Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Source Verification for Speech Deepfakes

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proposes source verification for speech deepfakes: a classifier trained on known generators, used as an embedding extractor, and scored by maximum cosine similarity, tells whether a test track comes from the same unseen…

desk verdict New task definition, solid baseline, but the open-set results need a speaker-overlap control before you can call it source verification. read the letter →

arxiv 2505.14188 v1 pith:KZUSQDCX submitted 2025-05-20 cs.SD eess.AS

Viola Negroni , Davide Salvi , Paolo Bestagini , Stefano Tubaro This is my paper · ORCID
classification cs.SDeess.AS
keywords sourceverificationspeechdeepfakeattributionopen-setgeneralizationembeddingsimilarityaudioforensicsanti-spoofingspeakerdiversitypost-processingrobustness
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

This paper tries to establish that deepfake speech attribution can be reframed as source verification: instead of putting a name on the generator, the system answers whether a test track was made by the same generator as a small set of reference tracks. The proposed pipeline trains a classifier on known generators, uses its penultimate-layer embeddings as a forensic representation, and decides by maximum cosine similarity to the reference set; the paper claims this generalizes to generators never seen in training. If that claim holds, forensic source tracing would no longer require retraining for each new synthesis model, because adding a generator would only require collecting a few reference recordings. The empirical anchor is a ResNet18 embedding model with a 13.0% average equal error rate and 92.2% average AUC across the MLAAD, ASVspoof 2019, and TIMIT-TTS corpora in open-set conditions, with clear degradation under speaker mismatch, language shift, and post-processing.

What carries the argument

The machinery has four parts: a multi-class classifier $C$ trained for source attribution on a fixed set of $N$ generators; the embedding vector $f = C(x)$ taken from the last dense layer before the classification head; a reference set $R$ of five tracks from one generator unseen during training; and the decision statistic $\max_{r \in R} \cos(f_x, f_r)$, the maximum cosine similarity between the test embedding and the reference embeddings. The paper adopts this statistic from reference [17] and uses five reference tracks based on an ad hoc experiment described as omitted for brevity. The load-bearing assumption is that $f$ encodes generator identity rather than speaker, language, or content; Section 4.2 demonstrates the assumption is fragile by showing that models trained on multi-speaker data can lean on speaker cues.

What would settle it

Build a test set in which every unseen generator synthesizes the same single speaker saying the same script, so speaker and content carry no information, and run the source verification pipeline described in the paper; if the average equal error rate approaches chance, the embeddings were carrying speaker or content identity rather than generator identity and the central generalization claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that deepfake source attribution can be recast as source verification, a same-source comparison that works even for generators the model has never seen. A classifier $C$ trained for multi-class attribution on $N$ known generators is reduced to an embedding extractor; a test track is accepted as coming from the same generator as a reference set $R$ when the maximum cosine similarity between its embedding and the reference embeddings is high. On open-set test partitions of MLAAD, ASVspoof 2019, and TIMIT-TTS, the strongest backend reported is ResNet18, with an average equal error rate of 13.0% and an average AUC of 92.2%. The paper's robustness analyses show the score is unevenly sensitive to post-processing, with speech enhancement the most damaging and impulse-response convolution the least, and its speaker-diversity and language experiments show that what the embeddings capture depends on the composition of the training data.

Load-bearing premise

The load-bearing premise is that the embedding space of a classifier trained on known generators keeps separating generators by synthesis identity, not by speaker, language, or content, and that this separation holds for generators the classifier has never seen.

Editorial extensions

If this is right

  • A new deepfake generator can be added to a forensic system by supplying a few reference tracks, with no retraining of the embedding model.
  • Source attribution output becomes a yes/no same-source decision, so the system stays open-set and does not need a fixed catalog of generator classes.
  • Post-processing is a first-order failure mode: speech enhancement degrades verification much more than mobile-phone impulse-response convolution, so deployments must be tested under realistic transmission and enhancement conditions.
  • Training data composition steers what the embeddings capture: multi-speaker training helps on multi-speaker test data, single-speaker training helps on single-speaker test data, and multilingual training helps on non-English test data.

