Improving Interpretable Genre Recognizing with Audio Future Statistics Based on Zygonic Theory

Automatic music genre recognition helps to organize music collections and discover new music pieces. However, state-of-the art classification models often lack of comprehensibility for listeners or musicologists, either because they are based on low-level audio descriptors or complex neural network architectures which operate as black boxes. In this work, we propose novel and interpretable statistics derived from zygonic theory and estimated for discretized audio features. These statistics describe repetitive or similar patterns with respect to meaningful musical properties like energy, harmony, instrumentation, tempo, and rhythm. Compared to the baseline from previous work, which builds models with comprehensible semantic features, we show that for most of tested genres and styles the classification performance increases when semantic features are combined with statistics based on zygonic theory, and the most relevant dimensions are stored after feature selection.