Approaches to speaker independent isolated word recognition

Abstract

The pattern recognition issues involved in speaker-independent isolated word recognition are considered. The discussion centers on results of experiments to assess four different techniques based on the hidden Markov model. These techniques are applied to the recognition of Afrikaan speech as uttered by seven male and three female speakers. The procedures for both classification of unknown speech and derivation of reference models are automated. Accuracy comparable with that of other researchers is obtained. Among the techniques discussed, those that utilize the continuous hidden Markov model give marked by better accuracy.<>