Examining Melodic Similarity Across Human, Computational, and Legal Perspectives

Examining Melodic Similarity Across Human, Computational, and Legal Perspectives Roser Batlle-Roca*,Lena M. Melo and Xavier Serra Music Technology Group, Universitat Pompeu Fabra, Barcelona Abstract As music-generative AI continues to evolve, critical concerns are emerging around originality, data replication, and the misuse of intellectual property. The Music Replication Assessment (MiRA) tool was proposed to detect potential data replication using audio-based similarity metrics. However, it remains unclear how its classifications align with human perception or legal judgments. Thus, this study aims to examine melodic similarity across human, computational, and legal perspectives. We conducted a perceptual experiment based on real-world copyright infringement cases, where participants completed two tasks: (1) a direct rating of melodic similarity, and (2) a forced-choice judgment on potential copying. Participants’ responses were compared with MiRA’s similarity scores and court rulings. Results revealed systematic discrepancies: legal decisions tended to align with forced-choice judgments, while MiRA corresponded more closely with rating patterns, particularly in ambiguous cases. These findings highlight the complexity of assessing music similarity and emphasise the need to incorporate human-centred evaluation into computational tools and legal discourse around AI-generated music. Keywords Melodic similarity, Music plagiarism, Music data replication 1. Introduction Generative AI is rapidly expanding into creative domains, often framed as democratizing access and empowering artistic expression [ 1 , 2 ]. In music, however, they also raise concerns about originality, training data replication, and intellectual property violations, such as plagiarism or copyright infringement [ 3 , 4 , 5 , 6 ]. Despite growing concerns, few tools directly assess data replication in AI-generated music from raw audio [ 6 , 7 ]. One relevant example is the Music Replication Assessment (MiRA) tool [ 8 ], which is built on four different audio-based music similarity metrics. Nonetheless, music similarity is shaped by melody, harmony, rhythm, timbre, and listener background, making it inherently complex and subjective [ 9 ]. Legal assessments further complicate matters by relying on substantial similarity, a vague notion that prioritises overall impression rather than element-by-element comparison [ 10 , 11 ]. Yet, MiRA has not been validated against human perception or legal rulings, limiting its interpretability. We aim to explore how assessments of music similarity align across human, computational and legal perspectives. In particular, we ask whether MiRA aligns with human perception, and how MiRA and human judgments relate to legal rulings. We focus on melodic similarity, often considered the most memorable musical feature [ 12 , 13 ], while acknowledging its perception HCMIR25: 3rd Workshop on Human-Centric Music Information Research, September 20th, 2025, Daejeon, Korea *Corresponding author. $roser[email protected]du (R. Batlle-Roca) ©2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). depends on interval patterns, rhythm, motifs, and the listener’s background [ 14 , 15 ]. We conduct a perceptual experiment using fragment pairs derived from real-world court cases involving music plagiarism, highlighting the convergence and divergence between perceptual, computational, and legal perspectives. 