Music Similarity Estimation with the Mean-Covariance Restricted Boltzmann Machine

Existing content-based music similarity estimation methods largely build on complex hand-crafted feature extractors, which are difficult to engineer. As an alternative, unsupervised machine learning allows to learn features empirically from data. We train a recently proposed model, the mean-covariance Restricted Boltzmann Machine, on music spectrogram excerpts and employ it for music similarity estimation. In k-NN based genre retrieval experiments on three datasets, it clearly outperforms MFCC-based methods, beats simple unsupervised feature extraction using k-Means and comes close to the state-of-the-art. This shows that unsupervised feature extraction poses a viable alternative to engineered features.