Experimental evaluation of structural features for a speaker-independent voice recognition system

Abstract

A preliminary investigation to evaluate a speaker-independent voice-recognition system based on speaker-invariant feature measurements is presented. The signal is represented as a trajectory in the first-order phase plane, from which a set of features is extracted including uncoded and coded intersection number with the x- and y-axes, respectively. Both nearest-neighbor and K-means clustering are used for classification. The feature set was evaluated for speaker-invariant recognition using an orthographic (written) vowel data. Among the features selected, the uncoded intersection number provided much tighter clustering in the decision space with an error rate of 10% using nearest-neighbor classification.<>