2. Methodology 2.1. The MiRA tool The MiRA tool [ 8 ] is a model-independent approach built on audio-based music similarity metrics to detect how closely a generated sample may resemble reference samples. MiRA computes similarity between reference and target samples using global and per-pair distances to assess replication. It is composed of four metrics: • Symmetric Kullback-Leibler (KL) divergence: A statistical measure between two probability distributions (reference and target). KL divergence is computed in both directions using the PaSST audio classifier proposed in Koutini et al. [ 16 ], trained on Audioset, and then takes the average to obtain a symmetric score. • Cover Song Identification (CoverID) distance [ 17 , 18 , 19 ]: A measure to detect whether two recordings share the same underlying composition, accounting for variations in tempo, structure, or instrumentation while retaining melodic and harmonic similarity. It relies on pitch-content features and local alignment, using Essentia’s implementation. • Contrastive Language-Audio Pretraining (CLAP) score [ 20 ]: A similarity measure based on embeddings from the pre-trained CLAP model, 1 where each music sample is projected into a latent space and compared via cosine distance. • Discogs-EffNet (DEfNet) score: A similarity measure based on embeddings from Essentia’s Discogs-EffNet model 2 [ 21 ], trained via contrastive learning on Discogs metadata. Cosine distance between embeddings provides an estimate of track-level similarity grounded in music-specific information. 2.2. Music Copyright Infringement Cases Dataset Our study builds upon the Music Copyright Infringement Cases (MCIC) dataset introduced by Park et al. [ 22 ], which compiles 116 real-world legal cases of alleged music plagiarism from 1915 to 2020 (Infringed: 32, Denied: 66, Settled: 18). The dataset includes metadata and detailed summaries of each case, including insights on the assessment of melodic similarity and the court ruling. We focused on cases involving melodic similarity, excluding those settled out of court or where the audio was unavailable or of insufficient quality. To ensure participant engagement and avoid bias, we selected a subset of 18 pairs: 9 cases ruled as non-plagiarism (Denied), which correspond to C01-C09, and 9 as plagiarism (Infringed), which correspond to C11-C19, balancing legal outcomes. 3 Each stimulus was 10–15 seconds long, except for C16 (22s) to preserve the full melodic content of that case. In selecting each excerpt, we reviewed the 1https://huggingface.co/lukewys/laion_clap/tree/main#:~:text=music_audioset_epoch_15_esc_90.14.pt 2https://essentia.upf.edu/models.html#discogs-effnet 3Samples are available in complementary page: https://tinyurl.com/melodic-similarity-exp-stimuli provided case summaries in the dataset, and then we manually identified the sections of each song containing the melodic material central to the infringement claim. While this approach relied on auditory judgment rather than automatic or annotated melody extraction, it ensured alignment with the legal context. All audio was normalised to -23 LUFS (EBU R128 [ 23 ]). We computed MiRA classifications for each pair of stimuli. 2.3. Perceptual Experiment To investigate how human perception aligns with legal rulings and computational metrics of melodic similarity, we designed a perceptual experiment where participants listened to pairs of melodies and rated how similar they sounded. The experiment consisted of two tasks: Task 1 (T1) - a direct scaling task in which participants rated the similarity of the two fragments on a 5-point scale (1 = Not at all, 5 = Very much), and Task 2 (T2) - a forced-choice task asking whether the second melody sounded like a copy or imitation of the first. To avoid bias and ensure clarity in the framing, specific legal terminology was excluded. The experiment was implemented using the webMUSHRA framework [ 24 ] with an estimated time of 20 minutes. It was available online both in English (en) and Catalan (cat) for 25 days (April 14–May 8, 2025). 2.3.1. Participants A total of 86 individuals took part in the experiment (en: 52, cat: 34). Participants were volunteers, recruited through different mailing lists and social media channels. To ensure task comprehension and response reliability, a validation procedure was applied using three reference cases with clear legal outcomes (C12, C16 as infringement; C04 as non-infringement). Participants who failed at least two of these were excluded, resulting in 79 valid responses: 48 en-participants and 31 cat-participants. Participants (75.5% Europe-based) ranged in age from 18 to 61 (M = 33). Musical background was self-reported across three levels: professionalintermediate, amateur-enthusiast and non-musician. Most participants had formal training: 45% of cat-participants and 52.1% of en-participants identified as professional-intermediate musicians, while another 45% and 43.8%, respectively, as amateur-enthusiasts. Few reported no musical experience (cat: 9.7%, en: 4.2%). These categories informed our analysis of how expertise affects similarity perception. 