Paper deep dive
Echoes: A semantically-aligned music deepfake detection dataset
Octavian Pascu, Dan Oneata, Horia Cucu, Nicolas M. Muller
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 98%
Last extracted: 3/26/2026, 1:49:24 AM
Summary
Echoes is a new, semantically-aligned music deepfake detection dataset containing 3,577 tracks (110 hours) generated by 10 diverse AI music systems. It addresses the limitations of existing benchmarks by enforcing semantic alignment between bona fide and synthetic audio, resulting in a more challenging and robust dataset that improves cross-dataset generalization for detection models.
Entities (6)
Relation Signals (3)
Echoes â benchmarkedagainst â AIME
confidence 100% · We evaluate Echoes in a cross-dataset setting against three existing AI-generated music datasets
Echoes â containsdatafrom â Free Music Archive
confidence 100% · built from bona fide songs sourced from the Free Music Archive [26].
Wav2Vec2 XLS-R 2B â usedtoevaluate â Echoes
confidence 100% · We evaluate Echoes in a cross-dataset setting... using state-of-the-art Wav2Vec2 XLS-R 2B representations.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 3,577 tracks (110 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.
Tags
Links
- Source: https://arxiv.org/abs/2603.23667v1
- Canonical: https://arxiv.org/abs/2603.23667v1
Trouble viewing inline? Open PDF directly â
Full Text
26,512 characters extracted from source content.
Expand or collapse full text
Echoes: A semantically-aligned music deepfake detection dataset Octavian Pascu 1 , Dan Oneata 1 , Horia Cucu 1 , Nicolas M. M Ì uller 2 1 National University of Science and Technology POLITEHNICA Bucharest, Romania 2 Fraunhofer AISEC, Germany octavian.pascu,dantheodor.oneata,horia.cucu@upb.ro, nicolas.mueller@aisec.fraunhofer.de Abstract We introduce Echoes, a new dataset for music deepfake detec- tion designed for training and benchmarking detectors under re- alistic and provider-diverse conditions. Echoes comprises 3,577 tracks (110 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 con- structed to be challenging, enforcing semantic-level alignment between spoofed audio and bona fide references. This align- ment is achieved by conditioning generated audio samples di- rectly 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; (i) detectors trained on existing datasets transfer poorly to Echoes; (i) training on Echoes yields the strongest generalization performance. These findings suggest that provider diversity and semantic align- ment help learn more transferable detection cues.Echoes is released publicly at https://huggingface.co/ datasets/Octavian97/Echoes under MIT license. Index Terms: music generation, deepfake detection, audio pro- cessing 1. 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 challenges for the music ecosys- tem: synthetic uploads can be used for impersonation, attribu- tion fraud, and automated monetization schemes, while stress- ing platform-level moderation. Streaming platforms have re- ported rapidly growing volumes of fully AI-generated submis- sions, motivating the need for scalable and reliable detection pipelines [1]. Music deepfake detection solutions are emerging in re- sponse.Current approaches are based on techniques from speech anti-spoofing. These include waveform-based models (e.g., RawNet2) [2], spectro-temporal attention models (e.g., AASIST) [3], and self-supervised encoders (e.g., wav2vec2- XLS-R) paired with lightweight classifiers and data augmen- tation to improve transfer [4, 5, 6]. However, as in speech anti- spoofing, generalization often suffers: models perform well 1 This paper was submitted for review to Interspeech 2026 when training and testing data follow a similar distributions, but performance drops sharply when evaluated on unseen gen- erators or different preprocessing pipelines [7, 8, 9, 10]. An important factor in achieving good generalization is data quality. Prior work found that datasets can exhibit unwarranted asymmetries [11, 12, 13], such as leading silences; these lead models to learn shortcut features that do not generalize. One solution is to generate fake samples that are as closely aligned to real samples as possible. In speech deepfake detection, for example, samples have been passed