Tutorial: Pattern recognition in acoustic signal processing.

Abstract

This tutorial presents a framework for understanding and comparing applications of pattern recognition in acoustic signal processing. Representative examples will be delimited by two binary features: (1) regression versus classification (inferred variables are continuous versus discrete); (2) instantaneous versus dynamic (inference algorithms consider only an instantaneous observation vector versus inference algorithms integrate observations with knowledge of system dynamics). (1. Regression) problems include imaging and sound source tracking using a device with unknown properties and inverse problems, e.g., articulatory estimation from speech audio. (2. Classification) problems include, e.g., classification of animal and human vocalizations and nonspeech audio events. Instantaneous classification and regression are performed using a universal approximator (neural network, Gaussian mixture, classification, and regression tree), regularized, if necessary, to reduce generalization error (resulting in a support vector machine, regularized neural net, pruned classification tree, or AdaBoost). Dynamic classification and regression are done by imposing a prior to characterize system dynamics. Depending on the prior, the resulting model may be called a hidden Markov model, finite state transducer, dynamic Bayesian network, or conditional random field (dynamic classification), or a Kalman filter, extended Kalman filter, or switching Kalman filter (dynamic regression).