Multimodal Dataset for Music-Taste Correspondences: Cross-Modal Features, Food Chemistry Vectors, and Perceptual Ratings

This dataset accompanies the paper Multimodal Dataset Normalization and Perceptual Validation for Music-Taste Correspondences (Spanio, Frezzato, Rodà, 2026). It provides all data needed to reproduce the two experiments reported in the paper. Experiment 1 — Cross-modal transfer analysis. A small corpus of 257 music tracks with human-annotated flavor ratings (aggregated from 22 published studies) and a large FMA-derived corpus of ~49,300 30-second segments with Audio Spectrogram Transformer-generated flavor labels. Both corpora include extracted audio features (MFCCs, chroma, spectral descriptors) and five-dimensional taste annotations (sweet, bitter, salty, sour, spicy). A genre taxonomy file supports text-flavor association analyses. Experiment 2 — Perceptual validation. Food chemistry resources for constructing computational taste vectors: ~70,000 FooDB compound-level FART taste predictions, nutrient-to-taste mappings, aggregated food-level vectors (992 foods), and the final 20 five-dimensional target vectors used in the experiment. Raw perceptual ratings from an online listener study (49 participants, 20 tracks, 5 taste dimensions, 7-point Likert scale) and stimuli metadata are included. The 20 audio stimuli (30-second MP3 excerpts from the Free Music Archive, Creative Commons licensed) are provided in the audio-stimuli/ folder. Companion code for reproducing all analyses and figures is available at https://github.com/CSCPadova/music-flavor-analysis.