile, enabling fine-grained detection of synthetic artifacts. The proposed method achieves comparable accuracy to traditional approaches while offering superior robustness and interpretability, key aspects in multimedia forensics. By focusing on phoneme analysis, this work explores a novel direction for explainable, speaker-centric deepfake detection.

Figures

Figures reproduced from arXiv: 2507.08626 by the authors.

Figure 1
Figure 1. Proposed phoneme-based Person-of-Interest (POI) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of the proposed phoneme-level POI speech [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of average input signal duration (in sec [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case study: Phoneme-distance comparison between real [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 47 canonical work pages

  1. [1]

    Protecting world leaders against deep fakes

    Shruti Agarwal, Hany Farid, Yuming Gu, Mingming He, Koki Nagano, and Hao Li. Protecting world leaders against deep fakes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 1, 2

    work page 2019

  2. [2]

    Detecting deep-fake videos from phoneme- viseme mismatches

    Shruti Agarwal, Hany Farid, Ohad Fried, and Maneesh Agrawala. Detecting deep-fake videos from phoneme- viseme mismatches. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 2

    work page 2020

  3. [3]

    Deepfake media forensics: Status and future chal- lenges

    Irene Amerini, Mauro Barni, Sebastiano Battiato, Paolo Bestagini, Giulia Boato, Vittoria Bruni, Roberto Caldelli, Francesco De Natale, Rocco De Nicola, Luca Guarnera, et al. Deepfake media forensics: Status and future chal- lenges. Journal of Imaging, 11(3):73, 2025. 2

    work page 2025

  4. [4]

    Combining Automatic Speaker Verification and Prosody Analysis for Synthetic Speech Detection

    Luigi Attorresi, Davide Salvi, Clara Borrelli, Paolo Bestagini, and Stefano Tubaro. Combining Automatic Speaker Verification and Prosody Analysis for Synthetic Speech Detection. In International Conference on Pattern Recognition (ICPR), 2022. 1

    work page 2022

  5. [5]

    wav2vec 2.0: A framework for self-supervised learning of speech representations

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural In- formation Processing Systems (NeurIPS), 2020. 4

    work page 2020

  6. [6]

    Transformer ensemble for syn- thesized speech detection

    Emily R Bartusiak, Kratika Bhagtani, Amit Kumar Singh Yadav, and Edward J Delp. Transformer ensemble for syn- thesized speech detection. In Asilomar Conference on Sig- nals, Systems, and Computers. IEEE, 2023. 1

    work page 2023

  7. [7]

    Are Recent Deepfake Speech Generators Detectable? In ACM Workshop on Information Hiding and Multimedia Security, 2024

    Kratika Bhagtani, Amit Kumar Singh Yadav, Paolo Bestagini, and Edward J Delp. Are Recent Deepfake Speech Generators Detectable? In ACM Workshop on Information Hiding and Multimedia Security, 2024. 5

    work page 2024

  8. [8]

    Ljspeech phonemes dataset

    Bookbot. Ljspeech phonemes dataset. https : / / huggingface . co / datasets / bookbot / ljspeech _ phonemes, 2023. Accessed: 2025-02-

    work page 2023

Show all 48 references
  1. [9]

    Listen, attend and spell: A neural network for large vocab- ulary conversational speech recognition

    William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. Listen, attend and spell: A neural network for large vocab- ulary conversational speech recognition. In IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2016. 3

    2016

  2. [10]

    Anomaly detection and local- ization for speech deepfakes via feature pyramid matching

    Emma Coletta, Davide Salvi, Viola Negroni, Daniele Ugo Leonzio, and Paolo Bestagini. Anomaly detection and local- ization for speech deepfakes via feature pyramid matching. arXiv preprint arXiv:2503.18032, 2025. 2

    2025 arXiv

  3. [11]

    Id-reveal: Identity-aware deepfake video detection

    Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021. 2

    2021

  4. [12]

    Audio-visual person-of-interest deep- fake detection

    Davide Cozzolino, Alessandro Pianese, Matthias Nießner, and Luisa Verdoliva. Audio-visual person-of-interest deep- fake detection. In IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2023. 1, 2

    2023

  5. [13]

    Open challenges in synthetic speech detection

    Luca Cuccovillo, Christoforos Papastergiopoulos, Anasta- sios Vafeiadis, Artem Yaroshchuk, Patrick Aichroth, Kon- stantinos V otis, and Dimitrios Tzovaras. Open challenges in synthetic speech detection. In IEEE International Workshop on Information Forensics and Security (WIFS...