through the vocoders [14]; while in image deepfake detection, real images have been re- constructed using a latent diffusion model [15]. This deliberate alignment encourages detectors be focused on the generation artifacts rather than spurious differences. Data diversity is equally important. Different generator ar- chitectures and generation techniques produce distinct artifacts. This makes detectors trained on one generator to fail on an- other. Current architectures vary significantly, from autoregres- sive modeling over audio tokens [16, 17] to latent diffusion models [18, 19]. These methods produce differences in timbre, mixing, and post-processinw, further complicating generaliza- tion [20, 8]. In addition, the generation process itself introduces challenges [21, 22, 7]: a detector should detect a single forged component (e.g., synthesized vocals over real accompaniment), as well as an entirely generated song. However, existing deepfake music datasets do not em- phasise these aspects, and as such there remains a gap be- tween existing benchmarks and realistic evaluation require- ments. Among the existing options, AIME [23] targets large- scale human preference studies on short excerpts, prioritizing subjective evaluation over detector training; FakeMusicCaps [24] supports detection and attribution but remains largely clip- oriented and does not contain long-form structure; SONICS [25] emphasizes long-duration counterfeit-song detection, yet covers only two popular providers. To address these gaps, we introduce Echoes, 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 contains 3 577 tracks totaling 110 hours containing pop, rock, and electronic genres, built from bona fide songs sourced from the Free Music Archive [26]. To preserve high-level semantics, for each bona fide track we generate song-specific descriptors with LLMs (Figure 1) and use these descriptions (and, where supported, ref- erence audio) to prompt 10 popular music generators. Table 1 summarizes the existing datasets under three axes (provider di- versity, scale, and duration) and shows that Echoes fills in an important benchmarking gap. Finally, we report initial baseline results on our dataset us- ing a detector built on self-supervised Wav2Vec2 features. We arXiv:2603.23667v1 [cs.SD] 24 Mar 2026 Figure 1: Word cloud of generated song descriptions. The de- scriptions are produced by a large language model from songsâ metadata and are used to condition music generation in our dataset, Echoes. Table 1: Comparison of AI music datasets in terms of provider diversity, scale, and song duration. Duration NameProviders Total (h) Average (s) AIME [23]75810.18±0.3 SONICS [25]24 751 144.00±57.0 FakeMusicCaps [24]57710.08±0.1 Echoes (ours)10110 110.84±72.0 find that even in the in-domain setting (training and testing on Echoes), the performance is not saturated, leaving room for further improvement in subsequent work. In cross-dataset ex- periments, we observe the strongest transfer when training on Echoes and testing on the three datasets. This result highlights the value of a more diverse and better aligned training dataset. To summarize, our main contributions are: (i) a new semantically-aligned deepfake music dataset which contains both short-form and long-form synthetic songs; (i) coverage across major generator families by sourcing deepfakes from 10 providers; and (i) a cross-dataset baseline showing substan- tial generalization gaps and improved transfer when training on Echoes. 2. 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 C0, C-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 preserve the original style, we use each trackâs title and genre to prompt ChatGPT-5.0 Thinking to extract a list of stylistic characteristics. We query the large language model with the following prompt: Describe the following trackâs characteristics in as few words as possible with the idea of using the descriptions to generate a deepfake audio as close sounding as possible. Use commas between characteristics. The resulting descriptions are then used to condition the text-to-audio generation process. For providers that support Figure 2: Generation pipeline for the Echoes dataset. Spoofed music tracks are generated from corresponding bona fide audio in two ways: (1) by directly passing the bona fide track through an audio-to-audio model; (2) by providing its metadata (title and genre) to an LLM to produce stylistic descriptions, which is then used to condition a text-to-audio model. reference-audio conditioning, we also generate an audio-to- audio sample using the same descriptors together with the orig- inal track as reference. This pipeline is illustrated in Figure 2 and yields 3 577 generated (fake) tracks. 