Pith. sign in

REVIEW 2 major objections 1 minor 46 references

GRIDEX generates structured explanations for deepfake spectrograms by selecting top anomalous regions and mapping them to acoustic categories.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-26 19:50 UTC pith:UY4AOMIW

load-bearing objection GRIDEX claims to be the first to ground structured forensic explanations in specific spectrogram regions for deepfakes, but the abstract gives no metrics, dataset details, or validation of whether those explanations are accurate or useful. the 2 major comments →

arxiv 2606.18738 v1 pith:UY4AOMIW submitted 2026-06-17 cs.SD

GRIDEX: Grid-Grounded Forensic Explanations for Deepfake Spectrogram Analysis

Thi Ngan Ha Do , Tingmin Wu , Alsharif Abuadbba , Kristen Moore This is my paper
classification cs.SD
keywords deepfake detectionspectrogram analysisforensic explanationsanomaly localizationvision-language modelsaudio forensicsregional grounding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a pipeline that takes a deepfake spectrogram and produces explanations by first identifying the most anomalous patches and then describing each one according to a fixed set of acoustic fields. These fields cover temporal location, spectral properties, phonetic content, and an interpretive text summary. The goal is to replace resource-heavy manual review of entire spectrograms with focused, verifiable evidence that connects localized visual cues to acoustic meaning. Training uses supervised fine-tuning followed by a reinforcement step to improve the quality of the region selection and the resulting descriptions. Experiments compare the output against standard vision-language model approaches on the authors' dataset.

Core claim

GRIDEX is the first framework to generate structured forensic explanations using regional grounding for deepfake spectrograms. Given an input spectrogram, the pipeline selects the top-K anomalous regions and produces for each region an explanation that follows a schema of categorical acoustic fields including temporal, spectral, phonetic information and interpretation text.

What carries the argument

A two-stage pipeline that selects the top-K anomalous regions in the spectrogram and maps each to a fixed schema of categorical acoustic fields.

Load-bearing premise

Selecting the top-K anomalous regions and describing them with a fixed set of acoustic categories will produce explanations that are both verifiable by experts and better for forensic work than full manual analysis.

What would settle it

A controlled test in which forensic analysts are given either GRIDEX explanations or full spectrograms and asked to identify confirmed deepfake artifacts, measuring whether the structured outputs improve accuracy or speed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The method improves artifact localization accuracy over standard vision-language model baselines.
  • Explanation quality increases when the two-stage training combines supervised fine-tuning with Group Relative Policy Optimization.
  • Structured outputs in the acoustic schema make it easier to check whether a given region truly indicates a spoof cue.
  • Narrowing attention to top-K regions reduces the effort needed for manual forensic review of spectrograms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same region-selection plus schema approach could be tested on other types of audio manipulation beyond current deepfake generators.
  • If the categorical fields prove stable across datasets, the explanations might serve as input features for downstream detection models.
  • Integration with existing audio classifiers could allow the explanations to be generated on demand rather than after full classification.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript proposes GRIDEX, a pipeline that generates structured forensic explanations for deepfake spectrograms. Given an input spectrogram, it (i) selects the top-K anomalous regions and (ii) maps each to a fixed schema of categorical acoustic fields (temporal, spectral, phonetic information plus interpretation text). The pipeline is trained in two stages (supervised fine-tuning followed by Group Relative Policy Optimization) and is presented as the first framework to combine regional grounding with this explanation schema. Experiments on an internal dataset are said to demonstrate improved artifact localization and explanation quality relative to VLM baselines; the dataset and code are promised for release.