Reading between the lines

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

  • A direct extension the paper leaves implicit is threshold calibration: with reference tracks available, one could estimate a per-generator or global threshold on the maximum-cosine score and report the full operating curve, not only EER/AUC.
  • The paper's reliance on maximum cosine similarity suggests a cheap ablative test: compare max, mean, and top-$k$ pooling of reference similarities; if max is not consistently best, the decision statistic is not the load-bearing part of the pipeline.
  • If generator fingerprints live in a low-dimensional subspace, decomposing the embedding space (for example along eigenvoice-style directions, which the paper's conclusion gestures toward) could yield a more robust and more interpretable verification score; this is an editorial extrapolation, not a paper claim.
  • The generalization claim could be stress-tested with twin generators: the same architecture and the same training speaker set, differing only in random seed or training data; failure to separate them would mean the embeddings are encoding instance-level variation rather than a stable generator identity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces 'source verification' for speech deepfakes: given a reference set of tracks from an unknown generator and a test track, the system decides whether the test track was produced by the same generator. The method trains a classifier C for closed-set attribution on N known generators, extracts embeddings from its penultimate layer, and scores a test track by its maximum cosine similarity to the reference embeddings. Experiments on MLAAD, ASVspoof 2019 LA, TIMIT-TTS, and ADD 2023 evaluate four architectures (ResNet18, LCNN, AASIST, RawNet2) in open-set conditions, reporting EER/AUC. Additional analyses study speaker diversity, language mismatch, and post-processing robustness. The best system (ResNet18) achieves an average EER of 13.0% and AUC of 92.2% across DMLA, ASVspoof, and TIMIT-TTS.

Significance. If the open-set generalization results survive a speaker-confounded control experiment, the paper would provide a useful task formulation and a first systematic benchmark for source verification. Its strengths are the use of public datasets, multiple architectures, and a robustness analysis across post-processing operations. The speaker-diversity findings are candid but currently undermine the central claim, because they suggest the embeddings may encode speaker identity rather than generator identity. The omitted reference-set-size experiment and the absence of variance estimates further limit the strength of the empirical conclusions. This is a promising first exploration, but the evidence in its current form is insufficient to support the headline generalization claim.

major comments (3)
  1. [4.2, Table 2] The central claim that source verification generalizes to unseen generators is not yet established, because the speaker-diversity experiment provides direct evidence that the embeddings may encode speaker identity rather than generator identity. In Table 2, R18-MS outperforms R18-SS on all multi-speaker test sets but underperforms on the single-speaker DTIM_SS; this is exactly the pattern expected if the model uses a speaker-set proxy for each generator. Since in MLAAD each TTS generator is tied to a specific voice dataset, reference and test tracks from the same generator can share voices while impostor generators have different voices. Under these conditions, max cosine similarity can succeed via speaker verification. The paper should add a controlled experiment in which reference and test sets are built with non-overlapping speakers for the same generator (e.g., using different partitions of the underlying voice dataset), or explicitly remove speaker information from the embeddings, and report the resulting EER/AUC. Without such a control, the Table 1 open-set numbers cannot be unambiguously attributed to source verification rather than speaker or voice-dataset recognition.
  2. [3.3] The reference set size R=5 is a free parameter that directly controls the decision statistic (maximum over R cosine similarities), yet the experiment that motivated this choice is explicitly omitted ('omitted here for brevity'). The reported results in Table 1 are therefore conditional on an unreported experiment, and no sensitivity analysis (e.g., R in {1, 3, 5, 10}) is provided. Consequently, the headline EER/AUC values are not reproducible from the manuscript as written. Please either include the R-selection experiment, provide a sensitivity analysis over R, or replace the ad hoc choice with a fixed R justified by published results.
  3. [4.1 and all tables] No variance estimates or number of seeds are reported for any of the EER/AUC values in Tables 1-4. Given that the central claim is empirical generalization, the absence of error bars or a seed variance analysis makes it impossible to assess whether the reported differences between systems (e.g., ResNet18 vs AASIST in Table 1) are statistically meaningful. At minimum, report the mean and standard deviation over at least three training runs or test-set bootstraps for the headline results.
minor comments (5)
  1. [4.3, Table 3] The language transfer experiment is partially confounded by the admitted generator overlap between training and test sets ('some degree of generator overlap between training and test sets, despite the language differences'). Since the goal is to study language robustness, please either ensure zero generator overlap or remove the confounded DMLA_ML result from the generalization claim.
  2. [Throughout] There are several typos and formatting issues: Table 1 shows 'ResNet184.8' (missing space); Section 4.1 writes 'V oice Conversion' with a stray space; Section 4.3 has 'the two systems models' (extra word). Please proofread the manuscript.
  3. [Figure 1] The pipeline diagram (Figure 1) is referenced in Section 2.2 but is not visible in the submitted text. Please ensure the figure is included and legible.
  4. [3.1] The notation for dataset subsets is inconsistent (e.g., 'DMLA, Spk' vs 'DTIM, SS' and 'DMLA, ML' vs 'DMLA, EL'). Please define all subset names in a table or in Section 3.1 to improve readability.
  5. [7. References] The decision statistic is adopted from [17], but no comparison is made with alternative scoring rules (e.g., average cosine similarity, or a calibrated logistic-regression score on the embeddings). A brief ablation would strengthen the claim that the maximum-cosine choice is appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the open-set protocol and externally anchored decision rule keep the central claim empirically independent.

