Cluster-Weighted Modeling for Time Series

9/23/98

Click here to start

Table of Contents

Cluster-Weighted Modeling for Time Series (How to build a Digital Strad) 

Outline

Mimic Synthesis

The goal: A digital violin from observation 

Sensor Violin and Bow

Bowing Styles

Modeling the statistics of the violin

Modeling Sequence - Training

Modeling Sequence - Synthesis

Cluster-Weighted Modeling

Density-Estimation and Prediction

Forecast of y and ?y

CWM Prediction Space

Expectation-Maximization

EM: an information geometric view (Amari, Neural Computation 7, 1994)

EM for Gaussian Mixture Models

EM for Gaussian Mixture Models

EM (continued M-Step)

A posteriori cluster estimation 

Maximum Likelihood Derivation of Model Parameters 

Predicted Spectral Data

Off-line Sound Examples

One-tone violin

One-string violin

Fine

Author: Bernd Schoner 

Email: schoner@media.mit.edu

Home Page: http://www.media.mit.edu