ClusterWeighted Probabilistic Networks for AudioSynthesis.pdfVIP

Cluster-Weighted Probabilistic Networks for Audio-Synthesis Bernd Schoner, Charles Co op er and Neil Gershenfeld  MIT Media Lab oratory Cambridge, MA 02139 fschoner,cmc,neilgg @ Abstract A cluster-weighted framework is presented that allows p owerful and transparent non-linear mo dels to b e built by integration of simple lo cal mo dels. Audio synthesis algorithms such as sampling, lin- ear predictive co ding and sp ectral synthesis b ecome globally non- linear synthesis mo dels that integrate control and sound generation within a single framework. These networks are added additional levels of probabilistic abstraction, e.g. hierarchical structures or Hidden-Markov structures. A system has b een implemented that computes the sound of a violin in real time, given the gesture input of a violinist. 1 Intro duction The b eauty and descriptive p ower of probabilistic network architectures and their graphical representation have b een widely appreciated [5]. Graphical mo dels inte- grate essentially all known machine learning, function approximation and prediction to ols in a single framework. Despite their generality, however, the practical use of these networks has b een limited, since new applications require sp ecic architectures which then must b e designed from scratch. In this pap er we integrate a variety of well-known signal pro cessing algorithms for audio synthesis in a cluster-weighted probabilistic framework, carefully intro ducing just enough generality and terminology as needed for this sp ecic ap