Finding Latent Sources in Recorded Music With a Shift-Invariant HDP
Matthew D. Hoffman, David M. Blei, and Perry R. CookMuch interesting work has been done in recent years on analyzing and finding good representations of music audio data for tasks such as content-based recommendation, audio fingerprinting, and automatic metadata generation. However, popular "bag-of-feature-vector" approaches fail to take into account the way that individual sounds evolve over time, and can only model the qualities of the mixed audio signal, not of individual sounds that occur simultaneously. In this paper, we will present the Shift-Invariant Hierarchical Dirichlet Process (SIHDP), a generative model that allows us to represent songs in terms of the instruments and other sounds that generated them.