Pith. sign in

REVIEW 4 major objections 5 minor 66 references

Blending a 230-hour, 30-API corpus of synthetic speech into training shrinks the gap between lab benchmarks and real-world spoofing, and a local-attention network built on top of it sets new state-of-the-art error rates.

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 · deepseek-v4-flash

2026-08-03 17:57 UTC pith:H6FGN7IN

load-bearing objection The MultiAPI Spoof dataset is the real contribution; the local-attention model is incremental, and the unseen-API robustness claim needs a leakage analysis before you trust it. the 4 major comments →

arxiv 2512.07352 v5 pith:H6FGN7IN submitted 2025-12-08 cs.SD

MultiAPI Spoof: A Multi-API Dataset and Local-Attention Network for Speech Anti-spoofing Detection

Xueping Zhang , Zhenshan Zhang , Yechen Wang , Linxi Li , Liwei Jin , Ming Li This is my paper
classification cs.SD
keywords speech anti-spoofingdeepfake detectionmulti-API datasetlocal attentionNes2Net-LAAPI tracingcross-domain generalizationsynthetic speech
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 argues that existing speech anti-spoofing benchmarks under-represent the diversity of real-world synthetic speech because they are built from a handful of public TTS/VC models. To close this gap, it introduces MultiAPI Spoof, a 230-hour English corpus of spoofed audio generated through 30 distinct commercial, open-source, and web-based APIs, paired with an equal amount of bona fide speech. It then proposes Nes2Net-LA, a local-attention extension of the nested Nes2Net feature extractor, and claims this model reaches state-of-the-art error rates on several benchmarks while improving detection of audio from unseen APIs. The paper also defines an API tracing task that attributes spoofed audio to its generating API, useful for forensic source identification.

Core claim

On its own terms, the paper asserts that the narrow range of synthesis systems in existing anti-spoofing benchmarks causes a measurable domain gap to real-world commercial TTS/VC APIs, and that closing this gap is a matter of training-data breadth plus a modest architectural change. Concretely, it claims that adding MultiAPI Spoof to the training pool cuts equal error rate on the MultiAPI Spoof test set from 7.30% to 0.70% for one strong baseline, and that the proposed local-attention variant Nes2Net-LA reaches 1.42% EER on ITW and 5.64% on AI4T without any data augmentation. It further claims the local-attention mechanism improves generalization to unseen APIs, and that the API tracing task

What carries the argument

MultiAPI Spoof is a corpus of ~230 hours of spoofed English speech generated through 30 distinct APIs (commercial services, open-source models, online TTS websites), balanced 1:1 with bona fide CommonVoice speech, partitioned by API into seen (A0-A20), development-unseen (A21-A23), and evaluation-unseen (A24-A29) groups. Nes2Net-LA is the paper's architectural contribution: it wraps the nested multi-scale Res2Net feature extractor with a sliding-window local self-attention operator, so each channel subset attends only to its K neighbors (K=1 in experiments) rather than the global sequence.

Load-bearing premise

The held-out 'unseen' APIs are assumed to be genuinely novel relative to the training APIs, with no shared accidental acoustic features (codec, text, recording chain) that make them easy to separate; the paper does not quantify acoustic overlap between seen and unseen API conditions.

What would settle it

If a randomly selected new TTS API not in the 30-API list yields an EER close to chance for a model trained on MultiAPI Spoof, the claimed cross-domain generalization to unseen generators would be falsified. A simpler check: measure the embedding-space distance between seen and unseen APIs; if unseen examples already sit inside seen clusters in a nearest-neighbor sense, the 'unseen' evaluation is not measuring generalization to truly new conditions.

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

If this is right

  • Adding MultiAPI Spoof to the training pool improves detection on existing benchmarks like ITW and AI4T, not just on the new corpus itself.
  • Models trained with MultiAPI Spoof generalize to spoofed audio from previously unseen APIs, as shown by the improved unseen-split EERs.
  • The API tracing task can attribute spoofed audio to one of 21 seen APIs with high F1 (~0.94), establishing a new fine-grained benchmark.
  • Nes2Net-LA outperforms the prior Nes2Net and AASIST back-ends under identical training conditions and without data augmentation.
  • The corpus provides a broader diversity of audio quality (as measured by Scoreq) than earlier datasets, which the paper argues reduces overfitting to narrow acoustic conditions.

