Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

The paper introduces Echoes, a music deepfake dataset where each AI-generated track is semantically matched to a real song, and shows that detectors trained on it generalize better across datasets than those trained on any current benchmark

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

2026-08-02 17:31 UTC pith:ZTM7SKQX

load-bearing objection Echoes is a genuinely useful new music-deepfake dataset with a solid baseline protocol, but the paper's central claim that semantic alignment drives the gains is not actually validated — the dataset is still worth publishing after fixes. the 4 major comments →

arxiv 2603.23667 v3 pith:ZTM7SKQX submitted 2026-03-24 cs.SD cs.AIeess.AS

Echoes: A semantically-aligned music deepfake detection dataset

Octavian Pascu , Dan Oneata , Horia Cucu , Nicolas M. Muller This is my paper
classification cs.SD cs.AIeess.AS
keywords audio deepfake detectionAI-generated musicdatasetsemantic alignmentcross-dataset generalizationshortcut learningprovider diversityWav2Vec2
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.

This paper introduces Echoes, a dataset of 3,577 AI-generated music tracks (110 hours of audio) built from 310 bona fide songs and ten different generators, and argues that it fixes a gap in existing benchmarks: unlike clip-based or low-diversity collections, Echoes pairs each fake track with a real song of matching style and theme. On the authors' own SSL-plus-logistic-regression baseline, Echoes is the hardest in-domain dataset, with an EER of 9.36% versus 2.06-8.61% for AIME, SONICS, and FakeMusicCaps, yet the model trained on Echoes achieves the best average out-of-domain EER (21.02%) when transferred to the other three. If the results hold, Echoes is simultaneously a stiffer test for detectors and a better training ground, because its semantic alignment removes content-based shortcuts while its provider diversity covers more generation artifacts.

Core claim

The central claim is that dataset construction—semantic alignment plus provider diversity—drives both benchmark difficulty and transferability. Echoes is built by generating, for each bona fide track, LLM-derived stylistic descriptions (from title and genre) that condition text-to-audio generation, with reference-audio conditioning for models that support it. In cross-dataset experiments with a frozen Wav2Vec2 XLS-R 2B encoder and a linear classifier, Echoes is the hardest to classify in-domain (9.36% EER), and training on Echoes yields the lowest average out-of-domain EER (21.02%) versus 21.36% (AIME), 28.65% (SONICS), and 37.10% (FakeMusicCaps). The authors interpret the asymmetry—existing

What carries the argument

The semantic-alignment pipeline: for each bona fide track, a large language model (ChatGPT) turns the track's title and genre into a short comma-separated list of stylistic characteristics; that descriptor conditions the text-to-audio generators, and the original waveform is additionally passed as reference for audio-to-audio providers. This is the mechanism that equalizes content (style, genre, instrumentation, tempo) between real and fake tracks, so a detector that exploits only content differences cannot succeed. Provider diversity complements it: the ten generators (e.g., Suno, Udio, AudioLDM, Stable Audio, DiffRhythm) yield distinct architectures and artifacts, so the trained detector m

Load-bearing premise

The entire claim that Echoes is both harder and more transferable depends on the assumption that LLM-generated descriptors derived from title and genre successfully align each fake track's content with its bona fide reference; if the descriptors capture only coarse metadata, fake tracks may still differ systematically in content, letting detectors exploit content cues rather than generation artifacts.

What would settle it

Generate a control subset of Echoes where the descriptor conditioning is deliberately mismatched (each fake track is produced from a descriptor of a different genre or track). If in-domain EER on that subset drops substantially below the 9.36% of the aligned set, semantic alignment is confirmed as the cause of difficulty; if the EER stays near 9%, alignment is not doing the work claimed.

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

If this is right

  • A detector trained on AIME, SONICS, or FakeMusicCaps degrades sharply on Echoes (28.6-41.7% EER), so reported in-domain numbers on older benchmarks likely overstate deployed performance.
  • Training on Echoes gives the best average transfer (21.02% EER) to the other three datasets, suggesting it is a stronger default training corpus for general-purpose music deepfake detectors.
  • Per-provider accuracy on Echoes spans 87-94%, so benchmark results should be disaggregated by generator rather than pooled.
  • Text-to-audio and audio-to-audio generated segments are nearly equally detectable (90.6% vs 90.1% accuracy at the EER threshold), indicating LLM-descriptor alignment is on par with waveform conditioning.
  • Echoes' full-track structure (up to 480 s, median 106 s) supports future sequence-level detection that clip-based datasets cannot.