full rationale

The paper's derivation chain is: train an attribution classifier C on labeled known generators following [8]; freeze C and extract its penultimate-layer embeddings; compare a test embedding against R reference embeddings using maximum cosine similarity, a statistic adopted from external work [17]; and report EER/AUC on test generators that are not in C's training set. No step equates the reported prediction to an input by construction. The training labels are generator identities from the training partition, while the test labels come from held-out generators in DMLA, DASV, and DTIM. The decision rule is not fit to the test data: the reference count R=5 is borrowed from an ad hoc experiment in [17] and the max-cosine rule is likewise adopted from [17], not tuned on the present evaluation sets. The self-citations [8] and [21] provide the attribution training methodology and the TIMIT-TTS dataset; these are resources and dataset contributions, not a theorem that forces the verification outcome. Section 4.2 shows that speaker diversity changes performance, and Section 4.3 notes possible generator overlap; these are threats to construct validity or evaluation cleanliness, not circularity, because they concern whether the embeddings encode generator identity rather than whether the reported quantity was defined into existence. No specific reduction of the form Eq. X = Eq. Y can be exhibited, so the honest finding is no significant circularity.

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

The central claim rests on a transferable-embedding assumption, a cosine-similarity decision rule, and clean reference sets; these are modeled, not proven. No new physical entities are proposed. R=5 and the 4-second input window are hand-picked design parameters that the paper does not justify with data.

free parameters (2)
  • Reference set size R = 5
    Chosen based on an omitted ad hoc experiment from [17] (Section 3.3). Verification performance likely depends on this number, and no sensitivity analysis is reported.
  • Input segment length = 4 seconds
    All models are trained and evaluated on fixed 4-second audio segments (Section 3.2). No ablation shows how this affects verification accuracy.
assumptions (4)
  • domain assumption Embeddings from an attribution classifier encode generator-specific information that transfers to unseen generators.
    Stated in Section 2.2 ('we assume that the embeddings f will encode meaningful information about the model used to generate them'). If false, verification reduces to speaker or content matching. The paper probes this in Section 4.2 but does not establish it independently.
  • domain assumption Maximum cosine similarity to a reference set is a sufficient decision statistic for same-generator verification.
    Adopted from [17] in Section 2.2 without comparing to other scoring rules. The reference-set size R=5 is also inherited from [17] via an omitted experiment (Section 3.3).
  • domain assumption The reference set contains only tracks from the claimed generator and is free of systematic recording or speaker confounds.
    The problem formulation in Section 2.1 assumes R is homogeneous. Real forensic references may be contaminated or mixed, which the paper does not test.
  • domain assumption Penultimate-layer embeddings of cross-entropy trained classifiers are discriminative for generator identity.
    The method relies on the penultimate dense layer as a feature extractor (Section 2.2). This is a common deep learning assumption but is not ablated against other layers or pooling strategies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Source Verification for Speech Deepfakes." pith.science (2026). https://pith.science/paper/KZUSQDCX