Significance. If the central claims hold, the work could meaningfully advance forensic deepfake analysis by moving beyond classification accuracy to produce locally grounded, schema-structured explanations that are easier to verify against acoustic evidence. The two-stage training recipe and the commitment to release data/code are constructive elements. However, the significance cannot be evaluated without quantitative evidence that the generated explanations are accurate to underlying spoof cues or useful in forensic settings.

major comments (2)
  1. [Abstract] Abstract: the central claim that GRIDEX produces 'verifiable' forensic explanations rests on the untested assumption that top-K region selection plus mapping to the fixed acoustic schema yields outputs that align with known spoof artifacts; no expert-agreement scores, alignment metrics, or downstream forensic-utility results are supplied to support this.
  2. [Abstract] Abstract: the reported improvements in 'artifact localization and explanation quality' are stated without any numerical metrics, baseline configurations, dataset statistics, or error analysis, rendering it impossible to assess whether the experimental results actually substantiate the forensic-utility claim.
minor comments (1)
  1. [Abstract] Abstract: the dataset is referred to only as 'our dataset' with no description of its size, spoofing methods, or annotation protocol.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed feedback on the abstract. The comments correctly identify that the abstract makes claims about verifiability and improvements without supporting quantitative details. We will revise the abstract to address these issues by incorporating key metrics and clarifying the basis for our claims. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that GRIDEX produces 'verifiable' forensic explanations rests on the untested assumption that top-K region selection plus mapping to the fixed acoustic schema yields outputs that align with known spoof artifacts; no expert-agreement scores, alignment metrics, or downstream forensic-utility results are supplied to support this.

    Authors: We agree that the abstract's use of 'verifiable' would be strengthened by explicit supporting evidence. The schema is intended to enable verification by linking regions to established acoustic properties of spoofing artifacts, but the current manuscript does not include expert-agreement scores or alignment metrics. In the revision we will add a new subsection reporting inter-annotator agreement from a small-scale expert study on explanation accuracy and any available downstream forensic utility indicators. revision: yes

  2. Referee: [Abstract] Abstract: the reported improvements in 'artifact localization and explanation quality' are stated without any numerical metrics, baseline configurations, dataset statistics, or error analysis, rendering it impossible to assess whether the experimental results actually substantiate the forensic-utility claim.

    Authors: The referee is correct that the abstract reports improvements qualitatively without numbers. The full manuscript contains quantitative results in the experiments section (localization IoU, explanation quality scores against VLM baselines, dataset size and split statistics, and error analysis), but these were not summarized in the abstract. We will revise the abstract to include the principal numerical results, baseline details, and a brief reference to the error analysis. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; pipeline description is self-contained

full rationale

The paper presents GRIDEX as a two-stage pipeline (top-K anomalous region selection followed by schema-based explanation generation) trained via SFT + GRPO on an internal dataset, with claims of improved localization over VLM baselines. No equations, fitted parameters renamed as predictions, self-citations, uniqueness theorems, or ansatzes appear in the abstract or described approach. The central claim reduces to an empirical pipeline whose outputs are evaluated externally rather than derived by construction from its inputs. This matches the default non-circular case for descriptive ML frameworks without mathematical reduction steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no information on free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5775 in / 1077 out tokens · 38059 ms · 2026-06-26T19:50:54.014220+00:00 · methodology

0 comments
read the original abstract

The advancement of speech generation technologies has made artificial speech increasingly realistic. Although modern classification models can achieve high accuracy when it comes to deepfake detection, they do not produce evidences such as indicating where spoof cues appear in the spectrogram and what they imply acoustically, limiting their usefulness in forensic settings. Manual analysis of full spectrograms is resource-intensive, so evidence should narrow attention to the most diagnostic regions. Moreover, existing explainability methods have limited capabilities in connecting contextual attributes to localized evidence, making explanations harder to verify. To overcome this limitation, we propose GRIDEX, a pipeline that, when given a deepfake spectrogram, generates forensic explanations of its anomalies. The pipeline (i) selects top-K anomalous regions in the spectrogram and (ii) produces an explanation for each anomaly. The explanations follow a schema of categorical acoustic fields, including temporal, spectral, phonetic information and interpretation text. To our knowledge, this is the first framework to generate structured forensic explanations using regional grounding for deepfake spectrograms. GRIDEX is trained with a two-stage learning paradigm that combines supervised fine-tuning (SFT) with Group Relative Policy Optimization (GRPO). Experiments on our dataset show improved artifact localization and explanation quality over strong vision-language model (VLM) baselines. The dataset and code will be released upon publication.

Figures

Figures reproduced from arXiv: 2606.18738 by Alsharif Abuadbba, Kristen Moore, Thi Ngan Ha Do, Tingmin Wu.