3. Results We conducted two two-way ANOVAs to assess whether participant language affected responses in T1 and T2. Since no statistical significance was observed ( 𝑝𝑇1= 0.917 , 𝑝𝑇2= 0.658 ), we combined results across both groups. Table 1summarises the overall results, including human perceptual responses, MiRA assessments, 4 and legal rulings. MiRA assessments were derived by applying similarity thresholds specific to each metric to classify cases as infringement or not. Thresholds were set using percentiles of each score distribution: 70th percentile for CLAP and DEfNet ( ≥ 0.70), 15th percentile for KL divergence ( ≤ 0.35), and 30th percentile for CoverID ( ≤ 0.27). These values were chosen in correspondence with the distributions and overall result patterns, ensuring consistency across metrics. 4Note that similarity is indicated with low KL div/CoverID and high CLAP/DEfNet values. Table 1: Results of perceptual (Human Perception), computational (MiRA), and legal (Ruling) assessments. Table is split between Denied and Infringed cases. T1 represents the median ( 𝑀 ) across participants, and T2 describes the average of responses in percentage (no for Denied, yes for Infringed). In bold, we indicate when T1, T2 or MiRA metrics align with the court decision. Ruling: Denied Ruling: Infringed Human Perception Human Perception C T1 (𝑀) T2𝑛𝑜 (%) MiRA C T1 (𝑀) T2𝑦𝑒𝑠 (%) MiRA C01 2 83.54 KL div: 0.28 CoverID: 0.37 CLAP: 0.83 DEfNet: 0.53 C11 4 73.42 KL div: 0.36 CoverID: 0.42 CLAP: 0.75 DEfNet: 0.54 C02 4 36.7 KL div: 0.63 CoverID: 0.41 CLAP: 0.76 DEfNet: 0.61 C12 4 88.61 KL div: 0.92 CoverID: 0.25 CLAP: 0.82 DEfNet: 0.73 C03 2 78.48 KL div: 0.71 CoverID: 0.25 CLAP: 0.79 DEfNet: 0.61 C13 4 60.76 KL div: 1.02 CoverID: 0.16 CLAP: 0.55 DEfNet: 0.53 C04 2 93.67 KL div: 0.90 CoverID: 0.65 CLAP: 0.43 DEfNet: 0.19 C14 2 10.13 KL div: 5.07 CoverID: 0.38 CLAP: 0.34 DEfNet: 0.14 C05 2 98.73 KL div: 0.41 CoverID: 0.46 CLAP: 0.50 DEfNet: 0.20 C15 2 15.19 KL div: 0.53 CoverID: 0.27 CLAP: 0.82 DEfNet: 0.53 C06 362.02 KL div: 0.14 CoverID: 0.23 CLAP: 0.75 DEfNet: 0.47 C16 4 78.49 KL div:0.71 CoverID: 0.21 CLAP: 0.65 DEfNet: 0.39 C07 4 20.25 KL div: 0.41 CoverID: 0.51 CLAP: 0.86 DEfNet: 0.72 C17 4 72.16 KL div: 0.49 CoverID: 0.22 CLAP: 0.63 DEfNet: 0.40 C08 2 83.54 KL div: 0.81 CoverID: 0.40 CLAP: 0.69 DEfNet: 0.60 C18 2 26.59 KL div: 4.42 CoverID: 0.48 CLAP: 0.50 DEfNet: 0.25 C09 360.76 KL div: 0.52 CoverID: 0.16 CLAP: 0.70 DEfNet: 0.77 C19 2 7.51 KL div: 0.79 CoverID: 0.50 CLAP: 0.54 DEfNet: 0.29 The analysis reveals links between MiRA’s classifications, musical features, and perceptual judgments. For instance, C03 and C15, both rap songs, show high similarity in CoverID and CLAP, suggesting alignment in instrumental beats. C02 and C11, which only score highly in CLAP, share lyrical similarities, reflecting CLAP’s sensitivity to audio-text alignment. C13, C16 and C17 score highly only in CoverID, due to their near-identical melodic and harmonic structure despite tempo differences, highlighting CoverID’s sensitivity to these features. In these cases where only CoverID reports similarity, T2 𝑦𝑒𝑠 responses show strong agreement—especially among professional-intermediate musicians ( ∼ 70%)—while amateur-enthusiast responses are more varied. This suggests CoverID closely aligns with expert perception of musical similarity. When comparing perceptual, computational, and legal classifications, we can observe several patterns. For court-denied cases such as C04, C05, and C08, both human judgments and MiRA metrics consistently indicate no evidence of plagiarism. By contrast, C06 and C09 yield more ambiguous