Where Pith is reading between the lines

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

  • A direct test of the alignment mechanism: regenerate Echoes fakes with descriptors randomly reassigned across tracks; if in-domain EER falls well below 9.36%, semantic alignment is the active ingredient, while a similar EER would mean the dataset's difficulty comes from provider diversity or generator artifacts instead.
  • The 9.36% EER could serve as a lower bound for current commercial generator detectability under content-matched conditions, a number useful for platform moderation capacity planning.
  • If descriptors leak any source-specific information (e.g., a distinctive lyric theme that identifies the reference track), fakes could be matched to their bona fide source, inflating transfer scores; an oracle probing for such leakage would clarify the result.
  • Future work could extend the alignment protocol to partial/mixed content (AI vocals over real accompaniment), the harder setting named in the paper's conclusion.

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 Echoes, a dataset for music deepfake detection built from Free Music Archive bona fide tracks. For each reference track, an LLM (ChatGPT-5.0 Thinking) generates stylistic descriptors from title and genre, and ten music-generation providers produce fake tracks; audio-to-audio conditioning is used where supported. The full text reports 3,577 generated tracks / 110 hours (the abstract states 4,468 tracks / 131 hours). The baseline is frozen Wav2Vec2 XLS-R 2B embeddings plus logistic regression on 10-second segments, evaluated with EER. In-domain results give Echoes 9.36% EER vs AIME 6.40%, SONICS 2.06%, FakeMusicCaps 8.61%. Cross-dataset results show that models trained on existing datasets transfer poorly to Echoes (28.6–41.7% EER), while training on Echoes yields the best average out-of-domain EER (21.02%). The paper attributes this to provider diversity and semantic alignment.

Significance. If the dataset details are corrected and the causal claims are supported, Echoes would be a useful public benchmark: it covers ten providers, includes both short and long synthetic tracks, uses a reproducible SSL+LR baseline, performs track-level splitting to avoid leakage, and releases the data under an MIT license. The per-provider and per-modality analyses are also valuable. However, the current evidence for semantic alignment is indirect, the cross-dataset comparison is confounded by multiple dataset attributes, and the headline transfer advantage is small relative to the apparent experimental noise. These issues need to be resolved before the central claims can be accepted.

major comments (4)
  1. [Abstract, §1, §2.3, Table 1] The abstract states that Echoes contains 4,468 tracks and 131 hours of audio, while the full text consistently reports 3,577 tracks and 110 hours (e.g., §1, §2.3, Table 1, Table 2, §4). This is not a formatting nit: the dataset size is a core factual claim and the discrepancy prevents verification of the resource description. Please reconcile the numbers and ensure all instances match.
  2. [§2.1, §3.3, §4] Semantic alignment is asserted but never directly validated. The generation pipeline conditions on LLM descriptors derived from title and genre, and on reference audio where supported, but no check is reported that a generated track actually resembles its paired bona fide track in content (melody, instrumentation, tempo, lyrics, production style). The only evidence offered in §3.3 is that TTA and ATA subsets have similar per-segment accuracy at the EER threshold; this compares two conditioning modalities and does not measure content similarity. Without a direct alignment check, the abstract's causal sentence—'These findings suggest that provider diversity and semantic alignment help learn more transferable detection cues'—is unsupported.
  3. [Table 3, §3.2, §4] The cross-dataset comparison is confounded. Echoes differs from AIME, SONICS, and FakeMusicCaps in total hours, provider count, track duration distributions, and source corpus simultaneously. The observed transfer advantage of training on Echoes cannot be attributed specifically to 'provider diversity and semantic alignment' as claimed in the Abstract and Conclusion. Controlled ablations or subset-matching experiments (e.g., matching provider count, total duration, or segment content) are needed, or the causal claim should be substantially softened.
  4. [Table 3, §3.2] No error bars, confidence intervals, or multiple-seed experiments are reported for any EER value. The headline 'training on Echoes yields the strongest generalization performance' rests on a 0.34-percentage-point gap (21.02% vs 21.36% for AIME), which may be within experimental noise. Additionally, the FakeMusicCaps row lists an out-of-domain average of 37.10, but (31.03 + 30.16 + 41.66) / 3 = 34.28; the average is arithmetically incorrect. Please provide repeated runs with seeds and variability estimates, and correct the table.
minor comments (5)
  1. [§1] Typo: 'post-processinw' should be 'post-processing'.
  2. [Table 4] The header contains 'ACEStep Producer'; the provider name elsewhere is 'AceStep' (Table 2). 'Producer' appears to be a stray label. Also, the accuracy figures at the EER threshold would benefit from being tied to the threshold definition (0.91) with a brief explanation.
  3. [§2.3, §4] The dataset size is typeset inconsistently as '3 577' and '3577'; choose one format throughout.
  4. [§3.3] The terminology jumps between 'high-level intent (style/theme)' in §4 and 'semantic-level alignment' in the Abstract. These are not equivalent; please align the terminology or define what 'semantic' means operationally.
  5. [Table 3] The note that 'higher is better' for in-domain EER expresses a dataset-centric desirability judgment, not a model-quality measure. This should be clearly framed as an interpretation, not a property of the metric.