2.2. Additional metadata Apart from the audio files and their binary labels, we provide additional metadata to support a broader range of research tasks. Specifically, for each sample we include the generator name, generator type (text-to-audio or audio-to-audio), musical genre, and the prompt used for generation. This metadata facilitates reproducibility and enables richer analyses, such as attributing samples to the corresponding provider, studying the impact of the conditioning modality, or developing genre-aware models. 2.3. Dataset statistics Echoes contains 3 577 tracks totaling 6 629 minutes (or approx- imately 110 hours). Compared to existing datasets (Table 1), our dataset is larger than AIME (58 hours) and FakeMusicCaps (77 hours). Only SONICS contains more data (4751 hours), but it is less diverse as it includes samples from just two providers. Duration distribution. Track durations extend up to 480 seconds, with a median length of 106 seconds (Figure 3). The spikes in the distribution appear from providers that generate fixed-length outputs (around 30 s): for example, AudioLDM generates audio clips of 30.72 s on average, while SongGen out- put clips of 28.13 s (Table 2). Per-provider statistics. Table 2 lists the ten providers used to generate the samples, along with the underlying model (when available), number of generated tracks, average duration, and commercial status. Overall, eight out of the ten systems are commercial, and the number of tracks per provider ranges from 149 to 594. 3. 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. âUnknownâ means the provider does not publicly specify the model. The two generator types are either audio-to-audio (AâA) or text-to-audio (TâA). GeneratorNumberAverageTotal ProviderModelCommercial TâA AâAsongs duration (s) duration (h) AceStep [27] ACE-Step-v1-3.5BNoâ19666.12±20.545.4 AudioLDM [18] audioldm-s-fullNoâ58730.72±0.005.0 Brev [28] v4.5 ProYesâ298190.67±62.4415.8 DiffRhythm [29] DiffRhythm-v1.2Yesâ594121.16±14.6020.0 Mubert [30]UnknownYesâ149220.27±28.289.1 Stable Audio [31] AudioSparx 2.0Yesâ194180.00±0.009.7 SongGen [32] SongGen mixedproYesâ56128.13±3.134.4 Suno [33] v5Yesâ300179.30±54.8414.9 Riffusion [34] FUZZ 2.0Yesâ300178.97±39.4915.0 Udio [35]UnknownYesâ300130.94±0.1210.9 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 simple baseline that measures how much discriminative infor- mation is available from short local cues. Importantly, Echoes still differs from clip-only datasets because these segments are sampled from full-length songs and span diverse providers and generation settings, which increases variability even under a fixed-window evaluation. We view long-context modeling as a complementary direction and leave sequence-level evaluation on full tracks for future work. Train-test split. We randomly split the data into 80:20 training and evaluation splits. For non clip-centric datasets, in order to avoid trainâtest leakage, the split is performed at the track level: all 10-second segments extracted from the same full-length track are kept in the same split. Detection model. As the detection baseline, we use a self- supervised learning (SSL) front-end coupled with a lightweight classifier.We extract 1 920-dimensional embeddings using Wav2Vec2 XLS-R 2B [36, 37], which has shown strong perfor- mance in audio deepfake detection and related singing spoof- ing settings [6, 4]. We keep the Wav2Vec2 XLS-R 2B encoder frozen and obtain a fixed-dimensional embedding for each 10- second segment by applying a temporal average pooling over the frame-level representations from the final layer. On top of these embeddings, we train a logistic regression classifier to predict whether each segment is bona fide or AI- generated by minimizing binary cross-entropy with L2 regu- larization. We use scikit-learn logistic regression with a maximum of 5 000 iterations andC = 10 6 (all other param- eters are set to default), resulting in 1 921 trainable parameters (1 920 weights plus a bias). Performance is measured in terms of equal error rate (EER), a threshold-free metric commonly used in spoofing detection. 