results, as their median rating are neutral (T1 = 3 ), approximately 60% of participants indicated no in T2, and MiRA metrics provide mixed signals, with some pointing to similarity and others not. In C02, MiRA largely aligns with the court ruling (with 3 out of 4 metrics indicating no similarity), whereas human perception suggests the opposite (T1 = 4 , T2 𝑛𝑜 = 36.7 ). Instead, for court-infringed cases such as C14, C18 and C19, neither participant responses nor MiRA metrics indicate potential plagiarism, underscoring a notable misalignment with the legal ruling. Finally, C12 shows the strongest alignment across the three perspectives as all metrics except KL divergence indicate similarity, and participants rated this case with high similarity (T1 = 4 and T2𝑦𝑒𝑠 = 88.61). 4. Discussion and Conclusion 4.1. Comparing Human, Computational and Legal Perspectives Our study examined how melodic similarity is perceived across three lenses: human, computational, and legal. Results show that these perspectives converge in some clear-cut cases, but diverge notably in ambiguous or contextually complex examples. Subjectivity and variability in human perception: Participant responses varied by musical background and intuition, specifically in mid-scale similarity cases (e.g., C06, C09). While the direct similarity ratings (T1) and binary copy judgments (T2) aligned, especially in extreme cases, mid-scale cases revealed perceptual ambiguity. We observe that T1 provided broader perceptions, while T2 conditioned more conservative judgments for its binary nature. Professionals tended to align more closely with legal rulings, supporting prior claims that expert listeners bring domain-specific criteria to their judgments. Hence, perceptual similarity judgements are influenced by both musical expertise and the way the task is presented. Consistency and limitations of computational metrics: MiRA was generally consistent in identifying absence of similarity. For instance, in C04, C05 and C08, all MiRA metrics agreed on low similarity, as did most participants. Complete convergence also occurred in strong similarity cases (e.g., C12). Among embedding-based MiRA metrics, CLAP flagged more cases as similar, but DEfNet more often aligned with human judgments (12 cases) than CLAP (10 cases), despite providing fewer classifications above the high-similarity threshold. This may indicate DEfNet is stricter than CLAP but potentially more reliable, although both metrics align with human perception in 5 out of 9 court-infringed cases. In cases such as C13, C16 and C17, these computational assessments diverged from human judgment, revealing that embedding-based models may capture surface features but miss structural or stylistic nuances. CoverID, by contrast, was sensitive to melodic and harmonic structure despite tempo variations, and often aligned more closely with expert judgments. KL divergence rarely flagged high similarity (only two cases: C01 and C06), and those did not align with perceptual or legal judgments, suggesting it is unreliable for detecting similarity but useful as a conservative indicator of dissimilarity. Legal judgment thresholds: Court rulings sometimes diverged from both MiRA and participant judgments. In court-infringed cases such as C14, C18 and C19, both MiRA and participants indicated no similarity. Conversely, in cases dismissed as “too generic” by courts (e.g., C06, C09), participants showed signs of ambiguity, while MiRA flagged potential similarity. These mismatches suggest that legal judgments, although grounded in precedent and expert testimony, sometimes diverge from perceptual and computational evidence, possibly due to differences in thresholds for originality, intent, or access that extend beyond perceptual or algorithmic assessments. Alignment across perspectives: Complete alignment across all three perspectives was rare. Convergence occurred more consistently when indicating the absence of similarity (nonplagiarism), whereas high-similarity cases showed greater divergence. Notably, C12 represents a strong alignment, where MiRA, human