Circularity Check

0 steps flagged

No circular derivation: Echoes results are empirical benchmarks; self-citations are motivational only.

full rationale

The paper's core claims are empirical measurements obtained with a fixed protocol: identical frozen Wav2Vec2 XLS-R embeddings and a logistic-regression classifier are applied to all datasets, and no parameter is fitted to the reported EER outcomes. The dataset construction (LLM descriptors + reference-audio conditioning) is described independently of the evaluation results, and the in-domain/out-of-domain comparisons do not reduce to the construction by definition. The weak point flagged by the skeptic—that semantic alignment is asserted rather than directly validated—is a validity concern, not circularity: even if the causal attribution 'provider diversity and semantic alignment help learn more transferable detection cues' is under-supported, the EER numbers themselves are not derived from that assumption. The paper cites the authors' own prior work ([6], [13]) for motivation (SSL representation choice and shortcut-avoidance in datasets), but these citations are not load-bearing for the central transfer comparisons, which would stand unchanged if those citations were removed. There is an internal inconsistency in dataset scale (abstract 4,468 tracks/131 h vs. body 3,577 tracks/110 h), but that is an editing error, not a circular step. Overall, no prediction reduces by construction to a fitted input or to a self-citation chain; the score reflects only the presence of minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The comparison is an empirical benchmark, so there are no fitted parameters in the mathematical sense. The free parameters listed are protocol hyperparameters (segment length, regularization C). The claims rely on domain assumptions about descriptor fidelity, the fairness of the frozen XLS-R baseline, and the statistical adequacy of a single unseeded split. No invented entities are introduced.

free parameters (2)
  • Logistic regression regularization strength C = 1e6
    Chosen by hand in Section 3.1; fixed across all datasets, so it is a protocol constant rather than fitted to Echoes, but it is a tunable hyperparameter that can affect EER comparisons.
  • 10-second segment length = 10 s
    Chosen in Section 3.1 to standardize across clip-based datasets; a design choice that could influence which cues are available and therefore the difficulty ranking.
axioms (4)
  • domain assumption LLM-generated textual descriptions preserve enough stylistic information to make fake tracks semantically aligned with bona fide tracks.
    Section 2.1: descriptors from ChatGPT-5.0 Thinking are used to condition generation; if descriptors capture only coarse metadata, content-level shortcuts remain and the central rationale for Echoes is weakened.
  • domain assumption Wav2Vec2 XLS-R 2B frozen embeddings plus logistic regression is a fair common baseline for comparing dataset difficulty and transfer.
    Section 3.1: the same protocol is applied to all datasets; if this representation is particularly suited/unsuited to some providers, the EER ranking may reflect model fit rather than dataset properties.
  • domain assumption Songs from the Free Music Archive (310 CC0/CC-BY/public-domain tracks) are representative bona fide music for pop/rock/electronic genres.
    Section 2.1: dataset constructed from these tracks; distributional biases in FMA would propagate into both real and fake sides.
  • domain assumption The 80:20 random split at track level is statistically adequate without multiple seeds.
    Section 3.1: no seed or repeated splits are reported, so EER differences lack uncertainty estimates.

pith-pipeline@v1.3.0-alltime-deepseek · 7651 in / 13578 out tokens · 129134 ms · 2026-08-02T17:31:39.703311+00:00 · methodology