3.2. Cross-dataset evaluation The model is trained on each dataset in turn and evaluated on all four datasets (Table 3). This setup measures (i) how easy each dataset is in-domain (train and test on the same dataset) and (i) how well a model trained on one dataset transfers to the others. 3.2.1. In-domain detection We first look at the diagonal of Table 3 which represents in- domain results. We observe that Echoes is the most challeng- ing dataset under the same model and training protocol: the classifier achieves an EER of 9.36%, which exceeds those ob- tained on FakeMusicCaps (8.61%), AIME (6.40%), and SON- ICS (2.06%). From a dataset perspective, this is a desirable property: higher in-domain EER suggests fewer exploitable shortcuts and a more demanding benchmark for developing ro- bust detection methods. 3.2.2. Out-of-domain detection We next examine the off-diagonal entries in Table 3 which rep- resents out-of-domain testing. The first observation is that the out-of-domain performance is much lower than in-domain. This behavior is consistent with the broader audio deepfake detec- tion literature, where strong in-domain performance does not correlate with out-of-domain performance [7, 8, 9, 10]. Models trained on AIME, SONICS, or FakeMusicCaps perform poorly on Echoes, with EERs of 28.6%, 30.7%, and 41.7%, respec- tively. This indicates that cues learned from clip-centric datasets or low-diversity setups do not carry over well to Echoes. In the other direction, training on Echoes yields the best av- erage out-of-domain performance (21.0% average EER), and it achieves the best transfer to AIME (20.0%) and SONICS (14.8%) compared to the other training choices. Overall, these results support the goal of Echoes: it is difficult in-domain and it encourages learning cues that generalize better than those learned from the other benchmarks. Table 3: Equal error rate (EER; %) when training and testing on different combinations of datasets. In-domain results are shown in gray and best results are highlighted. Note that from a dataset quality perspective, higher is better (EER[%]â). Best cross-dataset performance (off-diagonal value) per training setup (i.e. row) is highlighted. Note that in a cross-dataset scenario, lower is better (EER[%]â). The average is computed on the out-of-domain combinations. Training datasetsâ Test datasetâ Average (Out-of-domain) AIMESONICSFakeMusicCapsEchoes AIME6.4015.1220.3628.6021.36 SONICS20.942.0634.3030.7228.65 FakeMusicCaps31.0330.168.6141.6637.10 Echoes (ours) 20.0214.8228.249.3621.02 Table 4: Per-provider accuracy on the Echoes dataset computed at the EER threshold achieving the best overall system performance. (threshold = 0.91) GeneratorDiffRhythm Suno Brev ACEStep Producer Udio SongGen AudioLDM Mubert StableAudio Accuracy (%)93.792.091.591.291.290.090.089.388.487.1 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%). Table 4 shows that performance varies across generators: accuracies range from 87.1% (StableAudio) up to 93.7% (DiffRhythm). This spread indicates that some providers produce outputs that are harder for the detector under the same training setup, and it supports the need for multi-provider benchmarks when reporting detection results. To understand whether the conditioning modality affects detec- tion difficulty, we also computed the accuracy at the optimal EER threshold on the two subsets of Echoes: one using only text-to-audio generated files and another using only audio-to- audio generated files. The model achieves 90.6% accuracy on TTA segments and 90.1% on ATA segments. The performance between the two is very similar, supporting the fact that the gen- erated descriptors are efficient in conveying the music charac- teristics. 