Speaker-independent vowel classification from inaccurate formant features
J. Scalkwyk, P. Vermeulen, Etienne Barnard · 2003
Formant extraction is a notoriously unreliable procedure. Neural networks on the other hand are able to deal with such inaccurate data. It is shown that a multilayer perceptron is able to classify five types of vowels with acceptable accuracy (approximately 74%) when operating on very simple formant-based features.>