0 comments
read the original abstract

We introduce Echoes, a new dataset for music deepfake detection designed for training and benchmarking detectors under realistic and provider-diverse conditions. Echoes comprises 4,468 tracks (131 hours of audio) spanning multiple genres (pop, rock, electronic), and includes content generated by ten popular AI music generation systems. To prevent shortcut learning and promote robust generalization, the dataset is deliberately constructed to be challenging, enforcing semantic-level alignment between spoofed audio and bona fide references. This alignment is achieved by conditioning generated audio samples directly on bona-fide waveforms or song descriptors. We evaluate Echoes in a cross-dataset setting against three existing AI-generated music datasets using state-of-the-art Wav2Vec2 XLS-R 2B representations. Results show that (i) Echoes is the hardest in-domain dataset; (ii) detectors trained on existing datasets transfer poorly to Echoes; (iii) training on Echoes yields the strongest generalization performance. These findings suggest that provider diversity and semantic alignment help learn more transferable detection cues.

Figures

Figures reproduced from arXiv: 2603.23667 by Dan Oneata, Horia Cucu, Nicolas M. Muller, Octavian Pascu.

Figure 1
Figure 1. Figure 1: Word cloud of generated song descriptions. The de￾scriptions are produced by a large language model from songs’ metadata and are used to condition music generation in our dataset, Echoes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Track duration distribution in Echoes. 3.1. Training setup Data processing. All audio files are split into non￾overlapping 10-second segments. This design choice is mo￾tivated by several reasons. First, AIME and FakeMusicCaps are clip-based and use 10-second excerpts; segmenting all cor￾pora to 10-second clips ensures a fair and standardized pro￾tocol across datasets. Second, 10-second segments provide a s… view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE visualization of our proposed Echoes dataset. 3.3. Per-provider performance on Echoes To better understand what drives the in-domain difficulty of Echoes, we also report per-model accuracy on the Echoes test split using the operating threshold 0.91 obtained through the EER (9.36%) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Finding the noise: Zero-shot AI Music Detection

    cs.SD 2026-07 conditional novelty 6.0

    A zero-shot method based on fakeprints, NMF and a blur-based reconstruction error detects unknown AI-music generators in one-class and clustering setups, working for most services but missing Mubert and pre-v9 Mureka.

Reference graph

Works this paper leans on

43 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Modern generators can pro- duce realistic music from short prompts, increasing the feasi- bility of large-scale synthetic music creation

    Introduction 1 Recent advances in music generation have enabled the syn- thesis of convincing songs that emulate genres, production aes- thetics, and vocal styles at scale. Modern generators can pro- duce realistic music from short prompts, increasing the feasi- bility of large-scale synthetic music creation. This shift cre- ates integrity and provenance ...

    2026

  2. [2]

    Generation pipeline To build our dataset, we start from bona fide music tracks from the Free Music Archive (FMA) [26]

    Dataset description 2.1. Generation pipeline To build our dataset, we start from bona fide music tracks from the Free Music Archive (FMA) [26]. We select 310 tracks li- censed under CC0, CC-BY or the public domain. For each bona fide track, we generate a corresponding fake samples using models from ten providers (Table 2). To obtain aligned samples that p...

  3. [3]

    Table 2:AI-music providers used to create Echoes

    Evaluation We provide initial results on our dataset, as well as cross- dataset generalization results using the three other major avail- able datasets (AIME, SONICS, FakeMusicCaps). Table 2:AI-music providers used to create Echoes. “Unknown” means the provider does not publicly specify the model. The two generator types are either audio-to-audio (A→A) or...

  4. [4]

    Echoes contains 3577tracks (110hours) across pop, rock, and electronic mu- sic genres

    Conclusion We presentedEchoes, a semantically-aligned dataset for AI- generated music detection built to support training and bench- marking under realistic provider diversity. Echoes contains 3577tracks (110hours) across pop, rock, and electronic mu- sic genres. For each bona fide song from the Free Music Archive, we generate AI-produced counterparts fro...

