Clustering Deterministic Acoustic Spectral Prototypes With an Adaptive Quadratic Form Using Trace Constraint
Original Publication Date: 1985-Sep-01
Included in the Prior Art Database: 2005-Feb-19
The present invention provides algorithms for obtaining a set of prototypes along with their respective quadratic measures in an attempt to improve labelling in a discrete parameter speech recognition system. In a discrete parameter speech recognition system, successive segments of speech are each characterized as a respective input vector, each component of which represents a spectral energy or some other speech parameter. The space in which the input vectors are defined is partitioned into a finite number of regions or clusters. Each input vector is then associated with one of the clusters and is labelled accordingly. The labels, which are generated as a string, are therefore decoded to recognize words therefrom.