@misc{pith2026250514188,
  author       = {Pith},
  title        = {Pith review of: Source Verification for Speech Deepfakes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZUSQDCX}},
  note         = {Machine review of arXiv:2505.14188}
}
read the original abstract

With the proliferation of speech deepfake generators, it becomes crucial not only to assess the authenticity of synthetic audio but also to trace its origin. While source attribution models attempt to address this challenge, they often struggle in open-set conditions against unseen generators. In this paper, we introduce the source verification task, which, inspired by speaker verification, determines whether a test track was produced using the same model as a set of reference signals. Our approach leverages embeddings from a classifier trained for source attribution, computing distance scores between tracks to assess whether they originate from the same source. We evaluate multiple models across diverse scenarios, analyzing the impact of speaker diversity, language mismatch, and post-processing operations. This work provides the first exploration of source verification, highlighting its potential and vulnerabilities, and offers insights for real-world forensic applications.

Figures

Figures reproduced from arXiv: 2505.14188 by the authors.

Figure 1
Figure 1. Pipeline of the proposed source verification method for speech deepfakes. A Feature Extractor C is leveraged as an embedding extractor for both the test track and the tracks in the reference set. Then, the cosine similarity is computed between the embedding of the test track and each reference embedding. The decision is made by taking the maximum similarity. 3. Experimental Setup In this section, we outline the expe… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Geometry-Limited Identification Floor and Its Consequences for Voice-Clone Attribution in Professional Voice Actors

    eess.AS 2026-07 conditional novelty 6.0 of 10

    On 1,168 professional voice actors, a misidentification floor in speaker embeddings survives calibration, normalization, and discriminative re-ranking, and the same floor makes fixed-threshold voice-clone attribution ...

  2. Towards Generalized Source Tracing for Codec-Based Deepfake Speech

    cs.SD 2025-06 conditional novelty 5.0 of 10

    SASTNet, which fuses Whisper semantic features with Wav2Vec2 and AudioMAE acoustic features, improves source tracing for codec-based deepfake speech on CodecFake+, while exposing that prior models overfit to silence.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages · cited by 2 Pith papers

  1. [1]

    Introduction The rapid advancement of AI-driven tools has made it easier than ever to create highly realistic synthetic content, including speech. While these advances enable exciting applications in fields like entertainment, accessibility, and human-computer in- teraction, they also pose serious security and privacy risks by fa- cilitating the diffusion...

    arXiv 2023

  2. [2]

    Problem Formulation The source verification task is formally defined as follows

    Proposed Method 2.1. Problem Formulation The source verification task is formally defined as follows. Let R = {x0, x1, . . . ,xR−1} represent a reference set of R syn- thetic speech samples, all generated by the same speech syn- thesis model. Given a test track x produced by an unknown generator, we aim to determine whether x originates from the same gene...

    work page

  3. [3]

    We first introduce the datasets used in our exper- iments, followed by a description of the models employed as classifiers C and their training strategies

    Experimental Setup In this section, we outline the experimental setup we used in our analyses. We first introduce the datasets used in our exper- iments, followed by a description of the models employed as classifiers C and their training strategies. Finally, we detail the methodology used for source verification. 3.1. Datasets We evaluate the proposed fr...

    work page 2019

  4. [4]

    Results In this section, we evaluate the proposed pipeline on the source verification task. We assess its performance under various ex- perimental conditions, analyzing its robustness to factors such as speaker diversity, language mismatches, and post-processing operations. 4.1. Source Verification As an initial investigation, we assess the feasibility of...

    work page

  5. [5]

    Conclusions In this paper, we introduced the source verification task, a novel approach to source tracing for speech deepfakes. Inspired by ASV, the proposed method determines whether a query track was generated by the same model as a set of reference sam- ples, eliminating the need for exhaustive training on every pos- sible generator. Our evaluation dem...

    work page

  6. [6]

    Telecommunications of the Future

    Acknowledgements This work was supported by the FOSTERER project, funded by the Italian Ministry of Education, University, and Re- search within the PRIN 2022 program. This work was par- tially supported by the European Union - Next Generation EU under the Italian National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.3, CUP D4...

