Making ACOUSTIC SPECTRAL PROTOTYPES by CLUSTERING WITH UNCORRELATED MARGINAL MIXTURE MODELS
Original Publication Date: 1985-Nov-01
Included in the Prior Art Database: 2005-Feb-19
The present invention relates to methods of defining clusters and acoustic prototypes associated therewith, according to probability distributions which conform more closely than the standard Gaussian distribution to actual speech data. In one approach to speech recognition, sound is defined according to a number d of features (such as spectral energy in a given frequency band). Each feature is represented as a corresponding component of a vector in d-dimensional space. In a training period during which known words are spoken, sets of points (hereafter referred to as "clusters") are determined such that each cluster corresponds to a particular sound type. In determining to which sound type an input utterance most closely corresponds, the input utterance is characterized as a d-dimensional vector.