Where Pith is reading between the lines

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

  • The 'unseen API' evaluation may be less about novel synthesis than the paper claims: the t-SNE plot shows unseen API embeddings overlapping seen ones, so the improvement on the unseen split could partly reflect acoustic proximity rather than true invariance. A direct test would measure nearest-neighbor distances between seen and unseen API embeddings.
  • Because each API group is a single label, the tracing task rewards models that memorize per-API artifacts (e.g., codec or prosody). An adversarial training objective that pushes representations to be invariant across API groups would be a natural extension and could improve the low unseen recall (~0.52).
  • The 30-API design makes the corpus a strong testbed for future zero-shot spoofing detection: if models trained on MultiAPI Spoof transfer to a brand-new TTS service, that would validate the dataset as a proxy for real-world deployment.
  • Re-encoding all spoofed audio with a common codec before training would isolate whether the gained generalization comes from genuine spoofing traces or from API-specific recording artifacts.

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

4 major / 5 minor

Summary. The paper introduces MultiAPI Spoof, a roughly 230-hour multi-API spoofed-speech dataset generated from 30 commercial, open-source, and web-based TTS/VC APIs, together with a held-out-API split for evaluating cross-source generalization. It also proposes Nes2Net-LA, a local-attention variant of the Nes2Net anti-spoofing backbone, and defines a fine-grained API tracing task. Experiments compare anti-spoofing models trained with and without the MultiAPI training set, report improvements on the MultiAPI test set and on the external ITW and AI4T benchmarks, and claim state-of-the-art performance. The dataset and code are publicly released.

Significance. If the central claims hold, MultiAPI Spoof would be a valuable addition to the anti-spoofing benchmark landscape, addressing a real gap between public-model-based training corpora and the diverse proprietary APIs used in practice. The API tracing task is a useful new evaluation direction, and releasing 230 hours of labeled multi-API speech is a concrete contribution. The paper also benefits from evaluating on an external in-the-wild benchmark (ITW) and from explicitly separating seen and unseen API conditions. However, the significance is currently limited by the absence of leakage analysis for the held-out APIs, the lack of a data-scale control in the training-set comparison, and the absence of any variance or significance information for the reported performance differences.

major comments (4)
  1. [Section 2.2, Table 1] The 'unseen API' evaluation assumes that APIs A24–A29 do not share accidental acoustic properties with the seen A0–A20, such as the same source speakers, identical text prompts, common codecs, or shared TTS backends. The paper provides no leakage analysis. Since Figure 3 shows unseen embeddings mixed with seen clusters, this concern is concrete. Without ruling out shared low-level cues, the reported 0.62% unseen EER may reflect recognition of familiar artifacts rather than robustness to genuinely novel APIs. Please provide per-API EERs, text/speaker-overlap statistics, and an acoustic-similarity analysis, or construct the split to control for these factors.
  2. [Section 5.2.1, Table 1] The comparison 'without vs. with MultiAPI Spoof' changes both the content diversity and the total training set size (adding roughly 230 hours). The improved performance on MultiAPI, ITW, and AI4T could be due to simply having more training data, not to multi-API diversity. To support the claim that the dataset's diversity is the cause, please add a control trained on the same additional hours drawn from a single API or a few APIs, or otherwise match training-set size. This is load-bearing for the paper's main contribution.
  3. [Section 5.2.2, Table 2] The 'state-of-the-art' claim is not supported by the reported comparison because the baselines in Table 2 were trained on different data collections (e.g., ASVspoof 2019 LA vs. Data Collection 2) and with different augmentation schemes (RawBoost, Sample Pruning, codec augmentation). Models trained on more or different data can outperform solely due to training conditions. Please either train all methods on the same Data Collection 2 without augmentation, or qualify the claim to 'under our training configuration.' Also, on AI4T, XLSR+Nes2Net-LA ties XLSR+Nes2Net at 5.64 EER, contradicting the 'across all benchmarks' phrasing.
  4. [Section 5.1 and all results tables] All reported metrics are single-run point estimates with no error bars, seeds, or significance tests. The claimed improvements are often small (e.g., ITW EER dropping from 1.70 to 1.42; Table 1), and without variance estimates it is impossible to tell whether these differences are meaningful. Please report the mean and standard deviation over at least three runs, or provide a statistical significance test. This applies to Tables 1, 2, and 3.
minor comments (5)
  1. [Section 4] The threshold for classifying an instance as 'unseen' in API tracing is mentioned but never described. How is this threshold selected? Is it tuned on the development set (A21–A23)? Please specify.
  2. [Section 2.2 vs. Figure 3] Section 2.2 says A21–A23 are development-only and A24–A29 are held out for evaluation, but Figure 3's caption says 'Unseen APIs are A24–A29.' Clarify which APIs constitute the unseen set in each experiment, and whether the dev unseen APIs are excluded from evaluation.
  3. [Section 5.2.1, Figure 2] The Scoreq distribution figure is described qualitatively. Please provide numerical statistics (e.g., mean, standard deviation, range) to support the claim that MultiAPI Spoof spans a broader quality distribution.
  4. [References] References [5] and [8] appear to be the same StableVC paper; please deduplicate. Also, several 2026 preprint citations may not be publicly verifiable; ensure they are indexed.
  5. [Throughout] There are minor typos, e.g., 'V oice' in the Introduction, and the table formatting in Table 1 is hard to parse because the AI4T column is not clearly separated from the MultiAPI Spoof columns. Please rework the table layout.

