Speaker-independent vowel recognition: comparison of backpropagation and trained classification trees

Experiments comparing the performance of trained classification trees to that of multilayer feedforward networks on speaker-independent vowel recognition, using information in a single spectra slice, are presented. The vowel stimuli were exemplars of 12 monophthongal vowels of American English taken from all phonetic contexts in spoken utterances. The training set consisted of 342 vowel tokens provided by 320 speakers, and the test set consisted of 137 tokens provided by a different 100 speakers. The classification trees and neural classifiers were trained and tested on identical data. In addition, experiments were performed to determine the most effective way to present vowel information for classification. The results show that neural nets trained with backpropagation produce better results than classification trees in all comparable experimental conditions