  5. [5]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,” inThe Speaker and Language Recognition Workshop (Odyssey 2022). ISCA, 2022

    2022

  6. [6]

    Towards generalisable and calibrated audio deepfake detection with self- supervised representations,

    O. Pascu, A. Stan, D. Oneata, E. Oneata, and H. Cucu, “Towards generalisable and calibrated audio deepfake detection with self- supervised representations,” inInterspeech, vol. 2024, 2024, pp. 4828–4832

    2024

  7. [7]

    Deezer: 28% of all delivered music is now fully AI-generated,

    Deezer, “Deezer: 28% of all delivered music is now fully AI-generated,” Deezer Newsroom (press release), Sep. 2025, accessed: 2026-01-11. [Online]. Available: https: //newsroom-deezer.com/2025/09/28-fully-ai-generated-music/

    2025

  8. [8]

    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,” in ICASSP, 2021

    2021

  9. [9]

    AASIST: Audio anti-spoofing using in- tegrated 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 in- tegrated spectro-temporal graph attention networks,” inICASSP, 2022

    2022

  10. [10]

    Singing voice graph modeling for singfake detection,

    X. Chen, H. Wu, R. Jang, and H.-y. Lee, “Singing voice graph modeling for singfake detection,” inProc. Interspeech 2024, 2024, pp. 4843–4847

    2024

  11. [11]

    Speech is silver, silence is golden: What do asvspoof-trained models really learn?

    N. M ¨uller, F. Dieckmann, P. Czempin, R. Canals, K. B ¨ottinger, and J. Williams, “Speech is silver, silence is golden: What do asvspoof-trained models really learn?” inProc. ASVSPOOF 2021, 2021, pp. 55–60

    2021

  12. [12]

    Is synthetic voice detection research going into the right direction?

    S. Borz `ı, O. Giudice, F. Stanco, and D. Allegra, “Is synthetic voice detection research going into the right direction?” inCVPRW, 2022

    2022

  13. [13]

    Detect- ing music deepfakes is easy but actually hard,

    D. Afchar, G. Meseguer-Brocal, and R. Hennequin, “Detect- ing music deepfakes is easy but actually hard,”arXiv preprint arXiv:2405.04181, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    From audio deepfake detection to AI-generated music detection–A pathway and overview,

    Y . Li, M. Milling, L. Specia, and B. W. Schuller, “From audio deepfake detection to AI-generated music detection–A pathway and overview,”arXiv preprint arXiv:2412.00571, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Asvspoof 2019: Future horizons in spoofed and fake audio detec- tion,

    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 detec- tion,” inInterspeech 2019. International Speech Communication Association, 2019, pp. 1008–1012

    2019

  16. [16]

    Does audio deepfake detection generalize?

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

    2022

  17. [17]

    Simple and controllable music gen- eration,

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y . Adi, and A. D ´efossez, “Simple and controllable music gen- eration,”Advances in Neural Information Processing Systems, vol. 36, pp. 47 704–47 720, 2023

    2023

  18. [18]

    Audioldm: Text-to-audio generation with la- tent diffusion models,

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. Plumbley, “Audioldm: Text-to-audio generation with la- tent diffusion models,” inProceedings of the 40th International Conference on Machine Learning, PMLR 2023, vol. 202. In- ternational Machine Learning Society (IMLS), 2023, pp. 21 450– 21 474

    2023

  19. [19]

    Circum- venting shortcuts in audio-visual deepfake detection datasets with unsupervised learning,

    S. Smeu, D.-A. Boldisor, D. Oneata, and E. Oneata, “Circum- venting shortcuts in audio-visual deepfake detection datasets with unsupervised learning,” inCVPR, 2025

    2025

  20. [20]

    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,” inICASSP, 2023

    2023

  21. [21]

    On the effec- tiveness of dataset alignment for fake image detection,

    A. S. Rajan, U. Ojha, J. Schloesser, and Y . J. Lee, “On the effec- tiveness of dataset alignment for fake image detection,” inICLR, 2025

    2025

  22. [22]

    MusicLM: Generating music from text,

    A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi et al., “MusicLM: Generating music from text,”arXiv preprint arXiv:2301.11325, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Benchmarking music generation models and metrics via human preference studies,

    F. Gr ¨otschla, A. Solak, L. A. Lanzend ¨orfer, and R. Wattenhofer, “Benchmarking music generation models and metrics via human preference studies,” inICASSP, 2025

    2025

  24. [24]

    supports detection and attribution but remains largely clip- oriented and does not contain long-form structure; SONICS

  25. [25]

    To address these gaps, we introduceEchoes, an evaluation- centric dataset for training and benchmarking AI-generated mu- sic detectors

    emphasizes long-duration counterfeit-song detection, yet covers only two popular providers. To address these gaps, we introduceEchoes, an evaluation- centric dataset for training and benchmarking AI-generated mu- sic detectors. Echoes is designed to be semantically aligned and covers a broad range of providers. The dataset contains3 577 tracks totaling 11...