Circularity Check

0 steps flagged

No significant circularity: the dataset, held-out API evaluation, and external benchmarks make the central empirical claims self-contained.

full rationale

This is an empirical systems/dataset paper, not a derivation. The central claims are (1) that adding the MultiAPI Spoof training set improves cross-domain robustness and (2) that the proposed Nes2Net-LA variant achieves strong results. Neither claim is obtained by construction: the unseen evaluation uses APIs A24-A29 that are held out from training (Section 2.2), and the cross-domain evaluation also reports results on the external ITW and AI4T benchmarks (Table 1, Section 5.2.1). The model definition in Equations (1)-(2) is an architectural extension of the published Nes2Net [30], and no fitted parameter is renamed as a prediction. The only self-citation is Ming Li's co-authorship of the Nes2Net baseline [30], but that work is a published, externally reviewed architecture used as a comparison baseline rather than as the justification for the paper's central claim. Concerns about possible acoustic leakage between seen and unseen APIs, or about missing error bars, are validity/robustness risks rather than circularity: the paper does not define any outcome in terms of the input. The honest finding is therefore no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The paper relies on a standard empirical ML setup: pretrained XLSR features, cross-entropy training, and held-out splits. The central results depend on dataset design choices (API set, split, threshold) and hyperparameters (J=8, K=1) that are not independently justified or ablated.

free parameters (5)
  • Local attention window radius K = 1
    Set in Section 5.1; no ablation over K, so the claimed improvement from local attention could be specific to this choice.
  • Channel splits J = 8
    Inherited from Nes2Net-X; fixed without ablation in this paper (Section 5.1).
  • Utterance length for evaluation = 4 seconds
    Section 5.1: samples are repeated/truncated to 4 s; this preprocessing choice can affect EER and is not analyzed.
  • API tracing unseen threshold = not reported
    Section 4: samples below a max-probability threshold are labeled unseen; threshold value and selection procedure are not specified, yet unseen precision/recall depend on it.
  • Optimizer hyperparameters = lr 1e-6/5e-6/1e-5, wd 1e-4
    Section 5.1; standard but fixed without sensitivity analysis.
axioms (5)
  • domain assumption XLSR-300M representations contain sufficient cues for anti-spoofing.
    All systems use XLSR-300M as feature extractor; if these features do not generalize to unseen API conditions, results are limited.
  • domain assumption CommonVoice speech is a valid bona fide distribution for English.
    Section 2: bona fide audio is taken from CommonVoice; mismatch with real attack scenarios is not analyzed.
  • domain assumption API labels A0-A29 are correct and each API is a homogeneous source.
    Section 2.1: dataset assumes per-API labels; mislabeled or internally diverse APIs would weaken tracing results.
  • domain assumption No data augmentation gives a clean comparison.
    Section 5.1: no augmentation applied; augmentation might change relative rankings.
  • standard math Cross-entropy loss and Adam optimization are sufficient for these tasks.
    Section 5.1: standard training setup, not the paper's contribution.
invented entities (2)
  • MultiAPI Spoof dataset independent evidence
    purpose: 230h spoofed audio from 30 APIs plus CommonVoice bona fide, for anti-spoofing training/eval and API tracing.
    Publicly released; independent researchers can download and re-benchmark.
  • Unseen-API class in tracing no independent evidence
    purpose: Aggregates all non-training APIs into a 22nd class via a max-probability threshold, to measure zero-shot source attribution.
    The class definition depends on the paper's threshold design; no external benchmark yet exists.