Figure 1
Figure 1. Figure 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spectrogram region segmentation methods. From left to right: raw spectrogram, uniform grid, SLIC superpixels, and SAM masks. We evaluate three segmentation strategies for defining can￾didate regions [18]: (1) uniform grids, (2) SLIC superpixels [19], and (3) SAM [20]. Each strategy produces a set of ar￾tifact regions by overlaying its division masks against that the difference maps (Sec. 3.1). As illustrat… view at source ↗
Figure 5
Figure 5. Figure 5: Top-3 region overlap distribution by vocoder. HiFi￾GAN and WaveGlow concentrate mass at overlap=2, whereas HN-NSF variants concentrate at overlap=1. while 7,103 (11.5%) match 2/3 fields, 3,087 (5.0%) match 1/3 fields, and 68 (0.11%) match none of the fields. Human correction. 83.3% of the samples exactly match all three fields under parsing rules. For the remaining cases, mismatches primarily arise from mi… view at source ↗
Figure 6
Figure 6. Figure 6: Query 1: Localization prompt. SYSTEM PROMPT You are an expert in deepfake speech spectrogram forensics. You are given a spectrogram and transcript, and you have already selected exactly three region IDs (ID1, ID2, ID3) in order. For each region, infer: time (T is one of {speech, non-speech}), frequency band (F is one of {low, mid, high}), phonetic category (P is one of {consonant, vowel, unvoiced}), and a … view at source ↗
Figure 7
Figure 7. Figure 7: Query 2: Structured explanation prompt [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generated explanation samples on most prominent artifacts. 5.4. Ablation Studies 5.4.1. Granularity levels Setting R@3↑ nDCG@3↑ mAP↑ Grid 0.309 0.333 0.264 SLIC 0.261 0.2911 0.214 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 13 canonical work pages · 3 internal anchors

  1. [1]

    Introduction Deepfake speech increasingly threatens trust in voice-based sys- tems and audio evidence [1, 2], making it harder for users and systems to distinguish genuine voices from synthetic ones. There is a growing need to understand not only whether au- dio is spoofed, but also where suspicious evidence appears in a spectrogram and what acoustic feat...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Task Overview The GRIDEX pipeline serves two objectives: Artifact local- ization and Artifact interpretation

    Problem Definition 2.1. Task Overview The GRIDEX pipeline serves two objectives: Artifact local- ization and Artifact interpretation. These two objectives are achieved through two queries which are sequentially inserted into a VLM. These queries will be referred to as Query 1 and Query 2 respectively. Given a deepfake spectrogram imageS, we first overlay ...

  3. [3]

    We therefore con- struct a dataset that provides region-level artifact supervision, structured metadata, and forensic explanations for each spec- trogram

    Data Construction and Region Annotation Prior deepfake audio datasets offer only spoof labels, providing limited value for interpretability research. We therefore con- struct a dataset that provides region-level artifact supervision, structured metadata, and forensic explanations for each spec- trogram. In the following, we detail how we derive artifact s...

  4. [4]

    Harmonic degradation: smeared harmonics

  5. [5]

    Formant fading: blurred vowel bands

  6. [6]

    Periodic texture: repeated stripe patterns

  7. [7]

    region 11

    Noise flattening: smooth unvoiced noise. TASK Given a region ID with time-frequency and phonetic context, describe the single most dominant artifact and its likely audible effect. THINKING STEPS (1) Compare expected genuine pattern vs. observed region. (2) Describe the main spectro-temporal distortion. (3) Write a forensic explanation linking cue to audib...

  8. [8]

    "); (T2=..., F2=..., P2=..., E2=

    GRIDEX System Design 4.1. Overview GRIDEX is a two-step VLM pipeline that first localizes the top- Kartifact regions on a spectrogram (Query 1) and then gener- ates structured explanations for those selections (Query 2). SYSTEM PROMPT As an expert in deepfake speech spectrogram forensics, detect regions containing spoof artifacts by analyzing spectrogram ...