3.4. t-SNE visualization Figure 4 provides a qualitative view of the AI-generated tracks from different providers. The t-SNE visualization was com- puted using the features extracted by the SSL model. Most providers overlap heavily, and there are no clean separations between clusters, which is consistent with the idea that provider identity is not trivially separable in this dataset. The main ex- ception is DiffRhythm, which forms a more distinct region, and is also the easiest model to predict when looking at Table 4. Bona fide tracks are scattered throughout the plot rather than forming a single compact cluster, suggesting that the repre- sentation space is mixed and that simple clustering-based rules would not be enough for reliable detection. 4. Conclusion We presented Echoes, a semantically-aligned dataset for AI- generated music detection built to support training and bench- marking under realistic provider diversity. Echoes contains 3577 tracks (110 hours) across pop, rock, and electronic mu- sic genres.For each bona fide song from the Free Music Archive, we generate AI-produced counterparts from 10 popu- lar music generators using LLM-derived, song-specific descrip- tors and, when available, audio-conditioned generation. This pairing strategy aims to keep high-level intent (style/theme) similar between bona fide and AI-generated tracks, making the task harder and reducing simple content-based shortcuts. Our experiments show two key results. First, Echoes is the hardest dataset in-domain under the same SSL+linear base- line (9.36% EER), compared with AIME (6.40%), SONICS (2.06%), and FakeMusicCaps (8.61%). Second, cross-dataset transfer is weak in general, and especially weak when testing on Echoes: models trained on AIME, SONICS, or FakeMusicCaps reach 28.6â 41.7% EER on Echoes. In contrast, training on Echoes yields the strongest average out-of-domain performance (21.0% EER), suggesting that combining semantic alignment with multi-provider coverage pushes detectors toward cues that transfer better across datasets. Echoes is released to enable more realistic evaluation of AI- generated music detection and to encourage reporting beyond in-domain results. Future work will evaluate detectors under settings closer to deployment (e.g., unseen providers, common audio post-processing, and partial or mixed real/AI content). 5. References [1] Deezer, âDeezer:28% of all delivered music is now fully AI-generated,â Deezer Newsroom (press release), Sep. 2025, accessed:2026-01-11. [Online]. Available:https: //newsroom-deezer.com/2025/09/28-fully-ai-generated-music/ [2] H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, âEnd-to-end anti-spoofing with RawNet2,â in ICASSP, 2021. [3] J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, âAASIST: Audio anti-spoofing using in- tegrated spectro-temporal graph attention networks,â in ICASSP, 2022. [4] X. Chen, H. Wu, R. Jang, and H.-y. Lee, âSinging voice graph modeling for singfake detection,â in Proc. Interspeech 2024, 2024, p. 4843â4847. [5] H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, âAutomatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,â in The Speaker and Language Recognition Workshop (Odyssey 2022). ISCA, 2022. [6] O. Pascu, A. Stan, D. Oneata, E. Oneata, and H. Cucu, âTowards generalisable and calibrated audio deepfake detection with self- supervised representations,â in Interspeech, vol. 2024, 2024, p. 4828â4832. [7] D. Afchar, G. Meseguer-Brocal, and R. Hennequin, âDetect- ing music deepfakes is easy but actually hard,â arXiv preprint arXiv:2405.04181, 2024. [8] Y. Li, M. Milling, L. Specia, and B. W. Schuller, âFrom audio deepfake detection to AI-generated music detectionâA pathway and overview,â arXiv preprint arXiv:2412.00571, 2024. [9] M. Todisco, X. Wang, V. Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T. Kinnunen, and K. A. Lee, âAsvspoof 2019: Future horizons in spoofed and fake audio detec- tion,â in Interspeech 2019. International Speech Communication Association, 2019, p. 1008â1012. [10] N. M. M Ì uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B Ì ottinger, âDoes audio deepfake detection generalize?â in In- terspeech, 2022. [11] N. M Ì uller, F. Dieckmann, P. Czempin, R. Canals, K. B Ì ottinger, and J. Williams, âSpeech is silver, silence is golden: What do asvspoof-trained models really learn?âin Proc. ASVSPOOF 2021, 2021, p. 55â60. [12] S. Borz ` ı, O. Giudice, F. Stanco, and D. Allegra, âIs synthetic voice detection research going into the right direction?