perception, and the court ruling all identified high similarity. However, this was the only court-infringed case where all three perspectives converged. In contrast, for court-denied cases, this happened more often (C02, C04, C05 and C08). Overall, these findings emphasise that melodic similarity is not an objective fact but a construct shaped by task framing, listener expertise, computational reduction, and legal interpretation. Combining these perspectives enriches our understanding of music similarity, underscoring both strengths and limitations of each approach in evaluating potential replication in AI-generated music. 4.2. Limitations and Future Work This work presents an exploration of perceptual, computational, and legal perspectives on music similarity, in the context of potential data replication in AI-generated music. While our approach provided valuable insights, it has also introduced certain limitations. The study focuses exclusively on melodic similarity, excluding other musical dimensions, such as harmony and rhythm. Despite efforts to diversify participants, most responses came from Europe, limiting cultural breadth and possibly influencing similarity judgments. Manual selection of melodic excerpts, guided by case summaries and authors’ auditory judgement, may also introduce bias. The relatively small sample (18 case pairs) constrains statistical depth. Future work should broaden both the participant demographics and musical features to explore their interaction with perception and legal reasoning. Expanding MiRA’s feature set and integrating perceptual feedback may improve alignment with human evaluations. While MiRA currently combines multiple metrics, including general-purpose (KL divergence, CLAP) and music-related (CoverID, DEfNet) metrics, further exploration of alternative music-specific representations, such as Musicset Unsupervised Large Embedding (MULE) [ 25 ], may help clarify whether more specialised embedding-based metrics would narrow the gap between perceptual, legal, and computational judgments. Finally, embedding MiRA in ethical and legal workflows may foster transparency in music-generative AI. Acknowledgments The authors would like to thank Perfecto Herrera from the Music Technology Group at Universitat Pompeu Fabra, for his insightful discussions and valuable feedback throughout the development of this research. References [1] K. Tatar, P. Ericson, K. Cotton, P. T. N. Del Prado, R. Batlle-Roca, B. Cabrero-Daniel, S. Ljungblad, G. Diapoulis, J. Hussain, A shift in artistic practices through artificial intelligence, Leonardo 57 (2024) 293–297. URL: https://doi.org/10.1162/leon%5C%5C_a% 5C%5C_02523. doi:10.1162/leon_a_02523. [2] B. Caramiaux, K. Crawford, Q. V. Liao, G. Ramos, J. Williams, Generative ai and creative work: Narratives, values, and impacts, 2025. URL: https://arxiv.org/abs/2502.03940. arXiv:2502.03940. [3] B. L. Sturm, M. Iglesias, O. Ben-Tal, M. Miron, E. Gómez, Artificial intelligence and music: open questions of copyright law and engineering praxis, in: Arts, volume 8, MDPI, 2019, p. 115. [4] Z. Yin, F. Reuben, S. Stepney, T. Collins, Measuring when a music generation algorithm copies too much: The originality report, cardinality score, and symbolic fingerprinting by geometric hashing, SN Computer Science 3 (2022) 340. [5] J. Barnett, The ethical implications of generative audio models: A systematic literature review, in: Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 2023, pp. 146–161. [6] R. Batlle-Roca, E. Gómez, W. L. S. R. X. Serra, Y. Mitsufuji, W. hsiang Liao, J. X. Serra, E. Gómez, W.