  9. [9]

    Dataset We use the V ocV4 vocoder-based dataset [31] to generate ground truth (Sec

    Evaluation 5.1. Dataset We use the V ocV4 vocoder-based dataset [31] to generate ground truth (Sec. 3). V ocV4 contains bona fide utterances from Table 1:Pipeline evaluation compares the performance of VLM baselines against GRIDEX for Query 1 and Query 2. Zero shot (ZS) reports Query 1 only, Oracle reports Query 2 only conditioned on ground-truth top-3 re...

  10. [10]

    Conclusion We propose GRIDEX, a two-stage VLM framework for grid- grounded forensic explanations of deepfake spectrograms. This framework performs two sequential tasks: Query 1 identifies the top-3 most suspicious regions on a fixed grid size, and Query 2 produces a structured per-region explanation tuple that maps each claim to a specific region ID and i...

  11. [11]

    Au- dio deepfake detection: A survey,

    J. Yi, C. Wang, J. Tao, X. Zhang, C. Y . Zhang, and Y . Zhao, “Au- dio deepfake detection: A survey,”arXiv, 2023

    2023

  12. [12]

    Audio deepfake detection: What has been achieved and what lies ahead,

    B. Zhang, H. Cui, V . Nguyen, and M. Whitty, “Audio deepfake detection: What has been achieved and what lies ahead,”Sensors, vol. 25, no. 7, p. 1989, 2025

    1989

  13. [13]

    Unmasking digital deceptions: An in- tegrative review of deepfake detection, multimedia forensics, and cybersecurity challenges,

    S. Singh and A. Dhumane, “Unmasking digital deceptions: An in- tegrative review of deepfake detection, multimedia forensics, and cybersecurity challenges,”MethodsX, p. 103632, 2025

    2025

  14. [14]

    Towards relatable explainable ai with the perceptual process,

    W. Zhang and B. Y . Lim, “Towards relatable explainable ai with the perceptual process,” inProceedings of the 2022 CHI confer- ence on human factors in computing systems, 2022, pp. 1–24

    2022

  15. [15]

    Anomaly detection and localization for speech deepfakes via fea- ture pyramid matching,

    E. Coletta, D. Salvi, V . Negroni, D. U. Leonzio, and P. Bestagini, “Anomaly detection and localization for speech deepfakes via fea- ture pyramid matching,” in2025 33rd European Signal Process- ing Conference (EUSIPCO). IEEE, 2025, pp. 820–824

    2025

  16. [16]

    Explainable deepfake and spoofing detection: an attack analysis using SHapley Additive exPlanations,

    W. Ge, M. Todisco, and N. Evans, “Explainable deepfake and spoofing detection: an attack analysis using SHapley Additive exPlanations,” inThe Speaker and Language Recognition Workshop (Odyssey 2022), Beijing, China, Jun. 2022, pp. 70–76. [Online]. Available: https://www.isca-archive.org/odyssey 2022/ ge22 odyssey.pdf

    2022

  17. [17]

    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,”arXiv preprint arXiv:2305.13774, 2023

    work page arXiv 2023

  18. [18]

    Phoneme Discretized Saliency Maps for Explainable Detection of AI- Generated V oice,

    S. Gupta, M. Ravanelli, P. Germain, and C. Subakan, “Phoneme Discretized Saliency Maps for Explainable Detection of AI- Generated V oice,” inInterspeech 2024, 2024, pp. 3295–3299

    2024

  19. [19]

    Explainable ai for forensic speech authentication within cognitive and computational neuroscience,

    Z. Cheng, H. Yang, Y . Xiong, and X. Hu, “Explainable ai for forensic speech authentication within cognitive and computational neuroscience,”Frontiers in Neuroscience, vol. V olume 19 - 2025,

    2025

  20. [20]

    Available: https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2025.1692122

    [Online]. Available: https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2025.1692122

    work page doi:10.3389/fnins.2025.1692122 2025

  21. [21]

    One wave to explain them all: A unifying perspective on post-hoc explainability,

    G. Kasmi, A. Brunetto, J. Parekhet al., “One wave to explain them all: A unifying perspective on post-hoc explainability,” 2024

    2024

  22. [22]

