Learning Musicality: Machine Classification and Generation of Music vs. Non-Music from Human-Labeled Sound Samples

This project investigates whether machine learning models can learn and reproduce human judgments about the boundary between “music” and “non‑music” in culturally diverse sound samples. Building on a labeled dataset from a prior behavioral study, the computational component focuses on two aims: (1) training and evaluating supervised classification models to predict human musicality judgments from acoustic features, and (2) developing generative models capable of synthesizing novel audio conditioned on the learned “music” vs. “non‑music” distinction. The classification phase involves extracting MFCCs, spectral descriptors, rhythmic features, and pitch‑related measures from each sound sample, followed by training multiple algorithms—including k‑nearest neighbors, decision trees, random forests, logistic regression, CatBoost, and gradient boosting. Model performance is assessed using accuracy, precision, recall, and F1 score on held‑out test data, with error analysis used to refine feature selection and identify systematic misclassifications. The generative phase explores whether neural audio‑generation architectures (e.g., MusicVAE, Jukebox, diffusion‑based models) can internalize perceptual boundaries of musicality. Models are trained on subsets of the labeled dataset to produce new audio samples conditioned on the binary label. These outputs are evaluated in a follow‑up behavioral task to determine whether human listeners classify the generated samples in ways consistent with the model’s intended conditioning. Together, these methods provide a computational framework for modeling human concepts of musicality, integrating supervised learning, generative modeling, and human evaluation. The project contributes to machine learning research on audio classification, representation learning, and cross‑cultural sound perception.