pith-pipeline@v1.3.0-alltime-deepseek · 10226 in / 13074 out tokens · 111835 ms · 2026-08-03T17:57:01.774090+00:00 · methodology

0 comments
read the original abstract

Existing speech anti-spoofing benchmarks rely on a narrow set of public models, creating a substantial gap from real-world scenarios in which commercial systems employ diverse, often proprietary APIs. To address this issue, we introduce MultiAPI Spoof, a multi-API audio anti-spoofing dataset comprising about 230 hours of synthetic speech generated by 30 distinct APIs, including commercial services, open-source models, and online platforms. Furthermore, we propose Nes2Net-LA, a local-attention enhanced variant of Nes2Net that improves local context modeling and fine-grained spoofing feature extraction. Based on this dataset, we also define the API tracing task, enabling fine-grained attribution of spoofed audio to its generation source. Experiments show that Nes2Net-LA achieves state-of-the-art performance and offers superior robustness, particularly under diverse and unseen spoofing conditions. Code \footnote{https://github.com/XuepingZhang/MultiAPI-Spoof} and dataset \footnote{https://xuepingzhang.github.io/MultiAPI-Spoof-Dataset/} have been released.

Figures

Figures reproduced from arXiv: 2512.07352 by Linxi Li, Liwei Jin, Ming Li, Xueping Zhang, Yechen Wang, Zhenshan Zhang.

Figure 1
Figure 1. Figure 1: Overall architecture of proposed Nes2Net-LA frame￾works. The model first extracts high-dimensional representa￾tions from the input audio and then processes them using nested multi-scale feature fusion. Nes2Net-LA further enhances cross￾block interactions through a sliding-window local attention mechanism. ‘WS’ represent Weighted Summation, and ‘ATT’ represent scaled dot-product self-attention. block output… view at source ↗
Figure 2
Figure 2. Figure 2: Scoreq [49] distribution comparison across datasets. The dashed vertical line in each curve marks the peak density value [PITH_FULL_IMAGE:figures/full_fig_p004_2.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

66 extracted references · 10 linked inside Pith

  1. [1]

    Introduction Recently, Text-To-Speech (TTS) [1, 2, 3, 4], V oice Conver- sion (VC) [5, 6, 7, 8], and generative modeling techniques [9, 10, 11, 12] have evolved rapidly. In particular, end-to-end dialogue systems [13, 14, 15, 16], speech continuation models [17, 18], and style- or emotion-specific speech generation mod- els [19, 20, 21] have advanced sign...

  2. [2]

    We show that there is a gap between previous research bench- marks and real-world spoofing scenarios; adding our API dataset in the training can also enhance the performance on current benchmarks

  3. [3]

    Furthermore, we propose a new anti-spoofing detection method, namely Nes2Net-LA, built upon Nes2Net [30]. By integrating local attention modules between Nested blocks, Nes2Net-LA enhances local context modeling and fine- grained spoofing feature extraction, thereby improving ro- bustness and discriminative capability. The Nes2Net-LA achieves state-of-the-...

  4. [4]

    Finally, we introduce the API tracing task, which aims to identify the generation API of spoofed audio and establishes a benchmark for fine-grained source attribution

  5. [5]

    It contains approximately 230 hours of spoofed audio and an equal amount of bona fide speech from CommonV oice, maintaining a 1:1 balance between the two

    MultiAPI Spoof Dataset MultiAPI Spoof is a new multi-API audio anti-spoofing dataset designed to bridge the gap between research benchmarks and real-world synthetic speech. It contains approximately 230 hours of spoofed audio and an equal amount of bona fide speech from CommonV oice, maintaining a 1:1 balance between the two. All recordings are in English...