    Audio explainable artificial intel- ligence: A review,

    A. Akman and B. W. Schuller, “Audio explainable artificial intel- ligence: A review,”Intelligent Computing, vol. 3, p. 0074, 2024

    2024

  23. [23]

    Towards faithful model explanation in nlp: A survey,

    Q. Lyu, M. Apidianaki, and C. Callison-Burch, “Towards faithful model explanation in nlp: A survey,”Computational Linguistics, vol. 50, no. 2, pp. 657–723, 2024

    2024

  24. [24]

    Explainable artificial intelligence for deepfake detection: Pipeline, open source and comparisons,

    H. Li, J. Chen, B. Wang, S. Garc´ıa, and D. Camacho, “Explainable artificial intelligence for deepfake detection: Pipeline, open source and comparisons,”Expert Systems, vol. 43, no. 3, p. e70222, 2026, e70222 EXSY-Oct-25-4700. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.70222

    work page doi:10.1111/exsy.70222 2026

  25. [25]

    AI-synthesized voice de- tection using neural vocoder artifacts,

    C. Sun, S. Jia, S. Hou, and S. Lyu, “AI-synthesized voice de- tection using neural vocoder artifacts,”arXiv, 2023, accepted at CVPRW 2023

    2023

  26. [26]

    Explainable deepfake detection: A multi-model framework with human-interpretable rationales for legal investigation purposes,

    N. Bharati, P. Wong, S. K. Most ´efaoui, D. Kbaier, and J. Collie, “Explainable deepfake detection: A multi-model framework with human-interpretable rationales for legal investigation purposes,” Machine Learning with Applications, p. 100819, 2025

    2025

  27. [27]

    Media foren- sics considerations on deepfake detection with hand-crafted fea- tures,

    D. Siegel, C. Kraetzer, S. Seidlitz, and J. Dittmann, “Media foren- sics considerations on deepfake detection with hand-crafted fea- tures,”Journal of imaging, vol. 7, no. 7, p. 108, 2021

    2021

  28. [28]

    A data- driven diffusion-based approach for audio deepfake explanations,

    P. Grinberg, A. Kumar, S. Koppisetti, and G. Bharaj, “A data- driven diffusion-based approach for audio deepfake explanations,” arXiv preprint arXiv:2506.03425, 2025

    work page arXiv 2025

  29. [29]

    GPT-4V-AD: Exploring grounding potential of VQA- oriented GPT-4V for zero-shot anomaly detection,

    J. Zhang, H. He, X. Chen, Z. Xue, Y . Wang, C. Wang, L. Xie, and Y . Liu, “GPT-4V-AD: Exploring grounding potential of VQA- oriented GPT-4V for zero-shot anomaly detection,”arXiv, 2024, arXiv:2311.02612v2

    work page arXiv 2024

  30. [30]

    Slic superpixels compared to state-of-the-art su- perpixel methods,

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk, “Slic superpixels compared to state-of-the-art su- perpixel methods,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 11, pp. 2274–2282, 2012

    2012

  31. [31]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Gir- shick, “Segment anything,”arXiv, Apr. 2023

    2023

  32. [32]

    Performance of the segment anything model in various rfi/events detection in radio astronomy,

    Y . Yang, F. Zhao, R. Liang, Q. Guo, J. Gu, Y . Huang, and Y . Yu, “Performance of the segment anything model in various rfi/events detection in radio astronomy,”Publications of the Astronomical Society of Australia, vol. 42, p. e019, 2025

    2025

  33. [33]

    Interpretable all-type audio deepfake detec- tion with audio llms via frequency-time reinforcement learning,

    Y . Xie, X. Guo, J. Zhou, T. Wang, J. Liu, R. Fu, X. Wang, H. Cheng, and L. Ye, “Interpretable all-type audio deepfake detec- tion with audio llms via frequency-time reinforcement learning,” arXiv preprint arXiv:2601.02983, 2026

    work page arXiv 2026

  34. [34]

    Where llm agents fail and how they can learn from failures

    K. Zhu, Z. Liu, B. Li, M. Tian, Y . Yang, J. Zhang, P. Han, Q. Xie, F. Cui, W. Zhanget al., “Where llm agents fail and how they can learn from failures,”arXiv preprint arXiv:2509.25370, 2025

    work page arXiv 2025

  35. [35]