    2022

  6. [14]

    Au- dio transformer for synthetic speech detection via multi- formant analysis

    Luca Cuccovillo, Milica Gerhardt, and Patrick Aichroth. Au- dio transformer for synthetic speech detection via multi- formant analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1

    2024

  7. [15]

    Using self attention DNNs to discover phonemic fea- tures for audio deep fake detection

    Hira Dhamyal, Ayesha Ali, Ihsan Ayyub Qazi, and Agha Ali Raza. Using self attention DNNs to discover phonemic fea- tures for audio deep fake detection. In IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) ,

  8. [16]

    Can spoofing countermeasure and speaker verification systems be jointly optimised? In IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2023

    Wanying Ge, Hemlata Tak, Massimiliano Todisco, and Nicholas Evans. Can spoofing countermeasure and speaker verification systems be jointly optimised? In IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2023. 2

    2023

  9. [17]

    Privacy and security concerns in generative ai: A compre- hensive survey

    Abenezer Golda, Kidus Mekonen, Amit Pandey, Anushka Singh, Vikas Hassija, Vinay Chamola, and Biplab Sikdar. Privacy and security concerns in generative ai: A compre- hensive survey. IEEE Access, 2024. 1

    2024

  10. [18]

    The LJ Speech Dataset

    Keith Ito and Linda Johnson. The LJ Speech Dataset. https://keithito.com/LJ-Speech-Dataset/ ,

  11. [19]

    Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks

    Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu, and Nicholas Evans. Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks. In IEEE International Conference on Acoustics, Speech and Signal Process...

    2022

  12. [20]

    A survey on speech deepfake detection

    Menglu Li, Yasaman Ahmadiadli, and Xiao-Ping Zhang. A survey on speech deepfake detection. ACM Computing Sur- veys, 2025. 2

    2025

  13. [21]

    Speaker-phonetic vector estima- tion for short duration speaker verification

    Jianbo Ma, Vidhyasaharan Sethu, Eliathamby Ambikaira- jah, and Kong Aik Lee. Speaker-phonetic vector estima- tion for short duration speaker verification. In IEEE Inter- national Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2018. 3

    2018

  14. [22]

    Deepfakes genera- tion and detection: State-of-the-art, open challenges, coun- termeasures, and way forward

    Momina Masood, Mariam Nawaz, Khalid Mahmood Malik, Ali Javed, Aun Irtaza, and Hafiz Malik. Deepfakes genera- tion and detection: State-of-the-art, open challenges, coun- termeasures, and way forward. Applied intelligence, 53(4): 3974–4026, 2023. 1

    2023

  15. [23]

    Speech is silver, silence is golden: What do ASVspoof- trained models really learn? In Interspeech, 2021

    Nicolas M Müller, Franziska Dieckmann, Pavel Czempin, Roman Canals, Konstantin Böttinger, and Jennifer Williams. Speech is silver, silence is golden: What do ASVspoof- trained models really learn? In Interspeech, 2021. 6

    2021

  16. [24]

    Does audio deep- fake detection generalize? In Interspeech, 2022

    Nicolas M Müller, Pavel Czempin, Franziska Dieckmann, Adam Froghyar, and Konstantin Böttinger. Does audio deep- fake detection generalize? In Interspeech, 2022. 5

    2022

  17. [25]

    Replay attack detection using dnn for channel discrim- ination

    Parav Nagarsheth, Elie Khoury, Kailash Patil, and Matt Gar- land. Replay attack detection using dnn for channel discrim- ination. In Interspeech, 2017. 3

    2017

  18. [26]

    Leveraging mixture of ex- perts for improved speech deepfake detection

    Viola Negroni, Davide Salvi, Alessandro Ilic Mezza, Paolo Bestagini, and Stefano Tubaro. Leveraging mixture of ex- perts for improved speech deepfake detection. In IEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2025. 1

    2025

  19. [27]

    Acoustic-to-word recog- nition with sequence-to-sequence models

    Shruti Palaskar and Florian Metze. Acoustic-to-word recog- nition with sequence-to-sequence models. In IEEE Spoken Language Technology Workshop (SLT), 2018. 3

    2018

  20. [28]

    Librispeech: an ASR corpus based on public domain audio books

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an ASR corpus based on public domain audio books. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015. 5

    2015

  21. [29]

    Deepfake audio detection by speaker ver- ification

    Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Deepfake audio detection by speaker ver- ification. In IEEE International Workshop on Information Forensics and Security (WIFS), 2022. 1, 2

    2022

  22. [30]

    Training-free deepfake voice recognition by leveraging large-scale pre-trained models

    Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Training-free deepfake voice recognition by leveraging large-scale pre-trained models. InACM Workshop on Information Hiding and Multimedia Security, 2024. 1, 2, 4, 5

    2024

  23. [31]

    Deepfake detection: A systematic lit- erature review

    Md Shohel Rana, Mohammad Nur Nobi, Beddhu Murali, and Andrew H Sung. Deepfake detection: A systematic lit- erature review. IEEE Access, 2024. 1

    2024

  24. [32]

    TIMIT-TTS: a Text-to-Speech Dataset for Multimodal Synthetic Media Detection

    Davide Salvi, Brian Hosler, Paolo Bestagini, Matthew C Stamm, and Stefano Tubaro. TIMIT-TTS: a Text-to-Speech Dataset for Multimodal Synthetic Media Detection. IEEE Access, 2023. 5

    2023

  25. [33]