  6. [6]

    Commercial TTS APIs: Speech synthesized by proprietary text-to-speech services widely used in industry

  7. [7]

    Open-Source Models: Speech generated using publicly avail- arXiv:2512.07352v3 [cs.SD] 5 Mar 2026 able neural TTS or voice conversion systems

    Pith/arXiv arXiv 2026

  8. [8]

    Each API corresponds to one labeled group (A0–A29), forming a comprehensive representation of modern TTS and generative pipelines

    TTS Websites: Audio collected from online platforms pro- viding web-based speech synthesis interfaces. Each API corresponds to one labeled group (A0–A29), forming a comprehensive representation of modern TTS and generative pipelines. 2.2. Dataset Split The MultiAPI Spoof dataset is partitioned by the APIs. APIs A0–A20 are used to construct the training, d...

  9. [9]

    Seen evaluation, where systems are tested on spoofed sam- ples generated from APIs that also appear in training (A0– A20)

  10. [10]

    Unseen evaluation, where systems are evaluated on spoofed samples from completely unseen APIs (A21–A29), allowing assessment of cross-source generalization

  11. [11]

    Prior Knowledge: Nested Res2Net (Nes2Net-X) The Nes2Net-X architecture [30] is a multi-scale feature extrac- tor for high-dimensional speech representations

    Local Attention Enhanced Anti-spoofing Network 3.1. Prior Knowledge: Nested Res2Net (Nes2Net-X) The Nes2Net-X architecture [30] is a multi-scale feature extrac- tor for high-dimensional speech representations. An audio seg- mentx i is encoded intox ′ i ∈R C×T ′ and split into channel- wise subsetsx i,1, . . . , xi,J. Each subset is processed hierarchi- ca...

  12. [12]

    Unlike conventional anti-spoofing, which only distinguishes bona fide and spoofed speech, API tracing provides fine-grained attribution

    Anti-Spoofing API Tracing Task The anti-spoofing API tracing task aims to identify which API generated a given spoofed audio sample. Unlike conventional anti-spoofing, which only distinguishes bona fide and spoofed speech, API tracing provides fine-grained attribution. APIs are divided into seen and unseen sets. The seen set, consisting of 21 APIs (A0–A20...

  13. [13]

    SP” denotes Sample Pruning [42]; “RB

    Experiments 5.1. Experimental Setup DatasetThe anti-spoofing experiments are conducted on a col- lection of six public datasets: TIMIT [39], ODSS [40], FoR [41], AI4T [42], ASV5 [35], and MLAAD [43]. These cor- pora cover a wide variety of spoofing sources, including real- world collected data, text-to-speech (TTS), and voice conver- sion (VC). We conside...

    2019

  14. [14]

    Experiments show that in- corporating MultiAPI Spoof into training significantly improves cross-domain robustness

    Conclusion In this paper, we present MultiAPI Spoof, a multi-API speech anti-spoofing dataset, and further introduce the API tracing task for fine-grained source attribution. Experiments show that in- corporating MultiAPI Spoof into training significantly improves cross-domain robustness. We also propose a local-attention enhanced anti-spoofing network, n...

  15. [15]

    The LLMs were not used for ideation, methodology, experimental design, data analysis, or result interpretation

    Generative AI Use Disclosure Large LanguageModels (LLMs) were used solely for manuscript polishing (e.g., rephrasing and grammar checks) to improve clarity and readability. The LLMs were not used for ideation, methodology, experimental design, data analysis, or result interpretation. All scientific content was produced and verified by the authors

  16. [16]

    Qwen3-tts technical report,

    H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guoet al., “Qwen3-tts technical report,” arXiv preprint arXiv:2601.15621, 2026

    Pith/arXiv arXiv 2026

  17. [17]

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

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. JianZhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, vol. 1, 2025, pp. 6255–6271

    2025

  18. [18]

    Restyle-tts: Relative and continuous style control for zero-shot speech synthe- sis,

    H. Li, C. Jin, C. Li, W. Guan, Z. Huang, and X. Chen, “Restyle-tts: Relative and continuous style control for zero-shot speech synthe- sis,”arXiv preprint arXiv:2601.03632, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    E1 tts: Simple and fast non-autoregressive tts,