    Uncertainty propagation on llm agent,

    Q. Zhao, D. Li, Y . Liu, W. Cheng, Y . Sun, M. Oishi, T. Osaki, K. Matsuda, H. Yao, C. Zhaoet al., “Uncertainty propagation on llm agent,” inProceedings of the 63rd Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 6064–6073

    2025

  36. [36]

    LoRA in LoRA: Towards Parameter-Efficient Architecture Expansion for Continual Visual Instruction Tuning

    C. Che, Z. Wang, P. Yang, Q. Wang, H. Ma, and Z. Shi, “Lora in lora: Towards parameter-efficient architecture expan- sion for continual visual instruction tuning,”arXiv preprint arXiv:2508.06202, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    Understanding catastrophic forgetting in language models via implicit inference,

    S. Kotha, J. M. Springer, and A. Raghunathan, “Understanding catastrophic forgetting in language models via implicit inference,” arXiv preprint arXiv:2309.10105, 2023

    work page arXiv 2023

  38. [38]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “Deepseek-r1: Incentivizing reason- ing capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Bridging the training-inference gap in llms by leveraging self-generated tokens,

    Z. Cen, Y . Liu, S. Zeng, P. Chaudhari, H. Rangwala, G. Karypis, and R. Fakoor, “Bridging the training-inference gap in llms by leveraging self-generated tokens,”arXiv preprint arXiv:2410.14655, 2024

    work page arXiv 2024

  40. [40]

    Cumulated gain-based evalua- tion of ir techniques,

    K. J ¨arvelin and J. Kek ¨al¨ainen, “Cumulated gain-based evalua- tion of ir techniques,”ACM Transactions on Information Systems (TOIS), vol. 20, no. 4, pp. 422–446, 2002

    2002

  41. [41]

    Swift: a scalable lightweight infras- tructure for fine-tuning,

    Y . Zhao, J. Huang, J. Hu, X. Wang, Y . Mao, D. Zhang, Z. Jiang, Z. Wu, B. Ai, A. Wanget al., “Swift: a scalable lightweight infras- tructure for fine-tuning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 28, 2025, pp. 29 733–29 735

    2025

  42. [42]

    Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,

    X. Wang and J. Yamagishi, “Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,”arXiv, 2022, arXiv:2210.10570; accepted at ICASSP 2023

    work page arXiv 2022

  43. [43]

    PASS (phonetic assessment of spoofed speech): To- wards a human-expert-based framework for spoofed speech detection,

    D. D. Lee, K. McDougall, F. Kelly, and A. Alexan- der, “PASS (phonetic assessment of spoofed speech): To- wards a human-expert-based framework for spoofed speech detection,” inBook of Abstracts: 31st IAFPA Confer- ence, Zurich, Switzerland, Jul. 2023, pp. 31–32, inter- national Association for Forensic Phonetics and Acoustics (IAFPA). [Online]. Available:...

    2023

  44. [44]

    Univnet: A neu- ral vocoder with multi-resolution spectrogram discriminators for high-fidelity waveform generation,

    W. Jang, D. Lim, J. Yoon, B. Kim, and J. Kim, “Univnet: A neu- ral vocoder with multi-resolution spectrogram discriminators for high-fidelity waveform generation,”arXiv, 2021

    2021

  45. [45]

    Upsampling arti- facts in neural audio synthesis,

    J. Pons, S. Pascual, G. Cengarle, and J. Serr `a, “Upsampling arti- facts in neural audio synthesis,” inICASSP 2021-2021 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 3005–3009

    2021

  46. [46]

    Av- ocodo: Generative adversarial network for artifact-free vocoder,

    T. Bak, J. Lee, H. Bae, J. Yang, J.-S. Bae, and Y .-S. Joo, “Av- ocodo: Generative adversarial network for artifact-free vocoder,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 11, 2023, pp. 12 562–12 570. A. Appendix A.1. Details of Training Hyperparameters We report shared training settings (Table 6) and list stage- speci...

    2023