â in CVPRW, 2022. [13] S. Smeu, D.-A. Boldisor, D. Oneata, and E. Oneata, âCircum- venting shortcuts in audio-visual deepfake detection datasets with unsupervised learning,â in CVPR, 2025. [14] X. Wang and J. Yamagishi, âSpoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,â in ICASSP, 2023. [15] A. S. Rajan, U. Ojha, J. Schloesser, and Y. J. Lee, âOn the effec- tiveness of dataset alignment for fake image detection,â in ICLR, 2025. [16] A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi et al., âMusicLM: Generating music from text,â arXiv preprint arXiv:2301.11325, 2023. [17] J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y. Adi, and A. D Ì efossez, âSimple and controllable music gen- eration,â Advances in Neural Information Processing Systems, vol. 36, p. 47 704â47 720, 2023. [18] H. Liu, Z. Chen, Y. Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. Plumbley, âAudioldm: Text-to-audio generation with la- tent diffusion models,â in Proceedings of the 40th International Conference on Machine Learning, PMLR 2023, vol. 202.In- ternational Machine Learning Society (IMLS), 2023, p. 21 450â 21 474. [19] Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, âStable audio open,â in ICASSP, 2025. [20] L. Cros Vila, B. Sturm, L. Casini, and D. Dalmazzo, âThe AI mu- sic arms race: On the detection of AI-generated music,â Transac- tions of the International Society for Music Information Retrieval, vol. 8, no. 1, p. 179â194, 2025. [21] Y. Zang, Y. Zhang, M. Heydari, and Z. Duan, âSingfake: Singing voice deepfake detection,â in ICASSP, 2024. [22] Y. Zhang, Y. Zang, J. Shi, R. Yamamoto, T. Toda, and Z. Duan, âSVDD 2024: The inaugural singing voice deepfake detection challenge,â in SLT, 2024. [23] F. Gr Ì otschla, A. Solak, L. A. Lanzend Ì orfer, and R. Wattenhofer, âBenchmarking music generation models and metrics via human preference studies,â in ICASSP, 2025. [24] L. Comanducci, P. Bestagini, and S. Tubaro, âFakeMusicCaps: A dataset for detection and attribution of synthetic music generated via text-to-music models,â Journal of Imaging, vol. 11, no. 7, p. 242, 2025. [25] M. A. Rahman, Z. I. A. Hakim, N. H. Sarker, B. Paul, and S. A. Fattah, âSonics: Synthetic or not-identifying counterfeit songs,â in The Thirteenth International Conference on Learning Repre- sentations, 2024. [26] M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, âFma: A dataset for music analysis,â in 18th International Society for Music Information Retrieval Conference, 2017. [27] J. Gong, S. Zhao, S. Wang, S. Xu, and J. Guo, âACE-step: A step towards music generation foundation model,â arXiv preprint arXiv:2506.00045, 2025. [28] Brev, âBrev AI Music Generator,â Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://brev.ai/ [29] Z. Ning, H. Chen, Y. Jiang, C. Hao, G. Ma, S. Wang, J. Yao, and L. Xie, âDiffRhythm: Blazingly fast and embarrassingly sim- ple end-to-end full-length song generation with latent diffusion,â arXiv preprint arXiv:2503.01183, 2025. [30] Mubert, âMubert AI Music Generator,â Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://mubert.com/ [31] Stability AI, âStable Audio 2.5,â Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://stability.ai/stable-audio [32] Z. Liu, S. Ding, Z. Zhang, X. Dong, P. Zhang, Y. Zang, Y. Cao, D. Lin, and J. Wang, âSonggen: A single stage auto-regressive transformer for text-to-song generation,â in International Confer- ence on Machine Learning. PMLR, 2025, p. 38 351â38 364. [33] Suno, âSuno:AI Music,â Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://suno.com/ [34] Seth Forsgren and Hayk Martiros, âRiffusion â Stable Diffusion for Real-Time Music Generation,â GitHub repository, Dec. 2022, Accessed:2026-01-13. [Online]. Available:https: //github.com/riffusion/riffusion-hobby [35] Udio, âUdio: AI Music Generator,â Website, Jan. 2026, Accessed: 2026-01-13. [Online]. Available: https://w.udio.com/ [36] A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, âwav2vec 2.0: A framework for self-supervised learning of speech representa- tions,â in NeurIPS, 2020. [37] A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y. Saraf, J. Pino et al., âXls-r: Self- supervised cross-lingual speech representation learning at scale,â in Proc. Interspeech 2022, 2022, p. 2278â2282.