    Z. Liu, S. Wang, P. Zhu, M. Bi, and H. Li, “E1 tts: Simple and fast non-autoregressive tts,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1– 5

    2025

  20. [21]

    Adaptvc: High quality voice conversion with adaptive learning,

    J. Kim, J.-H. Kim, Y . Choi, T. D. Nguyen, S. Mun, and J. S. Chung, “Adaptvc: High quality voice conversion with adaptive learning,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  21. [22]

    Onevoice: One model, triple scenarios-towards unified zero-shot voice con- version,

    Z. Wang, T. Li, W. Ge, Z. Cui, S. Zhang, and J. Feng, “Onevoice: One model, triple scenarios-towards unified zero-shot voice con- version,”arXiv preprint arXiv:2601.18094, 2026

    Pith/arXiv arXiv 2026

  22. [23]

    Stablevc: Style controllable zero-shot voice conversion with con- ditional flow matching,

    J. Yao, Y . Yuguang, Y . Pan, Z. Ning, J. Ye, H. Zhou, and L. Xie, “Stablevc: Style controllable zero-shot voice conversion with con- ditional flow matching,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 24, 2025, pp. 25 669–25 677

    2025

  23. [24]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Linet al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    Pith/arXiv arXiv 2024

  24. [25]

    Unifying speech recognition, synthesis and conversion with autoregressive transformers,

    R. Cai, Y . Lin, Y . Wang, C. Fu, and X. Zeng, “Unifying speech recognition, synthesis and conversion with autoregressive transformers,” 2026. [Online]. Available: https://arxiv.org/abs/ 2601.10770

    arXiv 2026

  25. [26]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  26. [27]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabil- ities,

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen et al., “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabil- ities,”arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  27. [28]

    Fun-audio-chat technical report,

    T. F. Team, Q. Chen, L. Cheng, C. Deng, X. Li, J. Liu, C.-H. Tan, W. Wang, J. Xu, J. Yeet al., “Fun-audio-chat technical report,” arXiv preprint arXiv:2512.20156, 2025

    arXiv 2025

  28. [29]

    Qwen3-omni technical report,

    J. Xu, Z. Guo, H. Hu, Y . Chu, X. Wang, J. He, Y . Wang, X. Shi, T. He, X. Zhuet al., “Qwen3-omni technical report,”arXiv preprint arXiv:2509.17765, 2025

    Pith/arXiv arXiv 2025

  29. [30]

    Wavbench: Benchmarking reasoning, colloquial- ism, and paralinguistics for end-to-end spoken dialogue models,

    Y . Li, S. Ji, Y . Chen, T. Liang, H. Ying, Y . Wang, J. Li, J. Fang, and Z. Zhao, “Wavbench: Benchmarking reasoning, colloquial- ism, and paralinguistics for end-to-end spoken dialogue models,” arXiv preprint arXiv:2602.12135, 2026

    arXiv 2026

  30. [31]

    Baichuan-omni-1.5 technical report,

    Y . Li, J. Liu, T. Zhang, S. Chen, T. Li, Z. Li, L. Liu, L. Ming, G. Dong, D. Panet al., “Baichuan-omni-1.5 technical report,” arXiv preprint arXiv:2501.15368, 2025

    arXiv 2025

  31. [32]

    Baichuan-audio: A unified framework for end-to-end speech interaction,

    T. Li, J. Liu, T. Zhang, Y . Fang, D. Pan, M. Wang, Z. Liang, Z. Li, M. Lin, G. Donget al., “Baichuan-audio: A unified framework for end-to-end speech interaction,”arXiv preprint arXiv:2502.17239, 2025

    Pith/arXiv arXiv 2025

  32. [33]

    Mimo-audio: Audio language models are few-shot learners,

    L.-C.-T. Xiaomi, “Mimo-audio: Audio language models are few-shot learners,” 2025. [Online]. Available: https: //github.com/XiaomiMiMo/MiMo-Audio

    2025

  33. [34]

    Soulx- podcast: Towards realistic long-form podcasts with dialectal and paralinguistic diversity,

