Cross-cultural evaluation of taste–sound correspondences in AI-generated music

Archive accompanying the article Spanio M., Zampini M., Torri L., Migliavada R., Mesz B., Ohno M., Wada Y., Rodà A. Cross-cultural evaluation of taste–sound correspondences in AI-generated music. PLOS ONE, 2026. This deposit contains everything needed to reproduce the study end to end: the survey responses, the analysis code, every derived table and figure, the full analysis report, and the Supporting Information submitted with the manuscript. It is self-contained — no part of the analysis depends on a file that is not in this archive, except the audio stimuli, which are noted separately below. What the study did 361 participants in three countries — Japan (n = 140), Argentina (n = 104) and Italy (n = 117) — evaluated short musical excerpts generated by a text-to-music model(MusicGen) from four taste prompts: sweet, sour, bitter, salty. Each participant completed two listening tasks online, in their native language, via PsyToolkit: Task 1 — model preference. Five trials, each pairing a clip from the original public MusicGen model with a clip from the fine-tuned model for the same taste prompt.Responses are on a 0–10 slider with 5 as the neutral point; scores are stored already oriented so that above 5 means the fine-tuned clip was preferred. Task 2 — perception ratings. Three fine-tuned excerpts, each rated on twelve 1–5 scales: four taste descriptors (sweet, sour, bitter, salty), six emotiondescriptors (happy, sad, anger, disgust, fear, surprise) and two thermal descriptors (hot, cold). 3 excerpts × 12 descriptors = 36 ratings per participant Throughout the archive, prompt always means the taste given to the generative model and descriptor always means one of the twelve items participants rated. The two are kept lexically distinct even where they coincide. The stimulus set is 100 fine-tuned clips, 25 per prompt. Participation was voluntary and restricted to adults. The protocol was approved by the Ethics Committee of the University of Gastronomic Sciences of Pollenzo (Minutes19122025). No personally identifying information was collected; every file here is anonymous, with participants identified only by a synthetic per-cohort index.