    work page 2022

  7. [7]

    Multi-task learning improves synthetic speech detection,

    Y . Mo and S. Wang, “Multi-task learning improves synthetic speech detection,” in IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2022

    work page 2022

  8. [8]

    Lightweight voice spoof- ing detection using improved one-class learning and knowledge distillation,

    Y . Ren, H. Peng, L. Li, and Y . Yang, “Lightweight voice spoof- ing detection using improved one-class learning and knowledge distillation,” IEEE Transactions on Multimedia, 2023

    work page 2023

Show all 39 references
  1. [9]

    A robust audio deepfake detection system via multi-view fea- ture,

    Y . Yang, H. Qin, H. Zhou, C. Wang, T. Guo, K. Han, and Y . Wang, “A robust audio deepfake detection system via multi-view fea- ture,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    2024

  2. [10]

    Audio transformer for synthetic speech detection via formant magnitude and phase analysis,

    L. Cuccovillo, M. Gerhardt, and P. Aichroth, “Audio transformer for synthetic speech detection via formant magnitude and phase analysis,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    2024

  3. [11]

    Leveraging mixture of experts for improved speech deepfake de- tection,

    V . Negroni, D. Salvi, A. I. Mezza, P. Bestagini, and S. Tubaro, “Leveraging mixture of experts for improved speech deepfake de- tection,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  4. [12]

    Freeze and learn: Continual learning with selective freezing for speech deepfake detection,

    D. Salvi, V . Negroni, L. Bondi, P. Bestagini, and S. Tubaro, “Freeze and learn: Continual learning with selective freezing for speech deepfake detection,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  5. [13]

    Synthetic speech detection through short-term and long-term prediction traces,

    C. Borrelli, P. Bestagini, F. Antonacci, A. Sarti, and S. Tubaro, “Synthetic speech detection through short-term and long-term prediction traces,” EURASIP Journal on Information Security , vol. 2021, 2021

    2021

  6. [14]

    Exploring the synthetic speech attribution problem through data-driven detectors,

    D. Salvi, P. Bestagini, and S. Tubaro, “Exploring the synthetic speech attribution problem through data-driven detectors,” in IEEE International Workshop on Information Forensics and Se- curity (WIFS), 2022

    2022

  7. [15]

    Add 2023: the second audio deepfake detection challenge,

    J. Yi, J. Tao, R. Fu, X. Yan, C. Wang, T. Wang, C. Y . Zhang, X. Zhang, Y . Zhao, Y . Renet al., “Add 2023: the second audio deepfake detection challenge,” in DADA@ IJCAI, 2023

    2023

  8. [16]

    Synthetic speech attribution: Highlights from the ieee signal processing cup 2022 student com- petition [sp competitions],

    D. Salvi, C. Borrelli, P. Bestagini, F. Antonacci, M. Stamm, L. Marcenaro, and A. Majumdar, “Synthetic speech attribution: Highlights from the ieee signal processing cup 2022 student com- petition [sp competitions],” IEEE Signal Processing Magazine , vol. 40, no. 6, pp. 92–98, 2023

    2022

  9. [17]

    Deepfake algorithm recognition system with augmented data for add 2023 challenge

    X.-M. Zeng, J.-T. Zhang, K. Li, Z.-L. Liu, W.-L. Xie, and Y . Song, “Deepfake algorithm recognition system with augmented data for add 2023 challenge.” in DADA@ IJCAI, 2023

    2023

  10. [18]

    Detect- ing unknown speech spoofing algorithms with nearest neighbors

    J. Lu, Y . Zhang, Z. Li, Z. Shang, W. Wang, and P. Zhang, “Detect- ing unknown speech spoofing algorithms with nearest neighbors.” in DADA@ IJCAI, 2023

    2023

  11. [19]

    Deepfake algorithm recog- nition through multi-model fusion based on manifold measure

    Y . Tian, Y . Chen, Y . Tang, and B. Fu, “Deepfake algorithm recog- nition through multi-model fusion based on manifold measure.” in DADA@ IJCAI, 2023

    2023

  12. [20]