    H. Xie, H. Lin, W. Cao, D. Guo, W. Tian, J. Wu, H. Wen, R. Shang, H. Liu, Z. Jiang, Y . Jiang, W. Chen, R. Yan, J. Qian, Y . Yan, S. Yin, M. Tao, X. Chen, L. Xie, and X. Wang, “Soulx- podcast: Towards realistic long-form podcasts with dialectal and paralinguistic diversity,”arXiv preprint arXiv:2510.23541, 2025

    arXiv 2025

  34. [35]

    Flex- ivoice: Enabling flexible style control in zero-shot tts with natural language instructions,

    D. Chen, X. Zhang, Y . Wang, K. Dai, L. Ma, and Z. Wu, “Flex- ivoice: Enabling flexible style control in zero-shot tts with natural language instructions,”arXiv preprint arXiv:2601.04656, 2026

    arXiv 2026

  35. [36]

    Step-audio- editx technical report,

    C. Yan, B. Wu, P. Yang, P. Tan, G. Hu, Y . Zhang, Xiangyu, Zhang, F. Tian, X. Yang, X. Zhang, D. Jiang, and G. Yu, “Step-audio- editx technical report,”arXiv preprint arXiv:2511.03601, 2025

    arXiv 2025

  36. [37]

    Tango 2: Aligning diffusion-based text-to- audio generative models through direct preference optimization,

    N. Majumder, C.-Y . Hung, D. Ghosal, W.-N. Hsu, R. Mihal- cea, and S. Poria, “Tango 2: Aligning diffusion-based text-to- audio generative models through direct preference optimization,” inACM Multimedia, 2024

    2024

  37. [38]

    Kimi-audio technical report,

    D. Ding, Z. Ju, Y . Leng, S. Liu, T. Liu, Z. Shang, K. Shen, W. Song, X. Tan, H. Tanget al., “Kimi-audio technical report,” arXiv preprint arXiv:2504.18425, 2025

    Pith/arXiv arXiv 2025

  38. [39]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  39. [40]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  40. [41]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  41. [42]

    Res2net: A new multi-scale backbone architecture,

    S.-H. Gao, M.-M. Cheng, K. Zhao, X.-Y . Zhang, M.-H. Yang, and P. Torr, “Res2net: A new multi-scale backbone architecture,” IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 2, pp. 652–662, 2019

    2019

  42. [43]

    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, pp. 6367–6371

    2022

  43. [44]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst conference on language modeling

  44. [45]

    Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,

    T. Liu, D.-T. Truong, R. K. Das, K. A. Lee, and H. Li, “Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,”IEEE Transactions on Information Foren- sics and Security, vol. 20, pp. 12 005–12 018, 2025

    2025

  45. [46]

    The partialspoof database and countermeasures for the detec- tion of short fake speech segments embedded in an utterance,

    L. Zhang, X. Wang, E. Cooper, N. Evans, and J. Yamagishi, “The partialspoof database and countermeasures for the detec- tion of short fake speech segments embedded in an utterance,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 813–825, 2022

    2022

  46. [47]

    The database and benchmark for the source speaker tracing challenge 2024,

    Z. Li, Y . Lin, T. Yao, H. Suo, P. Zhang, Y . Ren, Z. Cai, H. Nishizaki, and M. Li, “The database and benchmark for the source speaker tracing challenge 2024,” inIEEE Spoken Lan- guage Technology Workshop (SLT), 2024, pp. 1254–1261

    2024

  47. [48]

    Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,

    A. Nautsch, X. Wang, N. Evans, T. H. Kinnunen, V . Vestman, M. Todisco, H. Delgado, M. Sahidullah, J. Yamagishi, and K. A. Lee, “Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,”IEEE Transac- tions on Biometrics, Behavior, and Identity Science, vol. 3, no. 2, pp. 252–265, 2021

    2019

  48. [49]

    Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kin- nunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautschet al., “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 2507–2522, 2023

    2021

  49. [50]

    Asvspoof 5: crowdsourced speech data, deepfakes, and adversar- ial attacks at scale,

    X. Wang, H. Delgado, H. Tak, J.-w. Jung, H.-j. Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. H. Kinnunenet al., “Asvspoof 5: crowdsourced speech data, deepfakes, and adversar- ial attacks at scale,” inProceedings of ASVspoof, 2024, pp. 1–8

    2024

  50. [51]

    Codecfake: Enhancing anti-spoofing models against deepfake audios from codec-based speech synthesis systems,

