DATA-DRIVEN RECOMPOSITION USING THE HIERARCHICAL DIRICHLET PROCESS HIDDEN MARKOV MODEL
Matthew D. Hoffman, Perry R. Cook, Blei · 2009
Hidden Markov Models (HMMs) have been widely used in various audio analysis tasks such as speech recognition and genre classification. In this paper we show how HMMs can be used to synthesize new audio clips of unlimited length inspired by the temporal structure and perceptual content of a training recording or set of such recordings. We use Markov chain techniques similar to those that have long been used to generate symbolic data such as text and musical scores to instead generate sequences of continuous audio feature data that can then be transformed into audio using feature-based and concatenative synthesis techniques. Additionally, we explore the use of the Hierarchical Dirichlet Process HMM (HDP-HMM) for music, which sidesteps some difficulties with traditional HMMs, and extend the HDP-HMM to allow multiple song models to be trained simultaneously in a way that allows the blending of different models to produce output that is a hybrid of multiple input recordings. 1.