TR-2002-02
Detecting and Interpreting Acoustic Features Using Support Vector Machines
Partha Niyogi; Chris Burges. 16 January, 2002.
Communicated by Partha Niyogi.
Abstract
An important aspect of distinctive feature based approaches to automatic speech
recognition is the formulation of a framework for the robust detection of these
features. We discuss the application of the Support Vector Machines (SVM) that
arise when the structural risk minimization principle is applied to such feature
detection problems. In particular, we consider in some detail the problem of
detecting stop consonants in continuous speech. We track dynamic acoustic properties
and discuss and SVM framework for detecting these sounds. In this paper, we
use both linear and non-linear SVMs and present experimental results to illustrate
the factors upon which superior performance depends. We also relate the detectors
to perceptual phenomena like cateogorical perception and the perceptual magnet
effect. We show how the detectors operate by comparing sounds in a transformed
space leading to many different distance metrics that may then be defined. Only one
of these is compatible with the perceptual magnet effect.
Original Document
The original document is available in Postscript (uploaded 16 January, 2002 by
Partha Niyogi).