    H. Wu, Y . Tseng, and H.-y. Lee, “Codecfake: Enhancing anti-spoofing models against deepfake audios from codec-based speech synthesis systems,” inProceedings of Interspeech, 2024, pp. 1770–1774

    2024

  51. [52]

    Automatic speaker verification spoofing and deepfake de- tection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. W. Evans, “Automatic speaker verification spoofing and deepfake de- tection using wav2vec 2.0 and data augmentation,” inProceedings of Odyssey, 2022

    2022

  52. [53]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  53. [54]

    Timit-tts: A text-to-speech dataset for multimodal synthetic me- dia detection,

    D. Salvi, B. Hosler, P. Bestagini, M. C. Stamm, and S. Tubaro, “Timit-tts: A text-to-speech dataset for multimodal synthetic me- dia detection,”IEEE access, vol. 11, pp. 50 851–50 866, 2023

    2023

  54. [55]

    An open dataset of synthetic speech,

    A. Yaroshchuk, C. Papastergiopoulos, L. Cuccovillo, P. Aichroth, K. V otis, and D. Tzovaras, “An open dataset of synthetic speech,” inIEEE International Workshop on Information Forensics and Se- curity (WIFS), 2023, pp. 1–6

    2023

  55. [56]

    For: A dataset for synthetic speech detection,

    R. Reimao and V . Tzerpos, “For: A dataset for synthetic speech detection,” inIEEE International Conference on Speech Technol- ogy and Human-Computer Dialogue (SpeD), 2019, pp. 1–10

    2019

  56. [57]

    Un- masking real-world audio deepfakes: A data-centric approach,

    D. Combei, A. Stan, D. Oneata, N. M ¨uller, and H. Cucu, “Un- masking real-world audio deepfakes: A data-centric approach,” in Proceedings of Interspeech, 2025, pp. 5343–5347

    2025

  57. [58]

    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,” inIEEE International Joint Conference on Neural Networks (IJCNN), 2024, pp. 1–7

    2024

  58. [59]

    Does audio deepfake detection generalize?

    N. M ¨uller, P. Czempin, F. Diekmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?”Pro- ceedings of Interspeech, pp. 2783–2787, 2022

    2022

  59. [60]

    Adam: a method for stochastic optimization,

    D. Kingma, “Adam: a method for stochastic optimization,” inIn- ternational Conference Learn Represent, 2014

    2014

  60. [61]

    Cross-entropy loss functions: Theoretical analysis and applications,

    A. Mao, M. Mohri, and Y . Zhong, “Cross-entropy loss functions: Theoretical analysis and applications,” inInternational confer- ence on Machine learning, 2023, pp. 23 803–23 828

    2023

  61. [62]

    Asvspoof 5 evaluation plan,

    H. Delgado, N. Evans, J.-w. Jung, T. Kinnunen, I. Kukanov, K. A. Lee, X. Liu, H.-j. Shim, M. Sahidullah, H. Taket al., “Asvspoof 5 evaluation plan,” https://www.asvspoof.org/file/ASVspoof5 Evaluation Plan Phase2.pdf, 2024, [Online]

    2024

  62. [63]

    A review of the f- measure: its history, properties, criticism, and alternatives,

    P. Christen, D. J. Hand, and N. Kirielle, “A review of the f- measure: its history, properties, criticism, and alternatives,”ACM Computing Surveys, vol. 56, no. 3, pp. 1–24, 2023

    2023

  63. [64]

    Scoreq: Speech qual- ity assessment with contrastive regression,

    A. Ragano, J. Skoglund, and A. Hines, “Scoreq: Speech qual- ity assessment with contrastive regression,”Advances in Neural Information Processing Systems, vol. 37, pp. 105 702–105 729, 2024

    2024

  64. [65]

    Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,

    H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” inIEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6382–6386

    2022

  65. [66]

    Audio deepfake detection with self-supervised xls-r and sls classifier,

    Q. Zhang, S. Wen, and T. Hu, “Audio deepfake detection with self-supervised xls-r and sls classifier,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 6765– 6773

    2024

  66. [67]

    Visualizing data using t-sne,

    L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of machine learning research, vol. 9, no. Nov, pp. 2579– 2605, 2008

    2008