A speech recognition system using a neural network model for vocal shaping

An automatic, isolated, limited vocabulary, multilevel speech recognition system is presented. The system uses a standard backpropagation neural network as the recognizer and linear predictive coding coefficients as the recognition feature. The recognition of an utterance involves the identity (class) and version (quality level). Multilevel classification involves using up to five discrete nonlinear levels that correspond to human assessment. The system software was developed using both Microsoft C and Think C. The result of the multilevel test using a vowel subset achieved 61.8% recognition, and it achieved an average classification of good. The consonant test achieved recognition of 46.5% and 48% for the vowels /a/ and /e/, respectively. The system is intended to be used as a vocal shaping tool by autistic individuals, thus requiring a multilevel recognition scheme