Music Genre Classification via Sequential Wavelet Scattering Feature Learning

Abstract

Various content-based high-level descriptors are used for musical similarity, classification and recommendation tasks. Our study uses wavelet scattering coefficients as features providing both translation-invariant representation and transient characterizations of audio signal to predict musical genre. Extracted features are fed to sequential architectures to model temporal dependencies of musical piece more efficiently. Competitive classification results are obtained against hand-engineered feature based frameworks with proposed technique.