-H. Liao, X. Serra, Transparency in music-generative ai: A systematic literature review, Research Square (preprint) (2024). URL: https://www.researchsquare. com/article/rs-3708077/v2. doi:10.21203/RS.3.RS-3708077/V2. [7] Z. Xiong, W. Wang, J. Yu, Y. Lin, Z. Wang, A comprehensive survey for evaluation methodologies of ai-generated music, arXiv preprint arXiv:2308.13736 (2023). [8] R. Batlle-Roca, W. Liao, X. Serra, Y. Mitsufuji, E. Gómez, Towards assessing data replication in music generation with music similarity metrics on raw audio, in: Proceedings of the 25th International Society for Music Information Retrieval Conference, ISMIR 2024, San Francisco, California, USA, November 10-14, 2024, 2024, pp. 1004–1011. URL: https: //doi.org/10.5281/zenodo.14877501. doi:10.5281/ZENODO.14877501. [9] A. Berenzweig, B. Logan, D. P. Ellis, B. Whitman, A large-scale evaluation of acoustic and subjective music-similarity measures, Computer Music Journal 28 (2004) 63–76. URL: https://doi.org/10.1162/014892604323112257. doi:10.1162/014892604323112257. [10] A. Latman, "probative similarity" as proof of copying: Toward dispelling some myths in copyright infringement, Columbia Law Review 90 (1990) 1187–1214. URL: http://www. jstor.org/stable/1122876. [11] S. Scheffler, E. Tromer, M. Varia, Formalizing human ingenuity: A quantitative framework for copyright law’s substantial similarity, in: Proceedings of the 2022 Symposium on Computer Science and Law, CSLAW ’22, Association for Computing Machinery, New York, NY, USA, 2022, p. 37–49. URL: https://doi.org/10.1145/3511265.3550444. doi: 10.1145/ 3511265.3550444. [12] D. Müllensiefen, M. Pendzich, Court decisions on music plagiarism and the predictive value of similarity algorithms, Musicae Scientiae 13 (2009) 257–295. URL: https://doi.org/10.1177/102986490901300111. doi: 10.1177/102986490901300111 . arXiv:https://doi.org/10.1177/102986490901300111. [13] P. Savage, C. Cronin, D. Müllensiefen, Q. Atkinson, Quantitative evaluation of music copyright infringement, 2018. [14] A. M. and, Interrogating melodic similarity: A definitive phenomenon or the product of interpretation?, Journal of New Music Research 41 (2012) 323–335. URL: https://doi.org/10.1080/09298215.2012.740051. doi: 10.1080/09298215.2012.740051 . arXiv:https://doi.org/10.1080/09298215.2012.740051. [15] J. Herre, E. Allamanche, C. Ertel, How similar do songs sound? towards modeling human perception of musical similarity, in: 2003 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (IEEE Cat. No.03TH8684), 2003, pp. 83–86. doi:10.1109/ASPAA.2003.1285825. [16] K. Koutini, J. Schlüter, H. Eghbal-zadeh, G. Widmer, Efficient training of audio transformers with patchout, in: Interspeech 2022, 23rd Annual Conference of the International Speech Communication Association, Incheon, Korea, ISCA, 2022, pp. 2753–2757. [17] J. Serrà, X. Serra, R. Andrzejak, Cross recurrence quantification for cover song identification, New Journal of Physics 11 (2009). doi:10.1088/1367-2630/11/9/093017. [18] J. Serrà, E. Gómez, P. Herrera, X. Serra, Chroma binary similarity and local alignment applied to cover song identification, IEEE Transactions on Audio, Speech, and Language Processing 16 (2008) 1138–1151. doi:10.1109/TASL.2008.924595. [19] J. Serrà, E. Gómez, P. Herrera, Transposing chroma representations to a common key, IEEE CS Conference on The Use of Symbols to Represent Music and Multimedia Objects (2008). [20] Y. Wu, K. Chen, T. Zhang, Y. Hui, T. Berg-Kirkpatrick, S. Dubnov, Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation, ArXiv (2023). arXiv:2211.06687. [21] P. Alonso-Jiménez, D. Bogdanov, J. Pons, X. Serra, Tensorflow audio models in Essentia, in: International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020. [22] S. Park, H. Kim, J. Jung, J. Park, J. Kim, J. Nam, Quantitative analysis of melodic similarity in music copyright infringement cases, in: Proceedings of the 25th International Society for Music Information Retrieval Conference (ISMIR), San Francisco, USA, 2024. [23] EBU-Recommendation, R 128: Loudness normalisation and permitted maximum level of audio signals, https://tech.ebu.ch/publications/r128, 2023. [24] M. Schoeffler, S. Bartoschek, F.-R. Stöter, M. Roess, S. Westphal, B. Edler, J. Herre, webmushra — a comprehensive framework for web-based listening tests, Journal of Open Research Software (2018). doi:10.5334/jors.187. [25] M. C. McCallum, F. Korzeniowski, S. Oramas, F. Gouyon, A. F. Ehmann, Supervised and unsupervised learning of audio representations for music understanding, arXiv preprint arXiv:2210.03799 (2022).