부산대학교 도서관

Audio understanding and classification tasks are often aided by a reduced dimensionality representation of the source observations. For example, a supervised learning system trained to detect the genre or artist of a piece of music performs better if the input nodes are statistically decorrelated, either to prevent overfitting in the learning process or to 'anchor' similar observations to cluster centroids in the observation space. We provide an alternative approach that decomposes audio observations of music into semantically significant dimensions where each resultant dimension corresponds to the perceived meaning of the audio, and only the most significant meanings (those which are most effective in describing music audio) are kept. We show a fundamentally unsupervised method to obtain this decomposition automatically and compare its performance in a music understanding task against statistical decorrelation approaches such as PCA and non-negative matrix factorization (NMF).