    Pith/arXiv arXiv 2026

  26. [26]

    Stable audio open,

    Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, “Stable audio open,” inICASSP, 2025

    2025

  27. [27]

    The AI mu- sic arms race: On the detection of AI-generated music,

    L. Cros Vila, B. Sturm, L. Casini, and D. Dalmazzo, “The AI mu- sic arms race: On the detection of AI-generated music,”Transac- tions of the International Society for Music Information Retrieval, vol. 8, no. 1, pp. 179–194, 2025

    2025

  28. [28]

    Singfake: Singing voice deepfake detection,

    Y . Zang, Y . Zhang, M. Heydari, and Z. Duan, “Singfake: Singing voice deepfake detection,” inICASSP, 2024

    2024

  29. [29]

    SVDD 2024: The inaugural singing voice deepfake detection challenge,

    Y . Zhang, Y . Zang, J. Shi, R. Yamamoto, T. Toda, and Z. Duan, “SVDD 2024: The inaugural singing voice deepfake detection challenge,” inSLT, 2024

    2024

  30. [30]

    FakeMusicCaps: A dataset for detection and attribution of synthetic music generated via text-to-music models,

    L. Comanducci, P. Bestagini, and S. Tubaro, “FakeMusicCaps: A dataset for detection and attribution of synthetic music generated via text-to-music models,”Journal of Imaging, vol. 11, no. 7, p. 242, 2025

    2025

  31. [31]

    Sonics: Synthetic or not-identifying counterfeit songs,

    M. A. Rahman, Z. I. A. Hakim, N. H. Sarker, B. Paul, and S. A. Fattah, “Sonics: Synthetic or not-identifying counterfeit songs,” inThe Thirteenth International Conference on Learning Repre- sentations, 2024

    2024

  32. [32]

    Fma: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “Fma: A dataset for music analysis,” in18th International Society for Music Information Retrieval Conference, 2017

    2017

  33. [33]

    ACE-step: A step towards music generation foundation model,

    J. Gong, S. Zhao, S. Wang, S. Xu, and J. Guo, “ACE-step: A step towards music generation foundation model,”arXiv preprint arXiv:2506.00045, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Brev AI Music Generator,

    Brev, “Brev AI Music Generator,” Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://brev.ai/

    2026

  35. [35]

    DiffRhythm: Blazingly fast and embarrassingly sim- ple end-to-end full-length song generation with latent diffusion,

    Z. Ning, H. Chen, Y . Jiang, C. Hao, G. Ma, S. Wang, J. Yao, and L. Xie, “DiffRhythm: Blazingly fast and embarrassingly sim- ple end-to-end full-length song generation with latent diffusion,” arXiv preprint arXiv:2503.01183, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Mubert AI Music Generator,

    Mubert, “Mubert AI Music Generator,” Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://mubert.com/

    2026

  37. [37]

    Stable Audio 2.5,

    Stability AI, “Stable Audio 2.5,” Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://stability.ai/stable-audio

    2026

  38. [38]

    Songgen: A single stage auto-regressive transformer for text-to-song generation,

    Z. Liu, S. Ding, Z. Zhang, X. Dong, P. Zhang, Y . Zang, Y . Cao, D. Lin, and J. Wang, “Songgen: A single stage auto-regressive transformer for text-to-song generation,” inInternational Confer- ence on Machine Learning. PMLR, 2025, pp. 38 351–38 364

    2025

  39. [39]

    Suno: AI Music,

    Suno, “Suno: AI Music,” Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://suno.com/

    2026

  40. [40]

    Riffusion – Stable Diffusion for Real-Time Music Generation,

    Seth Forsgren and Hayk Martiros, “Riffusion – Stable Diffusion for Real-Time Music Generation,” GitHub repository, Dec. 2022, Accessed: 2026-01-13. [Online]. Available: https: //github.com/riffusion/riffusion-hobby

    2022

  41. [41]

    Udio: AI Music Generator,

    Udio, “Udio: AI Music Generator,” Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://www.udio.com/

    2026

  42. [42]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inNeurIPS, 2020

    2020

  43. [43]

    Xls-r: Self- supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pinoet al., “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” inProc. Interspeech 2022, 2022, pp. 2278–2282

    2022