Phonetic classification using multi-layer perceptrons

Several extensions to the authors´ previously published results (Proc. IEEE IC ASSP, p.422-5, 1988) on the constrained task of using multilayer perceptrons to classify the vowels in American English spoken by many speakers and excised from continuous speech are described. For vowel classification, the use of linguistic features is investigated. How the choice of the number of hidden units affects classification accuracy is examined. The use of several initialization techniques to improve accuracy and reduce training is explored. The networks are modified in order to classify 38 vowels and consonants. Methods for input normalization and gain adaptation are investigated, leading to an accuracy of 70%