    Source tracing: detect- ing voice spoofing,

    T. Zhu, X. Wang, X. Qin, and M. Li, “Source tracing: detect- ing voice spoofing,” in IEEE Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2022

    2022

  13. [21]

    Source trac- ing of audio deepfake systems,

    N. Klein, T. Chen, H. Tak, R. Casal, and E. Khoury, “Source trac- ing of audio deepfake systems,” in Proc. INTERSPEECH, 2024

    2024

  14. [22]

    Deepfake audio detection by speaker verification,

    A. Pianese, D. Cozzolino, G. Poggi, and L. Verdoliva, “Deepfake audio detection by speaker verification,” in IEEE International Workshop on Information Forensics and Security (WIFS), 2022

    2022

  15. [23]

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

    ——, “Training-free deepfake voice recognition by leveraging large-scale pre-trained models,” in Proc. ACM Workshop on In- formation Hiding and Multimedia Security, 2024

    2024

  16. [24]

    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,” IEEE Interna- tional Joint Conference on Neural Networks (IJCNN), 2024

    2024

  17. [25]

    Using mlaad for source tracing of audio deepfakes,

    N. M ¨uller, “Using mlaad for source tracing of audio deepfakes,” https://deepfake-total.com/sourcetracing, Fraunhofer AISEC, 11 2024

    2024

  18. [26]

    ASVspoof 2019: Future horizons in spoofed and fake audio de- tection,

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T. Kinnunen, and K. A. Lee, “ASVspoof 2019: Future horizons in spoofed and fake audio de- tection,” in Proc. INTERSPEECH, 2019

    2019

  19. [27]

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

    D. Salvi, B. Hosler, P. Bestagini, M. C. Stamm, and S. Tubaro, “TIMIT-TTS: a Text-to-Speech Dataset for Multimodal Synthetic Media Detection,” IEEE Access, 2023

    2023

  20. [28]

    The VidTIMIT database,

    C. Sanderson, “The VidTIMIT database,” IDIAP, Tech. Rep., 2002

    2002

  21. [29]

    The LJ Speech Dataset,

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

    2017

  22. [30]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015

    2015

  23. [31]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  24. [32]

    A light CNN for deep face rep- resentation with noisy labels,

    X. Wu, R. He, Z. Sun, and T. Tan, “A light CNN for deep face rep- resentation with noisy labels,” IEEE Transactions on Information Forensics and Security, vol. 13, no. 11, 2018

    2018

  25. [33]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inIEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2021

    2021

  26. [34]

    Speaker recognition from raw wave- form with sincnet,

    M. Ravanelli and Y . Bengio, “Speaker recognition from raw wave- form with sincnet,” in IEEE spoken language technology work- shop (SLT), 2018

    2018

  27. [35]

    Aasist: Audio anti-spoofing using inte- grated spectro-temporal graph attention networks,

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using inte- grated spectro-temporal graph attention networks,” inIEEE inter- national conference on acoustics, speech and signal processing (ICASSP), 2022

    2022

  28. [36]

    Cross-lingual text- independent speaker verification using unsupervised adversarial discriminative domain adaptation,

    W. Xia, J. Huang, and J. H. Hansen, “Cross-lingual text- independent speaker verification using unsupervised adversarial discriminative domain adaptation,” inIEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2019

    2019

  29. [37]

    Attention is all you need in speech separation,

    C. Subakan, M. Ravanelli, S. Cornell, M. Bronzi, and J. Zhong, “Attention is all you need in speech separation,” in IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021

    2021

  30. [38]

    Poliphone: A dataset for smartphone model identification from audio recordings,

    D. Salvi, D. U. Leonzio, A. Giganti, C. Eutizi, S. Mandelli, P. Bestagini, and S. Tubaro, “Poliphone: A dataset for smartphone model identification from audio recordings,” IEEE Access, 2025

    2025

  31. [39]

    Eigenvoices for speaker adaptation

    R. Kuhn, P. Nguyen, J.-C. Junqua, L. Goldwasser, N. Niedzielski, S. Fincke, K. L. Field, and M. Contolini, “Eigenvoices for speaker adaptation.” in ICSLP, vol. 98, 1998, pp. 1774–1777

    1998

Pith tools

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