    The VidTIMIT database

    Conrad Sanderson. The VidTIMIT database. Technical re- port, IDIAP, 2002. 5

    2002

  26. [34]

    Hierarchi- cal structures of neural networks for phoneme recognition

    Petr Schwarz, Pavel Matejka, and Jan Cernocky. Hierarchi- cal structures of neural networks for phoneme recognition. In IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP), 2006. 3

    2006

  27. [35]

    Phonetic and speaker variations in automatic emotion classification

    Vidhyasaharan Sethu, Eliathamby Ambikairajah, and Julien Epps. Phonetic and speaker variations in automatic emotion classification. In Interspeech, 2008. 3

    2008

  28. [36]

    Phoneme specific modelling and scoring techniques for anti spoofing system

    Gajan Suthokumar, Kaavya Sriskandaraja, Vidhyasaharan Sethu, Chamith Wijenayake, and Eliathamby Ambikaira- jah. Phoneme specific modelling and scoring techniques for anti spoofing system. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019. 2

    2019

  29. [37]

    Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation

    Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee-weon Jung, Junichi Yamagishi, and Nicholas Evans. Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation. In Interspeech, 2022. 1

    2022

  30. [38]

    ASVspoof 2019: Future horizons in spoofed and fake audio detection

    Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Héctor Delgado, Andreas Nautsch, Junichi Yam- agishi, Nicholas Evans, Tomi Kinnunen, and Kong Aik Lee. ASVspoof 2019: Future horizons in spoofed and fake audio detection. In Interspeech, 2019. 5

    2019

  31. [39]

    Superseded-CSTR VCTK Corpus: English Multi- Speaker Corpus for CSTR V oice Cloning Toolkit.University of Edinburgh

    Christophe Veaux, Junichi Yamagishi, Kirsten MacDonald, et al. Superseded-CSTR VCTK Corpus: English Multi- Speaker Corpus for CSTR V oice Cloning Toolkit.University of Edinburgh. The Centre for Speech Technology Research (CSTR), 2016. 5

    2016

  32. [40]

    Media forensics and deepfakes: an overview

    Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE Journal of Selected Topics in Signal Pro- cessing, 14(5):910–932, 2020. 1

    2020

  33. [41]

    Asvspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detection us- ing crowdsourced speech

    Xin Wang, Héctor Delgado, Hemlata Tak, Jee-weon Jung, Hye-jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi Kinnunen, et al. Asvspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detection us- ing...

    2025

  34. [42]

    Mixture of experts fusion for fake au- dio detection using frozen wav2vec 2.0

    Zhiyong Wang, Ruibo Fu, Zhengqi Wen, Jianhua Tao, Xi- aopeng Wang, Yuankun Xie, Xin Qi, Shuchen Shi, Yi Lu, Yukun Liu, et al. Mixture of experts fusion for fake au- dio detection using frozen wav2vec 2.0. arXiv preprint arXiv:2409.11909, 2024. 1

    2024 arXiv

  35. [43]

    To what extent can ASV systems naturally defend against spoofing attacks? In Interspeech,

    Jee weon Jung, Xin Wang, Nicholas Evans, Shinji Watanabe, Hye jin Shim, Hemlata Tak, Siddhant Arora, Junichi Yamag- ishi, and Joon Son Chung. To what extent can ASV systems naturally defend against spoofing attacks? In Interspeech,

  36. [44]

    A robust audio deepfake detection system via multi-view feature

    Yujie Yang, Haochen Qin, Hang Zhou, Chengcheng Wang, Tianyu Guo, Kai Han, and Yunhe Wang. A robust audio deepfake detection system via multi-view feature. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024. 1

    2024

  37. [45]

    Hybrid trans- former architectures with diverse audio features for deepfake speech classification

    Khalid Zaman, Islam JAM Samiul, Melike Sah, Cem Di- rekoglu, Shogo Okada, and Masashi Unoki. Hybrid trans- former architectures with diverse audio features for deepfake speech classification. IEEE Access, 2024. 1

    2024

  38. [46]

    Phoneme-level feature discrepancies: A key to detecting sophisticated speech deepfakes

    Kuiyuan Zhang, Zhongyun Hua, Rushi Lan, Yushu Zhang, and Yifang Guo. Phoneme-level feature discrepancies: A key to detecting sophisticated speech deepfakes. In AAAI Conference on Artificial Intelligence, 2025. 3

    2025

  39. [47]

    Audio deep- fake detection with self-supervised xls-r and sls classifier

    Qishan Zhang, Shuangbing Wen, and Tao Hu. Audio deep- fake detection with self-supervised xls-r and sls classifier. In ACM International Conference on Multimedia, 2024. 1

    2024

  40. [48]

    One-class learn- ing towards synthetic voice spoofing detection

    You Zhang, Fei Jiang, and Zhiyao Duan. One-class learn- ing towards synthetic voice spoofing detection. IEEE Signal Processing Letters, 28:937–941, 2021. 1

    2021

Pith tools

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